GitHub Finds for AI Builders

Recursive describes Flint in its own words as "Recursive's mixture-of-experts language model for coding agents, with 39.8 billion total parameters and approximately 2.5 billion active per token." It is built to dig through a codebase, fix bugs, add features, and write tests, and it can hold a long back-and-forth while calling outside tools, which is what an agent needs to actually do work rather than just talk. In Recursive's own evaluations, it scores 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, both benchmarks built from real bugs in open-source projects. Unsloth, a team known for shrinking large models down to run on ordinary hardware, packaged it into ready local files; of the compressed options, the one unsloth labels Q4_K_XL runs about 25 GB, and a much smaller but noticeably weaker one runs about 11 GB. Because only a fraction of its parameters activate per token, it does not need to load the full 40 billion into memory at once, which is why it fits on a single well-equipped machine at all.

Black Forest Labs is a German team behind the whole FLUX family, one of the best-known open image-generation models. The authors describe FLUX 3 Image Klein as "an open model with 5.6 billion parameters for text-to-image generation and image editing." You write a text description and get a picture, or go further and hand it up to ten of your own photos, and it builds one new scene pulling details from each. This is the lightweight, ready-to-run version of the FLUX 3 release — the other version, Base, is meant for fine-tuning and research, while Klein is for people who just want to install it themselves. The license here is different from normal open-source code, so read the terms yourself before selling anything made with it.

VictorTaelin/uniichat

UniiChat is an AI agent that has a single chat with no “new chat” button. You can chat with it for days, weeks, or even years! Each message gets compressed down to a single line. These lines get further compressed into shorter lines over time. Information gets lost, but never deleted. When the agent wants to recall past information, it can expand a short line back into its initial message. This compression happens in the background using a second, cheaper AI model as you type your response. The same chat can be accessed both via a web browser and via a command line interface (CLI) tool named unii. In addition, on macOS, the agent is able to interact with the screen in a special mode. The agent is able to spawn background sub-agents that perform tasks and then send their results back to the main chat. The agent is able to import historical chats from Claude Code (Anthropic’s coding tool), Codex (OpenAI’s coding tool), or export conversations from ChatGPT. To install UniiChat, simply ask your agent to deploy UniiChat from github.com/VictorTaelin/uniichat following the instructions in its README (i.e., the project’s “setup guide”). Note that you will need Bun (a high performance JavaScript runtime) and your own Anthropic or OpenAI API key in order to run the server. This runs on a regular laptop without a GPU.

CopilotKit/OpenDots

OpenDots is a free template from CopilotKit, a company that builds AI assistant tools. It's for always-on AI 'coworkers' that move between chat, phone calls, and Slack, a workplace messaging app. Your documents live in 'Spaces' — shared workspaces. Each agent is called a 'Dot.' Every Dot has its own name, role, instructions, and list of allowed tools. Any Dot can get its own isolated computer — a separate browser and file system. This uses OpenBot, another CopilotKit project. Its browser profile and files are kept between restarts. Before a Dot saves a page or touches a real account, it shows a card with 'approve' and 'decline' buttons. It waits for your decision. This covers things like email, a calendar through a service called Composio, or any tool with write access. All chats and agent work are stored in CopilotKit Intelligence. It comes as a paid cloud service, a local test version, or a separately licensed self-hosted option. The local version runs in Docker, a tool for running software in isolated containers. The authors call it a starting point for building, not a finished product. It's in alpha, an early test stage. Some features, like calls and handing work off to Slack, are only tested locally so far. To install it, ask your coding AI agent to deploy OpenDots from github.com/CopilotKit/OpenDots. You'll need Node.js 24, a JavaScript runtime, and npm. You'll also need a CopilotKit Intelligence account and a model key, such as OpenAI's. Then have it run npm run dev and open http://127.0.0.1:5173. It runs on a regular laptop — no graphics card needed.

cheaper-inference/OmniRoute

OmniRoute is a free, open-source gateway (a program that routes your requests) you install on your own computer, giving you one local address instead of setting up each AI tool separately. Behind that address are 375 AI model providers, more than 150 of them with free access you sign up for yourself under their own terms, and the gateway can automatically switch between them if one is down or capped. It also compresses: a chain of filters shrinks, on the fly, the text that normally eats tokens for nothing — repeated command output, logs, long tool replies — and the project's own documentation claims this saves 15 to 95% of tokens per request, while leaving code, links, and structured data untouched. You connect it to regular coding agents and editors — Claude Code (an AI coding tool), Codex, Cursor, and others — through that same local address; it's also available as an MCP server (a way for an AI agent to control other tools), so the agent itself can manage the gateway: add providers, change compression settings, check status. Per the project, requests go straight to the chosen provider with no OmniRoute cloud server in between, and stored keys are encrypted. To install: ask your agent to set up OmniRoute from github.com/diegosouzapw/OmniRoute, following the project's own instructions page.

Meta has released their new open-source (downloadable and runnable) model, Muse Spark Open. It’s an 816 billion parameter model with 42 billion effective params per token (word piece) generated (MoE). This means less computation everywhere. With a context size of 1 million tokens, this model can process a lot of information at once (conversation/code). According to the paper, this model was designed for code/agent tasks but can also understand text/image/video/audio and respond back in text. They report a score of 76.4 on Terminal-Bench 2.1 (an agent benchmark in a terminal environment, i.e. command line interface). This benchmark hasn’t been verified by Hugging Face (model hosting service). Licensing is under the Muse Model License which allows you to deploy this model in your application without paying Meta, unless you or your affiliates generate more than $20M in revenue annually, then you will have to negotiate a separate contract with Meta. Play around with this model here in HuggingChat (no hardware required) by visiting https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B, giving it an agent task in a terminal and watching it reason through things and write some code.

wesbos/3d-model-anything

The project's author is Wes Bos, a well-known web developer and teacher many vibe-coders know from his courses. You scan a room with an iPhone or iPad that has LiDAR, a sensor that measures depth. That gives you a file with the walls and objects, plus photos linked to where they were taken. The project includes a 'skill' — step-by-step instructions for an AI agent. The agent reads those files. It builds the walls, openings and objects into an editable 3D scene. Then it checks that scene against the real photos, looking for mismatches in shape or position. The authors are upfront: this isn't a one-click scan-to-model button. It's a step-by-step process where the agent keeps refining the scene. For now it only handles rooms and interiors. Once it's built, you open the scene in a browser. You can walk through it in several modes, from an overhead view to a first-person walk. To try it, ask your agent to set up github.com/wesbos/3d-model-anything. It should run npm ci and npm run dev, then open the demo office in your browser. Any laptop with Node.js, a program that runs the code, works for that. You only need a LiDAR iPhone or iPad to scan your own room.

Nemotron 3.5 Bolt is an agent model developed by NVIDIA, a semiconductor company. It’s designed to perform common agent actions such as browsing the web, opening a terminal, invoking functions, reading documents, code, and screens. This model was created by compressing NVIDIA’s Nemotron 3.5 Super VL (a 120 billion parameter model) into a 74 billion parameter model with just 9 billion parameters active. It uses a hybrid architecture combining a Mamba-2 model and a transformer model along with a vision component allowing it to process both text and images/videos. With a context length of up to 256,000 tokens (word pieces), this model can handle quite a bit of information at once. This compressed version of the model weighs in at approximately 24 gigabytes. If you want to give this model a spin, simply tell your agent to install Nemotron 3.5 Bolt from NVIDIA using Ollama (an AI model execution framework that allows you to run AI models locally) by typing ‘ollama run nemotron-3.5-bolt’. Note that this requires a graphics card with at least 24GB of memory or a Mac with at least 32GB of memory (NVIDIA recommends using professional GPUs such as the A100, so this is definitely not something you can run on your average laptop).

Recursive’s first publicly released model, Flint, is available on the Hugging Face model hub (model handle: RecursiveSI). This is a mixture of experts model, meaning that only a subset of the model is used at any given time. With 39.8 billion parameters but only ~2.5 billion activated at any one time, Flint is relatively lightweight despite its size. You can ask Flint to complete any programming task, such as debugging, implementing new features, writing tests, etc., and watch as it completes these tasks autonomously, including interacting with external tools, verifying its output, etc. To use Flint, install the Ollama framework for running local AI models via the command ollama run flint. Note that the full model requires a datacenter GPU, while there is also a smaller model available for MacOS that requires at least 48GB of RAM.

This is a retrained version of Qwen3-4B-Instruct, a small open AI model. The habits that give away AI-written text have been trained out of it. It was made by erikkannike, who publishes on Hugging Face (a site for sharing AI models). They also write a blog at 100toolkit.com, where they lay out the whole process step by step. In the project's own description, the result is called "AI-assistant writing habits trained out of its weights." The habits are removed from the model itself, not hidden behind a system prompt (an instruction given before every answer). The work was done solo, in a home lab, on a MacBook Air. First, the author found the direction inside the model responsible for its "assistant" tone. Then they adjusted individual neurons — the model's smallest building blocks. Finally, a teacher version of the model answered 288 training questions. Only the answers free of typical AI clichés were kept for training. On the test topics, the AI-cliché count fell from 121 to 12.5 per 1,000 words. Filler dashes dropped from 20.5 to zero. The model still answered all 30 fact-and-math test questions correctly. To install it, run one command in Ollama (a tool for running AI models on your computer): ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M.

MOSS-Image-SFT is an open model for generating and editing images, made by OpenMOSS-Team, the same lab behind the MOSS language model. It's a fine-tuned version of their base model, MOSS-Image-Base — fine-tuned means retrained further on extra examples to sharpen one skill, and it's built on the same underlying design. A normal image model gets only a text description and decides on its own where to place letters, so text inside pictures often comes out blurry or garbled. Here, you can set text and objects separately, through a JSON layout — a structured note on what to draw, what text to write, and where to place it. The makers say this fine-tuned version groups text better and lays it out more neatly than the base model. It's 5 billion parameters in size, a rough measure of model size, under the Apache 2.0 license, free for commercial use with no restrictions. The community already built a plugin for ComfyUI, a no-code tool for running image models, so you don't have to write any code yourself. Ask your agent: install the model OpenMOSS-Team/MOSS-Image-SFT from Hugging Face, a site where people share free AI models, using ComfyUI or diffusers — a code library for running such models — whichever fits your computer better. Generate an image with readable text inside, like a cover with a caption, and show the result. You'll need a regular laptop with a graphics card of 8 GB or more; compressed, the model takes about 3 GB, in full quality about 10 GB.

Liquid AI is a company that builds AI models made to run on your own computer, not in the cloud. It calls its lineup LFM2.5, and LFM2.5-24B-A2B is the newest, biggest model in that family. It has 24 billion parameters, but only 2.45 billion are active on any task at once; the model switches between 64 smaller expert parts depending on what you ask it. It can reason before answering and call outside tools, which is what agents need. It handles 131,000 tokens of context at once (tokens are chunks of text the model reads), and Russian is one of its listed languages. The model's page on Hugging Face, a site that hosts AI models, already has ready-to-run files for an ordinary computer, no server or conversion needed: huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF. These are GGUF files, a format built for running models locally, run by a program called llama.cpp. The authors also made a smaller, four-bit compressed version and retrained it to recover the accuracy that kind of compression usually loses. You need a laptop or Mac with at least 32GB of unified memory; the smallest file is about 13.5GB, a more accurate one about 14.5GB. To set it up, tell your AI agent: "Install the LiquidAI LFM2.5-24B-A2B model locally: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF — pick the file that fits my memory, run it with llama.cpp, and show me how to talk to it and give it tools to call."

Jakeschincariol/arena-skill

/arena is a skill for Claude Code, Claude's coding tool. The command spins up 100 copies of Claude, each with its own way of reasoning, and pits them against each other on your task: they attack each other's answers, defend, get judged round by round, and the loser is cut — until one winner is left. To install it, paste the GitHub link into Claude Code and ask it to install the skill, then confirm /arena works. It needs Claude Code and Python, nothing else.

heygen-com/hyperframes

The people behind HyperFrames call it an open-source (freely available, editable) framework that turns an HTML page - the format used to build web pages - plus images, video clips and music into a finished MP4 video file, always frame-by-frame identical when you render the same file again. It was built by HeyGen, a company that makes AI-avatar video tools and uses HyperFrames itself. You describe a video as a regular HTML file: the title, a background video and music each get a timing tag, and the tool renders it on its own, running a hidden web browser in the background together with FFmpeg (a free tool that assembles and compresses video). Because the same file always gives the same result, you can run it inside automated pipelines - repeating steps with no one checking each render by hand. For Claude Code, Cursor, Codex and other AI coding agents there's a ready set of 21 skills (ready-made task scripts): the agent doesn't invent the video logic itself, it just picks the right scenario. One scenario is a product announcement. Another turns a GitHub pull request (a proposed code change) into a short explainer, with the changed lines of code highlighted and animated on screen. The license is Apache 2.0, with no fee per render. The closest comparable tool, a library called Remotion built on React (a popular toolkit for building web interfaces), works differently - its free use has a limit on commercial projects. To try it: ask your agent to add the HyperFrames plugin to Claude Code and render a short announcement video with a title and background music.

DeepSeek pushed out an updated version of its flagship model, V4.1 Flash, now carrying the index 1017, and the authors call it an improvement for reasoning, coding and agentic tasks. It weighs 763 billion parameters, and even a compressed version for local use takes several hundred gigabytes, so running it on a home computer is not realistic. On one of the autonomous coding benchmarks it scored 77 points, which does look strong on paper. I work with agents in Claude Code every day, and with Chinese models of this size the pattern usually repeats: good on paper, weaker once you give them real agentic work.

OpenBMB showed that their new model for understanding images and video, MiniCPM-V 4.7, runs right on an iPhone, with no server and no internet connection — the authors claim over 13 tokens per second on an iPhone 17 Pro Max, using about 3 GB on the device. The model reads a single image, several images together, or video, so you can hand it a screenshot or a clip and get a plain description back. For a desktop there is a GGUF version too, so setting it up on a regular computer is not a problem either. It is a neat technical trick, but I would not rely on it for serious work — this class of model is fine for simple, repeated tasks, not for real agentic work.

FishAudio released the third generation of its voice model, Drama 3, and they call it their new text-to-speech model built for multi-character scenes and long-form content. In one recording it can give up to 16 different voices, clone a voice from a short sample, and control tone through simple text tags, marking a line as angry or whispered, for example. The authors mention 83 supported languages and automatic splitting of long texts, so narrating a whole book chapter does not mean cutting the text by hand yourself. On heavy server GPUs they get under 100 milliseconds per line, but what a regular graphics card would give is not stated in the card.

Black Forest Labs describe FLUX 3 Image Klein as an open-weights model with 5.6 billion parameters for generating and editing images, and this time they packaged it for ComfyUI themselves, with ready-made workflows included, so there is no need to assemble the model from pieces yourself. You can generate an image from a description, edit a single photo by instruction, or compose a scene from several reference images. The lighter version runs on 8 GB of VRAM, which fits even an older RTX 3070. The weights are open, and under the license, commercial use of the output is free as long as your company's revenue is under 5 million dollars a year.

Lightricks describe LTX-2.5 as an open-weights model built for local execution and fine-tuning, with its established use being generating synchronized video and audio from a text, image, or video input. It comes as two 22-billion-parameter transformer versions, a fast distilled one and a full one for fine-tuning. For weaker GPUs the authors mention a lighter mode with memory offload, but they do not name an exact minimum card, which is an annoying gap if you are planning to run it at home. Under the license, commercial use is free as long as your company's revenue is under 10 million dollars a year.

Nvidia, a chip maker, describes this model as a 'vision-language model for reasoning, coding, tool use, document understanding, and video understanding.' In plain words, it reads text, images, and video, then uses that to write code, follow instructions, and work with documents. This release is compressed using a format called NVFP4, which shrinks how much memory the model needs. Nvidia built this smaller version to match its original 16-bit model as closely as possible. It can hold up to 1 million tokens of context — the text, images, and history it can consider at once. Results vary by test. On the coding test SWE-bench Verified it scores 71.07, almost identical to the original's 71.10. But on other tests the gap is real. On the browsing test WebVoyager it drops from 77.72 to 66.31. On MMLU-ProX, a general knowledge test, it drops from 76.91 to 73.76. You can't run this at home. It needs one DGX Spark station, one H200 GPU, two H100 GPUs, or a GB200 system — all server-grade hardware. To try it, ask your AI agent to open the NVIDIA Nemotron 3.5 Super VL model page on Hugging Face, a site for sharing AI models, and explain what code, document, and video tasks it can handle, since it's a server model you can't install at home.

trailhq/Trail-Code

Trail-code was previously known as Graft. Trail-code provides a roadmap of your code base to AI code agents such as Claude Code, Cursor, Codex and Gemini which can write code for you. This tool works offline on your machine. Tree sitter, a code parsing library, parses each file in your code base looking for functions and call graphs. There are no AI models or internet connections needed. Works with 23 different programming languages. The “learnings” feature allows you to save what you learned from correcting your agent or collaborating with it so that future sessions (or other users) can benefit from your experience. They ran this tool on their dataset of 162 runs and on an industry benchmark called SWE-bench Verified. They used Anthropic’s AI model Claude Sonnet 5 as their agent. Using the roadmap, the agent was able to complete 33 out of 50 real world coding problems. Without the roadmap, the agent completed 27. You can install this tool by running npm install -g @trailhq/trail-code followed by trail init in your project directory. There is no background process, everything is stored locally on your disk.

TencentCloud/Octop

Octop is an open-source AI assistant you install and run yourself, on your own computer. Tencent Cloud, the cloud arm of Chinese tech giant Tencent, built it. It supports several people and several "expert" agents working at once. You can talk to it through a web dashboard, Telegram, Discord, WeChat, QQ, or the Chinese workplace chat apps Feishu, DingTalk, and WeCom, plus a direct connection method for your own integrations. You can build a library of "experts" for different tasks, including ready-made personality templates the authors call MBTI personas, and switch between them. One admin account covers several users, which fits a family or a small team. A beta feature called AgentTeams lets one coordinator agent split a multi-step job across several experts. Octop has no model built in. You connect one through an OpenAI-compatible key, a connection format many AI providers support, through DashScope, a cloud service from Chinese company Alibaba that hosts its Qwen models, or by running a model locally with Ollama, a free tool for running AI models on your own machine. To install it, run the one-line setup command on the project's GitHub page, following its README, the project's instructions file.

storytold/filmcraft

FilmCraft is an open-source professional video editor, built by ArtCraft, an open-source creative-tools group, entirely in Rust, a fast, safety-focused programming language. It has its own codecs (the tech that compresses and plays back video and audio) for H.264, ProRes and AAC. It also does GPU-based image processing and frame-exact timing. The interface is familiar: panels, trimming, color correction with wheels and curves, titles and subtitles, like other pro editing apps. Here's the part that matters for this story: every menu, button and slider is also a separate command, with an ID and typed settings. There are over 650 of them. You can trigger the same commands from the interface, from the command line (text-only control), over a network channel in JSON (a common data format), or through FilmCraft's own MCP server (a standard way AI agents plug into software). That means Claude or another AI agent can edit, grade and export video as precisely as a person would. The authors say the trailer and color grading shown in their README (the project's own instructions page) were made by an agent controlling a running copy of the app, not by a human. Hardware speeds up decoding: VideoToolbox on Mac, and on Windows and Linux, NVIDIA's NVDEC chip, if you have the right graphics card. For export, NVIDIA's NVENC chip speeds up H.264 on both Windows and Linux. Sped-up HEVC export only works on Windows (via NVENC) and Mac (via VideoToolbox), not on Linux. FilmCraft's own code can decode HEVC footage but can't encode it; HEVC export only happens through that hardware path. The project calls itself an alpha. Some features, like importing Adobe Premiere projects or exporting to AV1 (a newer video format), aren't ready yet. It runs natively on macOS, Windows and Linux, and in the browser via WebAssembly (a way to run heavy code inside a browser). To install: grab a ready build from getartcraft.com/apps/filmcraft, or build it from source following the README, then open the demo project and screenshot the result.

paperclipai/paperclip

Paperclip is a Node.js server — a backend program — with a web dashboard. It's built to manage a team of AI agents doing real work. The project lays out three steps. First, you set a goal in plain words. Second, you hire a team of agents for roles like CEO, CTO, engineer, designer, marketer. Third, you review the plan, set a budget, and watch it run from the dashboard. Paperclip doesn't give you the agents itself. It connects to agents you already use. Claude Code is Anthropic's coding agent, and it's one of them. Codex is OpenAI's coding agent, and it's another. Paperclip links to each one through its own built-in connector. To install it, run `npx paperclipai@latest onboard --yes`. That command comes from the project's README, its instructions page on GitHub.

Niko1221/Strata

Strata is an installer and local server for Qwen3.8-Flash-Next, a 125-billion-parameter model from the Qwen team. It's normally built for server hardware. The first-run wizard checks your graphics card and memory, asks which version of the model to install and how big a context (how much text it can remember at once) you need, then downloads a ready-made build of the engine for your card — NVIDIA, AMD or Intel — or compiles one on the spot if none exists. After setup, the model runs as a local web server at 127.0.0.1:8080, speaking the same protocols as OpenAI's and Anthropic's APIs. That means you can point Claude Code, Codex CLI, OpenCode or similar coding agents at it just by changing the server address. The model reads pictures by default and can work with files and code. It's MIT-licensed, has no telemetry and needs no account; the repo has a separate document listing exactly what gets downloaded and when the program talks to the internet.

mistralai/timestral

Timestral is a family of models from Mistral AI, a company that builds AI models. These models forecast time series — numbers that change over time, like sales or temperature. There are four sizes, from 5 to 209 million parameters. They weigh between 21 and 837 MB. All of them sit on Hugging Face, a site that hosts AI models, under Mistral AI's own account. For each future step, the model predicts a range of likely values, not just one number. It can look back up to 32,768 steps of past data. On Mistral's own benchmarks at launch, the biggest version, Timestral-209M, ranked first among pretrained models. That was across three major open forecasting tests. This repo only holds the code that runs the model. Ask your agent to install it with pip install git+https://github.com/mistralai/timestral.git. Then ask it to run the demo forecast example from the README, the project's instructions page, on test data. The model's weights download on their own from Hugging Face the first time you run it.

Jakeschincariol/instagram-agent-skill

This is a set of thirteen skills for Claude, the AI model you talk to. They write Reel scripts, score the "hook" (the opening seconds of a video that decide if a viewer keeps watching) against known formulas, write captions that account for where Instagram cuts the feed text, score your profile out of 100, and strip invisible characters and typical AI-sounding phrases left over from AI text generation. Install by pasting the repo link into Claude and asking it to install the skills. Only you post anything.

diegosouzapw/OmniRoute

OmniRoute is a free, open-source gateway — a middleman server — that connects your coding tool to hundreds of AI providers through one address. Install it, point your tool at the address it gives you, and it replies right away through a free "auto" model, with no key or signup needed. You only connect your own provider account if you want more than that free routing.

alchaincyf/huashu-motion

This is a skill — a ready-made set of instructions — for AI coding agents like Claude Code (Anthropic's coding assistant) and Codex (OpenAI's coding agent). It works through the Agent Skills protocol, a shared format that lets different agents load the same skill. You tell the agent: make a 30-second promo video for my product. It reads about your product, then offers three style directions to pick from. It writes the full script and breaks the scenes into a beat-by-beat rhythm. For music, the agent doesn't use a stock track — it composes its own score. It plays that score with real piano, string, and drum samples — recordings of real instruments. If you haven't downloaded the samples, it uses synthesized sound instead. The finished video is just a web page with code-based animation, not a video file. It opens in a browser-based editor. You click any frame, type your complaint in one line, and the agent fixes only that spot. Everything else stays untouched. The README's author says he broke down 21 real promo videos from known companies and 142 popular animation videos, frame by frame. From that he extracted about 20 named "motion languages" for different product types. That library is the core of the skill, not ready templates. To install it, tell your agent: install the huashu-motion skill from github.com/alchaincyf/huashu-motion, using the command `npx skills add alchaincyf/huashu-motion`. Then ask it to build a demo promo video for your product and show you the result. It's only been tested on macOS. You'll need Node 22+ (a JavaScript runtime), Chrome, and FFmpeg (a video-encoding tool). No graphics card is needed — rendering runs through your own browser.

storytold/gridcraft

GridCraft is a free spreadsheet program, and the ArtCraft team writes it from scratch in the Rust language as a replacement for Microsoft Excel. Inside there is everything an Excel user is used to, from the formula bar to PivotTables and charts, and the native file format is the normal .xlsx. By the count of the authors, it covers 89% of the Excel ribbon commands in their catalog and 513 of the 523 Excel worksheet functions. The most important part for us is that every button, menu item and dialog exists as a command. An agent calls these commands through an MCP server, through a command-line tool or through a control channel to the running window. Through MCP the agent reads and writes ranges of cells, evaluates formulas, formats them, inserts charts, creates tables, sorts and filters the data. The program runs on macOS, Windows, Linux and BSD, and also in the browser, without a subscription, without a licence server and without telemetry. It is an alpha, the authors estimate it at about 50% ready for real work, and the fidelity of real-world .xlsx files is the first point in their plans.

edrisranjbar/lifeos

Edi Life OS is a self-hosted dashboard that puts ten workspaces behind one login and one database: an overview with a daily score, a growth section linking long-term goals to SMART goals and habits, a focus timer, a finance ledger for expenses and debts, a habit tracker, a kanban board, a calendar, goals, a markdown notepad and sticky notes. The workspaces are connected rather than separate: a habit drives goal progress, a kanban card with a due date shows up on the calendar, and the daily score is built from focus, habits, goals and work together. A built-in MCP server exposes 15 tools over a token-protected HTTP API, so Claude Desktop, Claude Code or any other MCP client can read the dashboard, create a goal, tick off a habit, or log an expense when you ask in plain language; the server never touches the database or credentials directly. It deploys with one `docker compose up`, with MySQL, the app and file attachments each in their own volume, or as a plain PHP+MySQL site on shared hosting.

darshi1337/apogee

Apogee is a browser extension that summarizes whatever you are looking at: an article, a selected passage, a PDF or DOCX file, pasted text, or a YouTube/Bilibili video. Open the popup or side panel, pick a format, and it produces a bullet summary where clicking any bullet scrolls the original page to that exact passage; for video it builds a clickable timeline of key moments instead. By default the model runs on your own graphics card through WebGPU, or on the CPU through WebAssembly if there is no GPU, so the extension needs no account, API key, backend server or cloud upload. For longer or trickier documents it can also talk to your own local Ollama or llama.cpp server over your machine's loopback address, which keeps everything local while giving you a bigger model to work with.

storytold/wordcraft

WordCraft is a free word processor — a program for writing and formatting documents. A team called ArtCraft is rebuilding it from zero in Rust, a programming language. None of Microsoft's own code is inside. It has Word's ribbon (the tabbed toolbar), styles, footnotes, and a references list in APA, MLA, Chicago and IEEE formats. It also has track changes with margin comments, and mail merge from a CSV spreadsheet file. All of it is free, no account needed. It runs on macOS, Windows, Linux, BSD, and in a browser, from one shared codebase. Here's the detail that matters for an audience like ours: each of its 422 actions is its own named command. Agents reach those same commands three ways: an MCP server, a command-line tool, or a JSON channel. MCP is a channel that lets AI tools control an app directly. An agent can type text, apply a style, insert a table, and save the file. Then it can check its own work by reading the document's structure back, with no screenshots involved. Paragraph layout is cached, so editing a 188-page document redraws in about 1.4 milliseconds. It's alpha software. The authors themselves put it at 88% of Word's ribbon features done. It's about 55-60% of the way to replacing Word for everyday work. To try it, open github.com/storytold/wordcraft, build it or grab an installer from the Releases page. Connect it to Claude Code with claude mcp add wordcraft -- wordcraft-cli mcp. Then ask the agent to open a test document, apply styles, and show you the result.

Liquid AI builds small models meant to run on everyday devices. They call d1-3B a decision model — 3 billion parameters, built on their LFM2.5-VL-3B model that reads text and images. You feed it a situation: text, JSON (a structured data format), an image, or a mix of these. Then you give it questions — yes/no, pick one option, or a rating from 1 to 10. It answers all of them in one pass, with zero written text. Each answer is a calibrated probability, not a sentence. The company says plainly: this isn't a chat model, it can't write text and isn't meant to. No installation needed to try it. Open Liquid AI's d1 Arcade demo at huggingface.co/spaces/LiquidAI/system-one-arcade and test a question right in your browser.

EmbeddingGemma 2 is a new model by Google DeepMind (Google’s AI research division). This model is NOT a chatbot (it won’t respond to your queries or generate text). It maps text, images, videos & audio to a common embedding space. An embedding is a numerical representation of the similarity between different items. Code can be included in the text. Supports more than 100 languages. Improves code search by 14% compared to previous model. Small model (only 740 million parameters), designed to be deployed on mobile/laptops rather than servers. Ask your agent to install EmbeddingGemma 2 using this link and then ask it to setup local semantic search for a directory containing your notes / docs. You will need a regular laptop with 8GB+ of memory.

Qwen3.8-27B is an AI reasoning & tool calling model developed by Chinese company Alibaba. This is a downsized version of Qwen3.8-27B. This model was created by Underdog, who published it on Hugging Face(a platform for sharing AI models) under the username ConwayResearch. The full size Qwen3.8-27B model is 54GB while this downsized model is only 7.89GB. This model is specially designed to preserve the tool calling ability of the original model instead of just reducing its size. In the Berkeley Function Calling Benchmark(a benchmark for testing tool calling ability), this downsized model completed 88/120 tasks compared to 84/120 tasks completed by the full size model. In 9 benchmarks, this model achieves an average score of 96% compared to the full size model. This model performs poorly when dealing with difficult math problems(e.g., on AIME 2025 math competition, this model got a score of 79.2 while the full size model got a score of 96.7). You can use this model to process images too. Just download the small vision plugin separately and put it together with this model. Here’s how to use this model: Copy the url of the model’s webpage, tell your code agent to download the model file, use llama.cpp(a program used to run model files locally) to run the model file and register this model file as a tool calling model for your code agent.

I put Clef from Cloudflare on my own test support tickets and gave it several questions at once: is the client angry, does it need escalation, what type of problem is it. The model does not write an answer, it returns a probability for every option right away, and it does this in one pass with 27 billion parameters. By Cloudflare's own table, on the BANKING77 test (sorting of banking requests) Clef scores 94.2 against 79.7 for the paid Jev, which it competes with directly. I liked that it does not get confused on narrow sorting, but for questions that need general knowledge I would not choose it, there the paid Jev wins by a wide margin.

I set up Qwen3.8-27B from Alibaba locally and gave it a screenshot of code plus a question in Russian. It natively understands images and video, not only text, and it can decide on its own whether to think fast or think long before answering. Compared with the previous Qwen3.6-27B the gain is noticeable, on the SWE-bench Pro coding test it is 61.7% against 53.5%. For local use there are already more than 700 quantized versions, including an official one from unsloth, so setting it up is not hard.

I set up Laya from Convai Innovations (the handle on Hugging Face is convaiinnovations) and ran a few test support emails through it with standard questions. It also answers not with text but with a probability in one step, the English version has 421 million parameters, and a separate multilingual version with 322 million covers over 100 languages. On a GPU the answer to one question takes 32.8 milliseconds, the paid Jev needs 236 to 276 milliseconds for the same question, and on a plain CPU, without a graphics card, Laya needs 0.2 to 0.5 seconds. The one thing that bothers me is that on questions with 20 or more answer options it starts to confuse them, and here fine-tuning on your own data raises accuracy from 0.362 to 0.766.

This model was released by Qwen, Alibaba's Chinese AI lab. Alibaba is a Chinese tech giant. Qwen-Image-2.1-Turbo is a faster version of their image model, Qwen-Image 2.1. It's the same 7-billion-parameter system for turning text into pictures and editing existing ones. It just builds the image in 8 steps instead of the full process. The settings are already tuned, so you don't adjust them yourself. To install it, ask your agent to get Qwen-Image-2.1-Turbo from Hugging Face — a hub site where AI models are shared. The page is huggingface.co/Qwen/Qwen-Image-2.1-Turbo — ask it to pick a compressed version that fits your computer and generate a test picture.

This model takes a finished text - a letter, an essay, a report, a forum post - and rewrites it so it sounds like a person wrote it, not an AI. It's made by an independent developer going by jialinyyzz, built on Gemma, a base AI model from Google, with 12 billion parameters (a rough measure of a model's size). When it rewrites, it keeps every number, date, name and quote exactly as they were - it only changes the style and rhythm of the sentences. Tested with a strict AI-detector called Originality.ai, 95% of rewritten drafts read as human-written; the detector flagged AI in only 11 of 210 texts, down from 26 for the previous version. The developer also checked for accuracy: all 420 rewritten English texts were fact-checked, and 376 of them came back with no factual errors at all. It runs offline, through a program for running AI models on your own computer, or directly through llama.cpp - install it on an ordinary laptop with 16 GB of memory; compressed, the model takes up about 7 GB.

Baidu — the Chinese tech giant behind a major search engine and cloud platform — made this model. It reads text out of images and documents: single photos, multi-page scans, PDF files. The authors call it a sequel to DeepSeek-OCR, another open OCR model, and say it can read a long multi-page document in one pass, without splitting it into pieces — that's the 'unlimited' in the name. It's small, 3 billion parameters, under an MIT license with no restrictions on use. Compressed versions are already out, ready to run with llama.cpp (a tool for running models locally), LM Studio, Jan, or Ollama.

Whistle is a small speech recognition AI model developed by Cactus-Compute. Speech recognition refers to converting spoken language into written text. This model can be used in mobile phones, wearables, robots, home automation systems, vehicles and microcontrollers (small computers with limited memory). This model is just a single file of size 16.9 megabytes. It uses the same processing core as another AI model developed by Cactus-Compute called Needle. This model doesn’t require any GPU (Graphics Processing Unit) or additional software. Whistle performs the following tasks on a device: Transcribes up to 30 seconds of audio in English, German, French, Spanish, Italian, Dutch and Polish. Provides time stamps for beginning and end of each word along with its confidence score. Generates audio fingerprints for searching and comparing audio files. For example, if your device plays back a recording of ‘turn off the kitchen lights’ then this model will output the exact same text. This model is available under Apache 2.0 License. This means that you can freely download and modify this model according to your needs including commercial purposes. This model can be easily installed as a python package (Python is a popular programming language). Also, there is a C interface for this model that can be used in embedded systems (tiny computers). Ask your agent to go to this link of Whistle on Hugging Face (a platform for sharing AI models) and then ask it to install the package, recognize some english audio file and print the recognized text along with time stamps for each word. You can run this model on any normal computer or even on your mobile phone. No GPU is required.

storytold/vectorcraft

VectorCraft is, in the authors' own words, a vector illustration program: an open-source, clean-room rebuild of Adobe Illustrator (Adobe's paid design app), written from scratch in Rust, a programming language. It runs natively on macOS, Windows, Linux and FreeBSD, and in the browser, and keeps Illustrator's familiar tool layout — pen, direct selection, Pathfinder, Smart Guides, and panels like Appearance, Swatches and Layers — so there's almost nothing to relearn. Rendering is multithreaded and uses the processor's SIMD instructions (a way to do many calculations at once) instead of the graphics card, so 20,000 shapes render in about 27 milliseconds at full retina resolution while the interface still holds 120 frames per second. Every tool, menu and dialog is also reachable through a JSON control channel and an MCP server (a standard way for AI agents to call outside tools), so an agent like Claude can draw, edit and export through the same command path a human uses, not through a layer of scripts bolted on top. The README states plainly that every illustration shown there, gradients and posters included, was made entirely through that same interface, which is the main proof that the agent route actually works and isn't just a demo. The authors are upfront about where the project stands: by their own estimate, roughly 69-75% of Illustrator's features exist and work, and only 40-55% reach the point where a professional couldn't tell the difference. What's missing is listed too: 3D and Photoshop-style raster effects, scripting, interaction-detail polish, and proper installers for Windows and Linux, so right now it's a solid foundation rather than a finished replacement.

storytold/pdfcraft

PdfCraft is a free, open-source program for reading and editing PDF files (a common document format). It's built in Rust (a programming language) by ArtCraft, a community of people who build creative apps together. The authors call it a PDF workbench — an open clone of Adobe Acrobat. It's built from scratch, without using any of Acrobat's own code. It runs on macOS (including Apple's own chips), Windows, Linux, FreeBSD, and even right in a browser. Your document never leaves your machine — no account, no tracking, no cloud step. Inside it there's an agent layer: an MCP server, the standard way AI agents talk to tools. It stays off by default and never opens a network port at all. It only talks through stdin/stdout, a local text pipe to the program on your own machine. So it never makes a network connection at all. Once an agent starts it, the agent gets the same tools a person has in the app. Search and pull out text, flip through and reorder pages, reply to comments. Fill form fields, password-protect, merge and split documents, undo and save. To try it, open the repository on GitHub. Ask your agent to set up PdfCraft, or build its pdfcraft-cli (command-line tool), following the README (the project's instructions file). Then turn on its MCP server, pointed at a folder with a few of your PDFs. Have the agent find a piece of text and list the form fields in one of them.

storytold/lightcraft

LightCraft is a free, open-source app for organizing and developing photos. A team called ArtCraft is building it from scratch in Rust, a programming language. It contains none of Adobe's original code. The authors call it a clean-room rebuild of Adobe Lightroom, the well-known photo-editing program. LightCraft runs natively on macOS, Windows and Linux. It also runs in the browser, using WebAssembly, a way to run native-speed code on web pages. There's no account, no cloud, and no telemetry, meaning no usage data gets sent back to the makers. It needs no subscription either. Your photo catalog is stored as a plain, readable log of edits, so you can compare or roll back changes line by line. On a 24-megapixel RAW photo — a camera's unprocessed image file — tested on an Apple M4 Pro chip, moving a slider redraws the image in about 4 milliseconds. A quick preview takes about 30 milliseconds, and a full export takes about 0.3 seconds. Without a graphics card, the same work runs on the regular processor instead, just slower. Every slider, brush stroke and menu click in LightCraft is its own named command. LightCraft also ships an MCP server — a connector that lets an AI agent send it those same commands. You start it with lightcraft-cli mcp. Through this server, an agent can pick out a photo. It can develop the RAW file, adjusting exposure and light. It can paint a mask over the sky. Then it can export the finished photo. The agent can do all this in the background. It can also do it inside the open app, clicking buttons and taking screenshots as it goes. The screenshots in LightCraft's own documentation were made exactly this way. By the authors' own count, LightCraft now covers about 79% of Lightroom's total features. For the core features only, that number is 98%. As an everyday replacement for daily editing work, the authors put it closer to 60-70%. It works best with JPEG and DNG files, and with most Nikon, Sony and older Canon RAW files. One honest limit: sky and subject masking look like an AI feature. In fact, it runs on classic computer-vision heuristics — fixed, hand-coded rules, not a trained AI model. LightCraft also has no AI-based noise reduction yet. It has no HDR either, which stands for high dynamic range, a way to show both bright and dark detail at once. The authors' roadmap lists a future option to plug in Meta's SAM 3 model for some mask types. That model would be a separate, user-approved download. It is not part of the app by default. To try it yourself: install LightCraft from github.com/storytold/lightcraft, following its README, a file that explains setup. The simplest route is the installer at getartcraft.com/apps/lightcraft. Import a test folder of RAW or JPEG photos, start the MCP server, and ask your agent to pick one photo, develop it, mask the sky, and export it.

storytold/deckcraft

DeckCraft is a project by ArtCraft, a small studio that makes open creative tools. It rewrites PowerPoint, Microsoft's presentation software, from scratch in Rust, a programming language, with no Microsoft code inside — the team calls it a clean-room clone, meaning it was built without looking at Microsoft's own code. Today it already has slide masters, eleven layout templates, eight themes, over 150 ready shapes, tables, charts, basic SmartArt (ready-made diagram shapes), transitions with a Morph effect, and animation by paragraph. You work in it with your mouse and menus, the same way you work in PowerPoint. A finished deck opens and saves as a real .pptx file, PowerPoint's own file format, or exports to PDF, PNG or JPEG. By the team's own estimate, DeckCraft covers about 79% of PowerPoint's features by breadth and 93% of the core ones, and is roughly 80% of the way to its first alpha version — a working draft, not a finished product. Every action in the program is also a command with undo, and there are over 200 of them. That means an AI agent can call these same commands too, through MCP (a protocol that lets AI agents control apps) or a JSON control channel, building slides, tables and charts on its own. The project appeared only a few days ago, so don't expect stability yet, but ready installers for Windows, macOS, Linux and FreeBSD are already in the Releases section on GitHub, so you can install it without touching any source code.

storytold/effectcraft

EffectCraft rebuilds Adobe After Effects from scratch. Adobe makes After Effects, the standard program for video effects and animation. EffectCraft is written in Rust, a programming language built for speed and safety. It comes from ArtCraft, a team with similar free, open programs: PhotoCraft for photos and FilmCraft for video. The first commit landed on October 1. Within a week the project had over 1,600 stars on GitHub, a way people bookmark and vote for projects. Inside it has layers, keyframes, 3D cameras and lights, and 306 effects. A keyframe sets a property's value at one exact moment; the program fills in the motion between keyframes on its own. The 306 effects cover all 298 effects from After Effects, plus more. You can export animations as Lottie, a format that drops straight onto a website with no plugin. Rendering — turning your project into a finished video file — skips FFmpeg, the usual open tool for this. EffectCraft wrote its own video encoders in pure Rust instead, for H.264, ProRes, HEVC and AV1. Project files are plain JSON, a readable text format. You can open one as text and see its edit history in Git, the tool that tracks file changes. Two optional features borrow from outside: cutting out objects (Roto Brush) can use MobileSAM, a small recognition model built on Meta AI's Segment Anything and Microsoft's TinyViT. Face tracking can use Google's Face Landmarker. The makers themselves say it isn't ready to replace After Effects on paid client work yet. Nobody has checked frame-by-frame if it matches the original, and Linux and Windows still have small interface bugs. To install: grab the installer for your system — Mac, Windows or Linux — from the project's Releases page on GitHub, or just open it in your browser with no install needed.

remotion-dev/skills

This is a pack of skills — instructions an AI coding agent reads before acting — for building videos with Remotion, a tool that makes videos out of code instead of a video editor. One skill starts a new video project or scene from a description, another covers layout: animation, text, images, sound, fonts, timing. Others trigger a render into an actual video file, or build maps with animated routes using Mapbox, MapLibre, or CesiumJS. Install with npx skills add remotion-dev/skills in a terminal, or pick skills while creating a project with bun create video.

multica-ai/andrej-karpathy-skills

This is a CLAUDE.md file — a settings file for Claude Code, Anthropic's AI coding tool. It's built from Andrej Karpathy's notes on common AI coding mistakes: guessing at your intent, hiding confusion, bloating code, and touching things it doesn't understand. It sets four rules — say assumptions out loud, write minimal code, change only what's asked, and check results with a test. It also works in Cursor, another AI coding tool. Install it by asking your agent to run /plugin marketplace add, then /plugin install, or by downloading the CLAUDE.md file into your project with curl.

666ghj/MiroFish

MiroFish is an engine built from thousands of AI agents — bots with their own memory and personality. You give it a real-world seed, like a news story, a policy draft, or a novel, and it builds a parallel digital world where those agents interact. You describe in plain language what you want predicted, and get back a forecast report plus an interactive digital world you can question. To run it, you need Node.js 18+, Python 3.11-3.12, and uv, a Python package manager. You get a paid API key for an LLM (the README suggests Qwen-plus, Alibaba's model) and a key from Zep, a memory-storage service — Zep's free monthly quota is enough for simple use. Put both in a .env file, run npm run setup:all, then npm run dev, or use Docker, a tool for running ready-made setups, with docker compose up -d.

The author describes the Space in one line: 'Play chess against the complete connectome of an adult fruit fly' — connectome means the complete wiring map of every neuron and connection in a brain. It's built on the FlyWire brain reconstruction, a brain-mapping project: 138,639 neurons and 15.1 million connections, taken from a real fly. Every move on the board activates 10,855 of the fly's visual neurons, the signal spreads across the whole network over five steps, and the move plus a win probability are read out from neurons in the fly's central brain. The neuron wiring was never altered — it matches the biological reconstruction exactly, including the sign of each connection. What got trained was only the strength of each connection, each neuron's individual sensitivity and firing threshold, and the final readout layer — on Lichess games rated by Stockfish, a chess engine. On top of that there's a search that looks three moves ahead, where one full level of that search is a single combined pass through the brain. The model and its weights are about 150 megabytes, downloaded once and then served from the browser's cache. To try it: open the Space page in Chrome or Edge, which support WebGPU, a browser feature for running AI.

AutoTrust, the AI lab behind it, describes JEV-27B-VL as 'JEV-27B that can see — the same System 1 and System 2 decisions, now on images.' System 1 is the fast part: give it text, up to eight pictures, and a question — yes/no, a score 0 to 5, or pick one from up to 256 options — and in one pass it returns a calibrated probability for each option, no paragraph of text. It's built on Qwen3.8-27B (an existing vision-language model); the vision part is unchanged, only the new decision head was trained. On a public ranking of 20 such vision-decision models it's first, ahead of Qwen3.5-397B, a model about 15 times bigger. On the independent VL-RewardBench test (CVPR 2025) it scores 78.3% accuracy, above GPT-4o and Claude 3.5 Sonnet — though that's not quite a fair fight, since those are older versions, not the current ones. To run it: ask your agent to set it up locally via vLLM or transformers (tools for running models), feed it a picture with a yes/no question, and show you the probability it returns.

GEV-26B-Decide is a small AI model made by AutoTrust, an AI lab. It answers a typed question with a calibrated probability for each option. Calibrated means the probability matches how often it's actually right. When reasoning is turned on, it only thinks step by step when it isn't sure. It's built on Gemma-4-26B-A4B, a Google AI model with 26 billion parameters, about 4 billion active per token. A normal answer takes about 45 milliseconds. To try it, go to its page on Hugging Face, a site that hosts AI models for download. There's no ready-made package for Ollama, a tool that runs models locally in one click. Instead, run it through vLLM or transformers, two developer tools for running AI models. You'll need a Mac with at least 32GB of memory, or a graphics card with 24GB.

Aleph Alpha (aleph-alpha on Hugging Face, a site that hosts AI models), a German research lab, calls Kolibri its reasoning model, built on a mixture-of-experts design with a focus on German and English. It has 78 billion parameters total, but only 3.46 billion switch on per word processed — that's what keeps running costs lower than the size suggests. Its context window, how much text it can read at once, is tested up to 1,048,576 tokens, though the makers recommend staying under 262,144 for speed on everyday tasks. It can run in an explicit step-by-step reasoning mode and call outside tools through an API (a way programs talk to the model) that matches OpenAI's format. The makers are upfront: in systems that help make decisions, it should sit on the advisory side, offering options and showing its reasoning, not act unchecked. License is Apache 2.0, free to use and change; released October 3, 2026. Aleph Alpha also signed the EU's code of practice for general-purpose AI — this is Europe's own bid for a big open model, not a copy of US or Chinese ones. If you have a server with the right cards, ask your agent to deploy Kolibri-1 via vLLM (a tool for running models) from its Hugging Face page, and test it in German and English with reasoning and tool-calling on.

storytold/photocraft

PhotoCraft is a free, open-source photo editing program. To try it, ask your coding agent to install it from https://github.com/storytold/photocraft: build it from source, or grab a ready build following the README, the project's instructions page. Then run a couple of test photos through photocraft-cli, the program's command-line tool, applying the same edit to each. It's written from scratch in Rust, a programming language, with none of Photoshop's own code inside. The authors call it a clean-room reimplementation of Photoshop, meaning it's rebuilt without copying any original code. It has layers, masks, brushes, vector shapes, and real PSD files, Photoshop's own format for layered images. It even uses Photoshop's keyboard shortcuts, so you won't need to relearn anything. Every action in the program is one of more than 500 shared commands. You can run these commands from the command line, typing instead of clicking, or hand them to an AI agent. That means one command can apply the same edit to a whole folder of photos at once. Re-saving a PSD file keeps the original look in 307 of 309 files from a test set called psd-tools. All processing, including smart area selection, runs locally on your computer, with no cloud and no account. The authors admit it's early alpha, not yet a replacement for Photoshop in daily professional work: it's missing generative AI features, around twenty tools, and enough depth in typography. macOS builds are signed and notarized by Apple, like any other Mac app.

rauchg/gdp-ts

gdp-ts is a tiny library and a skill for TypeScript projects. A skill here is a ready-made instruction pack that teaches your coding agent a specific method. It's built around an idea called Ghosts of Departed Proofs, which turns permission checks into part of the type system itself, not just a line of code buried inside a function. The author is Guillermo Rauch, founder of Vercel, a web hosting and deployment company. The idea came out of talks about security in their huge TypeScript codebase, where hundreds of engineers and their AI agents write code. Here's how it works: a function that needs protecting — say, turn off password protection on a project — now needs more than a user ID. It needs a ready-made proof that the right check already happened somewhere else in the code. If there's no proof, or it was made for a different user or a different resource, the code simply won't compile. It fails before it ever runs. You install the skill by telling your coding agent to run the command npx skills add rauchg/gdp-ts. It teaches the agent how to set up such a proof, where to get it, and how to turn on an automatic checker that scans the code so nobody can fake a proof. The authors say this doesn't catch fake or outdated proofs — that's a limit of TypeScript itself, not a guarantee against every hole.

Graphify-Labs/graphify

This is a tool that builds a knowledge graph. That's a map of how ideas, code and files connect across your whole project. It works inside Claude Code, Anthropic's AI coding tool. Code gets parsed locally; documents and images get read by Claude, the AI behind it. Install it with pip install graphifyy, then run graphify install, then type /graphify inside Claude Code.

bkingfilm/lapian-notes

Lapian Notes is a tool for studying a movie frame by frame. It cuts the video into frames about one second apart. It matches subtitles to them, then bundles it all into a package for an AI chatbot like ChatGPT. After you get the AI's result back, the tool builds story lines on separate tracks. It also draws a scene structure tree and a curve showing how engaged the audience is at each point. To install: download the zip from the project's Releases page on GitHub, a site that hosts code. Unzip it, then double-click run.bat on Windows or run.command on Mac — no coding needed.

alchaincyf/huashu-art-motion

huashu-art-motion is a skill — a set of instructions and code — for a coding agent. It turns a description of a video into real animation code, drawn with code instead of sliding pictures. You install it by running `npx skills add alchaincyf/huashu-art-motion` in your coding agent. It needs just a normal laptop, no graphics card. It comes with 35 art-style cards, from cave paintings to Van Gogh, Bauhaus, and pixel art. Each card has a ready scene built in Canvas, a code tool for drawing pictures. It also has a signature transition and known weak spots. There are also 9 grammars for narrated explainer videos — pop-science style, whiteboard drawing, finance charts, and more. Eight of these nine ship with demo clips and a JSON spec, a file that describes each shot. The agent follows that file frame by frame. The ninth, a presenter-led finance format, is only reference code; it needs your own character images. There's also a breakdown script. It takes someone else's reference video and turns it into a shot list, a rhythm grid, and a motion heatmap. You can then recreate that in code. The skill doesn't draw character frames itself. That goes to a separate image-generation model; the code only handles assembly, frame switching, and backgrounds. Before a video ships, an automatic check runs: one script measures stability, motion, and smoothness. Then a separate agent, one that didn't see the production, watches the finished video and flags issues.

I tried giving the model one invoice and two questions at the same time: whether it is paid or overdue, and whether the amount is above one thousand dollars. Clef-Flash from Cloudflare is a 9 billion parameter model that reads the record once and answers all the questions at once, instead of writing a paragraph that I would have to read through myself. On the test invoice for 1250 dollars with the overdue status, it gave ninety percent for the option overdue and only five percent for paid. I liked that it is compatible in format with the paid service Jev, and this version is open and free.

Last week I already tried this model and cut out a background in one request. In seven days the download count grew from 52804 to 90003: people keep installing this model. Qwen-Image-2.1 from Alibaba is a single model for drawing and editing pictures in the Qwen family, and in one request it accepts up to ten reference images, and an edit can be marked right on the picture with a circle or a mask, instead of a long text description. For the second week in a row this model holds a high place in the week's chart, and this does not look like a one-week accident.

I opened the ready demo space right in the browser and gave the model a floor plan with a question: which room is to the left of the kitchen, and it read the plan correctly. ZDTaichu5.0-9B from TaichuAI is a general 9 billion parameter model for understanding images, spatial tasks and acting as an agent, and by the vendor's own numbers it leads among models of about the same small size: 93.7 on the IFEval instruction test and 87.7 on the TAU2-Bench agentic test. The comparison here is only against other 9 billion parameter models, not against something at the level of Opus or Fable, which I use for real work every day. I liked that I could try it right in the browser, with no installation and no Ollama.

wy51ai/floorplan-3d

I opened the demo and found that the whole thing is one index.html file with no build step at all. You drag more than sixty pieces of furniture and appliances from a library onto a 2D floor plan, move walls, measure distances, and switch to a 3D scene with one click to walk through the apartment in first person. For each room you can swap the flooring material, and the tool estimates the renovation cost from the room area with an allowance for waste. The interface switches between Chinese and English, the plan autosaves in the browser, and you can export it as a PNG or a JSON file.

nykooi1/vibe-wise

I liked the idea itself. It is a Claude Code plugin that asks how you would design something before it writes any code. The README shows a notes app example: Claude asks what should happen to a note that sits in two folders when one of those folders gets deleted, you answer, it proposes a design through a links table, and it waits for your confirmation. The author writes that this is for anyone who wants to learn while building, from a beginner to an experienced engineer trying an unfamiliar stack, and there are three kinds of checkpoints, for approach, design, and the code itself. The README also notes that the plugin has already been approved for Anthropic's official Claude Code directory, even though it is not yet listed in the public marketplace.

whirlchat/whirl

This is an open AI chat app, and it already has a ready hosted version at whirl.chat, so you do not have to run it yourself, even though the full code is open under MIT. Through OpenRouter, one chat gives you access to every top model at once. What I liked is that it remembers your preferences and projects across separate conversations instead of forgetting everything in a new chat. The README separately describes 'living documents', editable pages and charts that stay next to the conversation in a side panel and can be shared by a link, and you can connect your own tools to it through MCP.

TailAdmin/free-nextjs-admin-dashboard

This is a free template called TailAdmin, built with Next.js and Tailwind CSS - tools for building and styling websites. It comes with ready-made pieces: tables, charts, login screens, a sidebar menu, profile pages, and dark mode. To install: ask your agent to clone the repo, install Node.js 20+, then run npm install and npm run dev.

Robbyant/lingbot-vla

This is a ready-made model for controlling a robot arm. It takes a camera image and a text command, and outputs arm movements. To install it, set up a Python 3.12 environment with conda, a tool for managing Python setups. Then clone the repo and run bash install.sh.

littledivy/mimic

mimic is a project on GitHub, a code-sharing site. It records an app's traffic — a token, device ids, session and cookies — once, then an AI writes a Python client from that recording: a ready file with named methods and step-by-step request chains. Install it by running sh install.sh, which sets up the needed tools and starts capturing traffic on first use.

dzhng/skills

This is a library of skills — ready-made instructions — for AI agents like Claude Code and Codex. Each skill breaks a task into checkable pieces, writes a plan from them, then builds unattended. Install it with npx skills add dzhng/skills in a terminal, on any computer that runs Claude Code or Codex.

Xing4.0 is an agent AI model — software that plans steps and uses tools on its own. It's built by China Telecom, China's big phone carrier, not an AI lab. They trained it only on Huawei chips, a Chinese chipmaker's hardware. They used zero Nvidia GPUs, the chips most AI normally trains on, and put the weights online in September. Inside, it uses a Mixture-of-Experts setup, a design that splits work among specialist sub-networks. It has 29 billion settings total, but only turns on 4 billion of them for each answer. That keeps it fast on normal hardware. Its context window — how much text it can read at once — is 256,000 tokens, or chunks of text. It can stretch to 512,000. Until now you could only run it on a server. A team called Venastine-Research took those same server weights. They converted them into GGUF files, a format for running AI locally. Apps like Ollama, llama.cpp and LM Studio can open these files. So now you can run it on your own computer. To install it, ask your agent to grab the GGUF build from Venastine-Research's page. You'll need a Mac or PC with at least 24GB of memory for the smallest version. 32GB is safer, since it fits the more accurate file. Have it pick the file size that fits your computer's memory. Then connect it through Ollama or LM Studio. Ask it to show you how to use it in the terminal, a text command window.

tt-a1i/archify

Archify is a skill for Claude Code, Cursor, Codex CLI, and OpenCode. To use it, simply type npx skills add tt-a1i/archify -g in your terminal and start using it within your chat interface! Simply explain your system with simple language, or have the agent inspect your repo, and let Archify determine the best block structure, connect the blocks with arrows, and generate an HTML file (self-contained web page) containing your diagram. This HTML file is fully interactive! Click on blocks, highlight a path between two blocks, toggle between light and dark mode, share an image or route card. Archify currently offers 5 different kinds of diagrams: Architecture, Workflow, Sequence, Data Flow, Lifecycle. Depending on your question, the agent will choose the most suitable diagram type. Once generated, this HTML file is fully self-contained, meaning others can view it by simply opening the HTML file in their browser without needing to install Archify. Early September, Archify hit #1 on GitHub’s weekly global trending repo (all languages) and the developer was featured on the Chinese news site QbitAI. As the developers note: “The diagram shows what the agent learned and reported, but it does not validate the runtime behavior of the code.”

charlesdove977/claude-x-jev

Jev is described by its makers as a tiny AI that makes one decision at a time - in about a third of a second, for a fraction of a cent - and tells you how sure it is. It isn't a chat model: you don't write it a prompt and wait for a paragraph back; you hand it a piece of text plus a fixed list of options, and it hands back a pick, a probability for each option, and a confidence score. Claude x Jev is a free Claude Code skill that puts this model in front of Claude: it can sort incoming emails, leads, or comments, or check a draft reply, and Claude only steps in where Jev itself isn't sure. To install: ask your agent to run `npx claude-x-jev install --with-commands` from the repo at github.com/charlesdove977/claude-x-jev, set up an OpenRouter key, then try /jev-classify on a few of your own emails.

BraveOPotato/FckSignups

NoSignups is a directory of zero setup web apps. This is an open source repository of curated open source tools which you can use without signing up or giving away any information. Built using React and TypeScript (building blocks for web apps). It currently contains over 200 tools divided into various sections such as Productivity, Design & Graphics, Development, Writing, Privacy, Utilities, Data & Analytics, Media, Education etc. Every tool has its name, description, url, category, tags. Some also have their GitHub repositories (code page) and licenses. There are also ‘featured’ tools which are unique compared to others. This repository is licensed under GPL-3.0 license. All the listed tools are licensed under their respective licenses. We don’t own these tools. This repo was created in May 2026. It already has nearly 4,400 stars on GitHub. An update was done recently. You can open this in your browser directly by clicking on the link. There is no downloading or running involved. If you wish to have it on your machine, then ask your agent to deploy the NoSignups catalog from https://github.com/BraveOPotato/FckSignups. Install dependencies and run it as mentioned in the README (instructions file) by using npm install and npm run dev. Then you will be able to see it in the browser. No keys, access required. You can do this even with your normal laptop and Node.js.

spinabot/brigade

Brigade is a framework to build your “personal enterprise intelligence”, described by its creator as “a team of agents structured like a company, complete with long term memory (Tideline), decaying facts/deduplication, agents with different credentials and interests, and rules for communicating between them”. It comes with 56 built-in skills and is trivial to connect to Gmail, Notion, Calendars, and thousands of other services using a single integration key. Your agents can open PDFs, Word documents, Excel spreadsheets, and images that you share with them, and they can return documents and reports back to you. You can interact with your agents using the command line, WhatsApp, Telegram, Slack, Discord, iMessage, or BlueBubbles. You can choose between Claude, OpenAI, Gemini, OpenRouter, Groq, Mistral, DeepSeek, local Ollama, or any other OpenAI compatible server (including mixing and matching models in the middle of a conversation!). Developed by Spinabot, licensed under the MIT license. To get started, simply ask your agent to install Brigade from github.com/spinabot/brigade. After following the instructions in the README to install Brigade using your Claude Code credentials (or inputting the API key the installer asks for), creating a Telegram bot, and confirming that your agent answers correctly when asked questions and given facts.

PostHog/jeeves

This is a reasoning classifier like Jev, but with a ‘draft’ diffusion stage (quick first draft answer), trained using SFT and CISPO training techniques. This follows on from previous ideas of using a small model to return probabilities for questions directly (what department does this belong to? Should we escalate? How angry is this customer?). Jeeves has the added benefit of being able to think about its answer first, according to its creators, taking ~0.3 seconds without thinking, and a median time of 3.3 seconds per question on a single H100 GPU when thinking. According to its creators, Jeeves performs better than both Kev-9B and Jev on unseen data (0.889 compared to 0.822 and 0.857 respectively) and on the publicly available tiers of the JevBench benchmark (0.935 compared to 0.866 for Jev). Note that these scores for Kev and Jev are self reported by the projects themselves, and so should be taken with a pinch of salt as they may not be completely trustworthy. Created by Product Analytics company PostHog, which was inspired by the open source Kev project.

openai/mcp-extensions

Recently, OpenAI released several improvements to the MCP protocol that enable third party services to be integrated directly into the ChatGPT UI (not just return text). The new features include side bars, custom file types, mentions (@), and image-based UI forms. This project is an officially sanctioned OpenAI effort based on the MCP protocol and MCP app SDK, but it is not a replacement for either of those things. There are SDKs available for both TypeScript and Python that you can install as a dependency on your current MCP server. Included is a demo parts catalog plugin that demonstrates how to use all 4 kinds of extensions.

oil-oil/beautify-github-readme

beautify-github-readme is a skill for your AI code agent that generates a new version of the README file of a GitHub repository (which is the main entry point when someone visits your repository) to make it easier to grasp the overall idea of your project. It analyzes your repository, discovers the most relevant parts of it, and then figures out how to present them. For instance, the layout for a CLI tool will be different than for an icon pack. It comes in two flavors: ‘whole README’ (reorders content and changes the overall appearance of the page, such as using your own header image) and ‘asset-only’ (doesn’t change your text, content order or links, but adds some images like a header, subheadings, a process flow chart, a badge or a small animated GIF). These images are editable SVG files (which means they are vector graphics and you can edit them further) and the text is kept in Markdown (which is the markup language used for writing README files). So you get a searchable and editable web page rather than just a static image. You can install this skill by telling your AI code agent: ‘Install beautify-github-readme from https://github.com/oil-oil/beautify-github-readme via npx skills add oil-oil/beautify-github-readme, use it in asset-only mode for my repository and show me a local preview first, do not publish the results without asking me!’

anthropics/claude-code

Claude Code from Anthropic is an “agentic” programming assistant for your terminal. It can answer questions about code, parse project files, automate repetitive tasks, explain code, handle git (commit, branch, merge), and more. Install the official install script for macos/linux (homebrew), winget (windows), or npm install (deprecated, see readme). Once installed, go into your project folder and run claude. You can also install plugins that give Claude Code extra capabilities and subagents. See the plugins folder for more information. Claude Code can also be integrated with an IDE or used via github comments (@claude)

Offline deployment of Swift 1.5, which was developed by startup UkisAI using Alibaba’s open source model Qwen3.8-27B. We fine-tuned it to be more impulsive when responding, theoretically reducing 58.5% tokens in thinking process, better performance than Qwen, and sometimes up to 9.18x faster than Qwen. It’s used for multi-step programming/agent tasks. We deploy its offline version at various compression ratios with smaller storage space, memory usage and slightly worse accuracy. You can download and run this model locally by telling your agent to “install this model locally” https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF, and select a suitable compressed version based on your memory size, then use llama.cpp/Ollama to query this model, and here’s how to use it: This model is completely free, there’s no subscription and no server required.

If you work with podcasts or videos, you may have heard of Pyannote and their open source speech processing models. They just published a new Community-1 model that can automatically transcribe audio into speaker diarization (i.e., who spoke when) without requiring any text transcription. Compared to their previous 3.1 model, the new Community-1 model shows a marked improvement in performance, lowering the error rate from 12.2% to 11.7% on the AISHELL-4 dataset and from 24.5% to 20.3% on the AliMeeting meeting recordings dataset. You can also supply the number of speakers in your audio if you know it. And best of all, this model runs entirely offline, so no internet required! Model weights are freely available under a CC-BY-4.0 license, but the model is currently gated, so you’ll have to go to their website, fill out a short form, and get a Hugging Face token to access the model.

Qwen has released their Qwen3.8-Flash-Next model, which consists of 176.9 billion parameters and was trained by multiple internal experts for serving purposes (it’s 354 gigabytes uncompressed). But a team from ISTA-DASLab at the Institute of Science and Technology Austria (who have been working on model compression techniques for years) has also developed a specialized version of the model for code. They completely pruned away half of the experts, and requantized the remaining ones to a lower bitwidth. The resulting GGUF file is just 58.4 gigabytes, which is about six times smaller than the full model and can run on a single GPU or even a 32GB mac! Admittedly, there’s a pretty big tradeoff in terms of performance; on the SWE-bench Verified benchmark for long agentic tasks on real world codebases, they’re able to keep 91.3% of the original model’s performance. On the LiveCodeBench benchmark for shorter and simpler tasks, they’re able to keep 98.7% of the original model’s performance, so you probably won’t notice any difference for most coding tasks. Note that the model is still able to handle both code and image inputs after expert pruning.

Krea 2 is an open-source image generation model built by Krea AI, creators of a web-based image editor. It comes in two versions — one for those who want to customize their own styles, and another for quickly generating final images. Until now, the primary ways to install this model were through the mac apps Draw Things and DiffusionBee. But thanks to the team behind the popular local image generator ComfyUI, there’s now a package available with both models, along with some pre-built templates and style packs, so you can install it on Windows or Linux machines with any GPU — not just Mac. You’ll need at least 12GB of GPU memory. Please ask your agent to install the krea 2 model onto ComfyUI (huggingface.co/Comfy-Org/Krea-2), choose a lightweight version suitable for your GPU, set up a template, and show me how to use this model to generate images.

Nemotron 3 Diarization is NVIDIA’s third iteration of their diarization (speech segmentation) model. This means that the model will be able to tell you which speaker spoke at what time, without having to transcribe what they said. It can differentiate between up to 8 simultaneous speakers, and can process either live audio or pre-recorded audio. You can install this model by saying “Install the NVIDIA Nemotron 3 Diarization model locally: https://huggingface.co/nvidia/Nemotron-3-Diarization. Run voice tagging using NeMo-Speech.cpp (lightweight local engine) on a multi-speaker test recording, and tell me who said what.”. You can use this on any laptop with 8GB of RAM or more. The model itself is only ~100MB, so you don’t even need a GPU!

CopilotKit/openmuse

OpenMuse is a personal assistant that uses its own browser, terminal and filesystem to get things done. Not only does it provide guidance, but it takes action too! OpenMuse consists of a server, a task worker, and a browser worker. You can self-host or deploy to Render with a single click. In this example, our agent reads an email regarding a class trip, selects appropriate exhibits at an aquarium, requests the necessary form fields, and generates a filled-out version of the PDF included in the email. Finally, it drafts a response for review. For mobile, our agent can select relevant news articles from Hacker News and summarize them briefly. There’s even a web diffing module (to detect changes on websites, such as price changes) and a module that imports a CSV file (a spreadsheet-like data file) containing financial transactions and generates a summary of expenses. The terminal is optional and runs in a Docker container in an isolated environment on Linux (it has no access to the host filesystem and can only connect to the browser worker).

TeleOCR is a compact model for parsing documents, built by the XingChen-AGI team - it used to be called NaviDC-OCR. It's trained to read more than clean digital scans: it also handles photos of documents shot at an angle or skewed, without a separate page-straightening program. TeleOCR picks up not just the text but the structure too - tables, formulas, and the reading order of multi-column pages. At around 1.2-1.4 billion parameters, it scored 96.87 on the OmniDocBench v1.6 test. That's higher than much bigger general-purpose models, including Gemini 3 Pro and GPT-5.2. The model is built for one job only - it doesn't hold a conversation or write code, it just takes in a document and hands back marked-up text.

I ran it on a live microphone stream and left it working without stopping it. Audio8 ASR Infinite from the Edge0 team is a speech recognition model that listens continuously, and it has no limit on recording length because its memory and delay do not grow no matter how many hours it keeps running. It can tell a thinking pause apart from stuttering and from a real end of a sentence, so it does not cut a person off in the middle of a word. On the English LibriSpeech test it makes about one error per thirty three words, and on the Chinese AISHELL-1 test it is under two errors per hundred.

I installed it through vLLM and gave it a task to read someone else's repository and find a bug from a ticket description. Xing4.0-29B-A4B from China Telecom is a model with 29 billion parameters, of which only 4 billion activate per token, so it is noticeably lighter to run than a typical model of this size. Its context is 256 thousand tokens natively and extends to 512 thousand, so a whole small project fits in at once. On the SWE-bench Verified benchmark, where a model fixes real bugs in open repositories, it scores 75 out of 100.

A week after I first installed it, I checked the download counter and it had grown by almost one and a half million, past three million in the past month. Ternary-Bonsai-2 from the Prism ML team is the same 27 billion parameter reasoning model that Qwen built, only the weights are recalculated so the file takes 5.95 gigabytes instead of 54. Across fourteen reasoning benchmarks it holds 98.2 percent of the full version's result, and on math and coding the difference is within the margin of error. I ran it through llama.cpp on my own Mac and got about 47 tokens per second, which is faster than I can read the answer as it streams in.

Parakeet Redux is a smaller version of NVIDIA’s speech recognition model parakeet-tdt-0.6b-v3 developed by Moondream team. It has the same architecture and vocabulary, but the weights of the encoder (the sound processing part) are quantized to [-1,0,+1] values only, making it 178 MB instead of 1.2GB. This model supports 25 languages, including Russian, and allows to get the transcript with time stamps for each sentence and word. The authors achieved 113x real time performance on 8 CPU cores (no GPU required) and you can request your agent to install Parakeet Redux by Moondream locally using moondream library, transcribe some test audio file and display the output with time stamps.

tobi/disktree

My Mac was telling me that I’m running out of disk space, but I don’t know where it goes. So I installed disktree to quickly visualize disk usage. It’s a simple treemap visualizer showing files and directories as colored squares proportional to their size. With it, you can easily spot large unnecessary files. After clicking around in some big directories, marking unwanted files, and checking the final list, I safely remove files. It’s a single binary without any services running in the background. It supports both Mac, Windows, and Linux.

yetone/magpie

I have three coding agents installed on my computer, and they all needed their own method of selecting a model. Magpie shows a menu bar with all available agents. Now I can easily switch between Claude Code and Kimi or Codex and DeepSeek with a single click instead of editing a configuration file. It’s also available as a standalone command line interface without the desktop application. Supported providers include Anthropic, OpenAI, DeepSeek, Kimi, GLM, Ollama, and many others. This allows using local or open source models alongside commercial options.

mikehasa/golive-skill

A coding agent can create a functional website or bot within hours, but then comes the hard part — setting up hosting, databases, domains, emails, payments, etc., manually via a dozen web interfaces. golive-skill is a skill for Claude Code or Codex that analyzes the completed project, determines required infrastructure, and displays a proposal without making any changes. Once approved, it creates hosting on Vercel or Netlify, a database on Supabase or Neon, and payment processing on Stripe using my accounts. There’s no golive account, backend, or telemetry. Everything happens in the terminal.

852wa/JIZURA

JIZURA is a browser-based tool that I launched without installation. Simply paste lyrics, play music, and it automatically detects tempo and beats. Then it generates a lyric video by stitching together premade clips and transitions like those found in textual music videos. Finally, it exports the video as an MP4 with audio or frame numbers for further editing. Although the user interface is in Japanese, button placement makes it easy to understand even without knowing the language. Also, there’s no uploading involved because it runs entirely on the client side within the browser.

dgreenheck/tidewater

Tidewater is a browser-based fishing game where you can fish from a pier, beach, or personal boat, pull in fish, sell them to Joe at his stand, and invest in better equipment from Marta. What intrigued me was that the developer created everything — water, sky, island — using a custom WebGPU engine implemented from scratch and in collaboration with Opus 5.5 instead of using an existing game engine. It requires a browser with WebGPU support and nothing else. To me, it’s not so much a game to play, but a demonstration of what one person can achieve with an agent over the weekend.

trailhq/Graft

Graft calls itself an open-source context layer for large codebases. It walks a repository once with tree-sitter, groups files into nodes by subsystem, and writes a plain-language explanation for each node plus typed links to the nodes around it. Everything lands as ordinary markdown files in a local folder next to the project, no database and no server. From there, a coding agent reads the node instead of grepping the repo blind, either through a bundled MCP server or by opening the files directly. On SWE-bench Verified, the official benchmark that grades real GitHub issues with the maintainers' own tests, an agent using Graft resolved 33 of 50 instances against 27 for a cold agent, with 23% fewer tokens and 32% less wall-clock time. The structural parsing is deterministic and needs no model key at all; only the richer, plain-language notes call a model, and for that you plug in your own key from any provider (Anthropic, OpenAI, OpenRouter, or a local one). Built by the team behind Trail (trailhq.com), and published on npm as @nanonets/graft.

GLiNER2.5-Decide, from the company Fastino, is a 340M classifier: you pass it any set of labels at call time — customer intent, ticket routing, sentiment, priority, spam or not — and it scores several of these at once in a single pass, without a prompt template and without fine-tuning per task. On the company's own benchmark, a suite of 17 domains with 300 held-out examples each, it scores 60.2% exact-match accuracy, ahead of the much larger Qwen3.5-4B at 56.4% — a model roughly twelve times its size. The authors are explicit about its limits: it is not a general model, it does not reason or explain, it only returns the labels you asked for. It runs on CPU or GPU through the small `gliner2` Python package.

Hemmingway-1 is a 27B chat model fine tuned by Altworld (an indie studio) on top of Qwen3.8-27B. This model is optimized for common tasks such as email writing or quick responses to colleagues. According to the creators, this model is more human sounding and performs better than other larger models on their internal tests, however those results are not independently verified and should be taken with a grain of salt. With a large context size of 262,144 tokens, this model is able to process large amounts of text at once, such as entire chat logs. The weights are available for public download and can be used offline for non-commercial purposes. Prebuilt android and mac applications are also available.

latent-spaces/brag

The team at latent-spaces built /brag as a skill for Claude Code: one command turns a finished project into a short launch video with music, motion, and ready post text. The classic /brag hands the job to a separate paid service, Hyperframes by HeyGen, which builds and renders the video. The newer /brag-slim, made for Opus 5.5, skips that — the agent writes the script, style, and animation itself, and a tool called ffmpeg assembles the file right on your computer. It installs as a skill or plugin from a marketplace inside Claude Code, and the same skill format also works in Codex CLI, opencode, Cursor, Copilot, and Gemini CLI. The project is open source under the MIT license, has 7 authors and almost 10,000 stars in three months, and it's still actively developed — /brag-slim was added only for the fresh Opus 5.5 release. To try it, tell your agent: 'Install the /brag skill (brag-slim variant, no Hyperframes) from https://github.com/latent-spaces/brag in Claude Code, run it on my latest project, and show me the finished video from brag-output/brag.mp4.' The result lands in a brag-output folder with the plan, a brief, and the video itself, brag.mp4.

MiMo-V2.6-Pro-RL is, in the XiaomiMiMo team's own words, the 'flagship model of the MiMo V2.6 line' (XiaomiMiMo is Xiaomi's AI division). It's the biggest model in the line: 1.02 trillion parameters total, with about 42 billion active per request — this setup is called a mixture of experts, meaning only part of the model switches on for each request. It can act as an agent: working step by step in a terminal, fixing bugs in real code, reading screenshots of an interface, and holding a context of a million tokens (how much text it can keep in memory at once). Compared to the line's previous flagship, V2.5 Pro, the jump is clearest on agent tasks: a test for acting independently on a computer, AutomationBench, rose from 16 to 53.1 out of 100, and a real code-fix test, DeepSWE, rose from 19 to 71.9. Xiaomi already released two smaller versions of this line — Flash-RL and a smaller Distill-Qwen-9B — that can run on your own computer. Pro-RL is the top, most powerful version, built only for server hardware. To look at it yourself, ask your AI agent: 'Open https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL and dig into what's new in agent and coding tasks, and whether there's a way to try it via API or cloud service, since you can't run it locally.'

It is an acceleration extension module of Alibaba’s Qwen team image model Qwen-Image-2.1. Developed by Viggle, an AI artificial intelligence and video generation company, through distillation training to learn how to quickly imitate the results of the original model. Base model: 40 steps, acceleration model: 6 steps. Both generation (generate image according to description) and modification (modify picture with 1~3 reference images, such as maintaining certain faces or products) functions are supported, while maintaining a high degree of similarity with the base model in most situations. However, there may be certain differences in some scenes, such as small font texts within the image and complex modifications such as multiple reference image face swap. This model is based on the weights of the original Qwen-Image-2.1, so you still need to keep the original model when using it.

Qwen-Image-2.1 repackaged by unsloth as a small local file. This model was created by Alibaba’s Qwen team. It is an image generation and editing model based on textual input, generating images up to 2048x2048 pixels. Compared to the original, the repackaged file uses significantly less disk space and can easily be used in local applications like ComfyUI. The model generates images with transparent backgrounds, which can then be layered on top of other images without further editing in a photo editing program. You can also use up to ten reference images for editing, which will all be merged into a single scene. The original model was released a couple of days ago, and now it is available as a pre-packaged installable model for your local device.

This is a low-latency and high-accuracy true streaming Automatic Speech Recognition (ASR) model, named Confucius4-R2T2. This model is developed by NetEase Youdao, which is an online dictionary and translation tool in China. Confucius4-R2T2 processes the input speech frame by frame, each frame ranges from 80ms to 2s, and generates the corresponding text in real time. It does not wait until the sentence is completed or modify the generated text. The average latency of the model is between 200–600ms, and the performance is comparable to the offline model. Confucius4-R2T2 is based on Qwen3-ASR, a 1.7 billion parameter model from Alibaba, and fine-tuned by Youdao team. In addition to Chinese and English, it supports approximately a dozen other languages such as Russian, etc. Confucius4-R2T2 contains around 2 billion parameters and can be installed on a normal computer.

Authors define it as “an agentic model with 9 billion parameters, fine-tuned on data generated by MiMo”. The fine-tuning process was carried out by the team XiaomiMiMo (Xiaomi’s AI department). This is a downsized version of the MiMo V2.6 family. The flagship model of the MiMo V2.6 family has 309 billion parameters and cannot be run locally. The model can write and correct code, take incremental actions via terminal interfaces, recognize UI screenshots, and perform cybersecurity tasks. Their performance on the AutomationBench test suite (which measures automated task completion on a personal computer) increased from 5 to 30.3 out of 100 points compared to Qwen3.5-9B of equal size. The SWE Pro test suite (which evaluates the ability to repair a code bug) increased from 32 to 44.6 out of 100 points. Authors consider this release as a beta version, a rough draft to train the agents.

helloianneo/ian-xiaohei-illustrations

A Codex Skill which helps you to teach AIs to create illustrations for Chinese articles, blog posts, Notion pages, methodology articles etc. This skill takes an article, understands the key points of it (key judgement, thought change, before-after, metaphor), comes up with a plan (4–8 images based on these points) and generates the images using built-in Codex image generator. Each image is generated according to some template (horizontal 16:9 aspect ratio, pure white background, black line art, some red/orange/blue handwriting in Chinese, etc). The skill uses the same character (Xiaohei — black silhouette with two white dots and thin legs) in all images and there are certain rules for generating images (Xiaohei should always be doing something, not just be present in the image) The skill is NOT suitable for creating illustrations for business use, PPT infographics, cartoon icons etc.

This is, in the creators' words, the "efficiency-balanced checkpoint of the MiMo-V2.6 series" from XiaomiMiMo, Xiaomi's AI division. It handles all types of input — text, images, video, audio — in one request, and fits agent tasks like writing code, using a web browser, or reading documents from photos of them. Context length is 1 million tokens. It has 309 billion parameters total, but only about 15 billion are active at any moment. It's already available through Xiaomi's own MiMo Open Platform API and through OpenRouter, so you don't need your own server to try it — ask your agent to find the current API or demo for MiMo-V2.6-Flash-RL (huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) and show a sample request with an image.

StepAudio 3 Music Studio is a browser-based application by StepFun (the stepfun-ai team), a Chinese AI company that focuses on audio models. According to the application description, it allows users to 'create songs, arrange vocals, and remix music.' The user describes a genre and mood, and the model generates a track. It is also possible to add a vocal arrangement on top of a generated piece or remix an existing recording. The application runs entirely in the browser via Gradio and requires no installation. It is publicly available on Hugging Face Spaces at no cost.

microsoft/ai-agents-for-beginners

Microsoft recently released 18 free lessons about AI agent design. It covers the fundamentals of designing AI agents, to multi-agent designs and putting them into production. There’s a video and code samples for each lesson (Python). Lesson 12 explains how to do context engineering, so your agent can stay on track during long conversations. Lesson 11 explains agentic protocols like MCP, A2A and NLWeb. There are other lessons on agent memories, the planning pattern, multi-agent orchestration, computer use agents and production security. These lessons are currently translated into over 50 languages using automatic translation and has 75,000 stars on Github.

cactus-compute/needle

Cactus Needle 3 is a compact model family ranging from 8 to 29 MB, designed to run tool calls and structured extraction on resource-constrained devices: phones, watches, smart home controllers, and browsers via WebAssembly. There is one set of weights where every depth from 2 to 20 layers is a self-contained model — you pick the size that fits your hardware. The model takes a user command, looks at a list of defined tools, and returns a typed JSON output saying which tool to call and with what arguments. Fine-tuning with custom training examples costs $19 through the Cactus hosted service. The founder is direct in the HN discussion: good results come after investing effort in defining the tool list well for your specific use case.

supabase/supabase

Supabase is “the Postgres development platform.” We’re building the Firebase features you know and love using rock-solid open source tools. You get a fully-managed PostgreSQL database, user authentication, file storage, instant REST and GraphQL APIs, realtime changes, and a vector toolkit for your AI app all from a single dashboard. Get started today with zero-setup: sign up or self-host and run locally. Supabase includes a built-in vector toolkit to power semantic search or agent memory in your AI app. Since 2019 we’ve had over 110,000 GitHub stars and are licensed under the Apache 2.0 license.

shadcn-ui/lint

@shadcn/lint is an agent-first linter for Tailwind CSS design systems from the shadcn team. AI agents can generate UI code that includes colors, spacing, or utility classes that violate your project’s design rules. This linter helps detect violations and gives error messages with details about what’s wrong and how to fix them, so that the agent can self-correct on the spot without needing multiple passes. For most tasks, Claude and GPT models needed only one round of feedback on their design rules to achieve zero violations. Sonnet 5, Haiku 4.5, and Opus 5 achieved zero violations in all eight tasks after only one correction. Linting with feedback reduced the number of tokens needed to correct violations by 10–48% compared to feedback without linting. @shadcn/lint supports React, Svelte, and Vue apps using Tailwind CSS v4 and works with both ESLint and Oxlint.

githubnext/localjev

The TypeSafe Jev API allows you to pass in a message and a set of typed questions about that message (e.g. “what team should I assign this ticket to?”, “how urgent is this problem?”, “how frustrated is the customer?”), and get back responses for each of the types of questions in a single request. LocalJev provides an implementation of the TypeSafe Jev protocol that maps questions onto a classification prompt, queries a local instance of DiffusionGemma via an oMLX server, and outputs the response in the same format as the paid TypeSafe Jev API. The primary difference between LocalJev and Jev is that the latter outputs probabilities from the model’s logit space, whereas LocalJev asks the model to output the probabilities as text, meaning that the results are only approximate. We caution that users should verify the model’s calibration with their own data before making important decisions.

agno-agi/agno

Agno is a framework and runtime for agent platforms developed by the Agno team over the past four years, with 42,000 stars on GitHub. It aims to abstract away the infrastructure and let you focus on your agent. With one Claude Code prompt you can deploy a full service, including a database for storing memory and traces, an MCP server, a web management UI and a production API (with 50+ endpoints). Your agent is ready to talk to Telegram, WhatsApp, Slack, Discord right out of the box. There’s built-in scheduling, human-in-the-loop decision making for important actions, 100+ pre-built integrations to Github, Slack, Postgres, etc. Deploy to Railway, AWS, GCP, Azure or any other container hosting platform.

I tested it through the DeepSeek API and it processed both a long document and an image in the same request. The context window is one million tokens, which is roughly the length of a full novel or a large codebase. During generation the model activates only about 16 billion of its parameters, which is why responses arrive quickly despite the overall scale. It does not run locally and is available through the DeepSeek API or HuggingChat.

I gave it a style description and song lyrics and it returned a finished track with vocals. YuE2 is an open music generation model from a research group called m-a-p, and the authors compare the quality to Suno v5 and v6. From my testing, that comparison appears accurate. There is also a cover generation mode where you provide an existing melody and a new style and receive a version in a different genre.

Comfy-Org released Alibaba’s Qwen-Image-2.1 for ComfyUI users. Workflow templates of text-to-image generation and photo editing are included. It is a unified 7B parameter model for text-to-image generation and photo editing. Input a piece of text, get an image output. Input a photo, give some descriptions about what should be edited, get an edited photo. Up to ten reference images can be input for guiding the image style and subject. Support RGBA transparent background, which is good for product photos and catalogs. Research use only, please refer to the license before using it commercially.

zai-org/ZCode

ZCode is an open-sourced code agent developed by the Chinese lab Zhipu AI, creators of GLM models, CogVideoX, and CogView. It’s released under Apache 2.0 License. Pre-built packages are offered for macOS (Apple Silicon & Intel), Windows (64-bit, ARM64), and Linux. There is a web UI and a CLI agent. It provides a goal task type, where users just need to state their expected results, the agent will split the task into sub-tasks and execute it continuously. Screenshot support and multimodal tasks are available. Users can interact with the agent using Telegram/Feishu/WeChat bot. Multi-llm support can be easily achieved by modifying a config file.

browser-use/jev-ultrafast

I asked this agent to book a flight from Zürich to London on a particular day. It went to Google Flights, took 7.1 seconds (including page loading time and typing), and got the job done. It took just 101 steps to complete this task compared to over a thousand steps that other tools would take. You can simply state your objective as a sentence and let the agent figure out the rest.

tamaratran/fast-jev-compaction

When you have a long Claude Code conversation, all the previous actions are eventually summarized. This summary may omit an important piece of information such as file location or error message. This plugin allows you to evaluate each step of the process by scoring the actions and keeping key results verbatim.

robbietilton/Compositor

I don’t want to pay monthly fees to Adobe, but GIMP feels clunky. So the developer faced the same challenge and created a free Photoshop alternative for Mac before open sourcing the project. All the essential features I use in Photoshop are available including layers with masks, Content-Aware Fill and curves. The keybindings mimic those of Photoshop so that you can continue using it without having to learn anything new.

TheoLeeCJ/SemIf

Imagine you’re developing a product that requires sorting user messages into categories such as spam, complaint or urgency. Typically, you would feed these messages into a language model and receive a sentence which you then parse through code. With SemIf, instead of getting a sentence from the model, you ask questions with multiple choices and receive probability distributions. The tool is accessible directly in your browser without any installation required.

incoai/splash

Inco AI built a local inference engine for Apple Silicon that runs 27B-parameter models entirely on-device, with no cloud required. It serves Qwen3.8-27B (17.4 GB) and Qwen3.6-35B (20.9 GB) through OpenAI- and Anthropic-compatible APIs, so coding agents connect without any reconfiguration. On an M5 Pro with 48 GB memory, it decodes Qwen3.8-27B at twice the speed of the next-fastest engine and achieves 282 ms time-to-first-token with a 32K cached context. Streaming, tool calls, JSON Schema constraints, images, and inline PDFs are all supported. Switching Claude Code, Codex, or OpenCode to the local engine is a single command — `splash claude`, `splash codex`, or `splash opencode`. Installation is via Homebrew, and the model server starts with one command.

trycua/cua

Cua is an open-source platform that gives AI agents direct access to computers for operating desktop applications and graphical interfaces. The main component is Cua Driver, which installs on macOS, Windows, and Linux with one command and connects your agent to native apps through CLI, MCP, or a typed SDK. It works with Claude Code, Codex, Cursor, and other common agent tools. After installation, the agent can see the screen, press buttons, type text, and read content from any application, not only from a browser. Cua Fleets provides isolated Linux cloud desktops that the agent takes for one task and releases when finished, keeping the work separate from your own machine. CUA-S1 is a family of small specialized models for fast interface decisions; model weights are available on Hugging Face. For Apple Silicon there is also Lume, which creates local macOS and Linux VMs. The project is MIT-licensed.

Swift-Qwen3.8-27B is a reasoning efficient version of Alibaba’s Qwen3.8-27B which is created by UkisAI. This model is able to process both text and image data. It has a context length of 262,144 tokens. After UkisAI found some tokens which trigger unnecessary reasoning chains, they fine-tune this model to reduce the frequency of using these tokens. In this way, the new model has 58.3% less thinking tokens, 1.95x faster reasoning speed and less than 1% loss of performance. UkisAI launched a free OpenAI compatible API service (ukisai.com) for people and companies whose annual income is less than 1 million USD.

Prism ML compressed the Qwen3.8-27B model from Alibaba using ternary (2-bit) weights. The resulting file is 8.60 GB on disk and runs natively on Apple Silicon. The authors report 98.2% of the original quality is retained. The model understands images alongside text, writes and explains code, handles long conversations, and supports tool use for agentic tasks. On an Apple M5 Max it outputs around 47 tokens per second. This format works only on Mac computers with M-series chips.

This space was created by observantdistressed on Hugging Face with the Wan-AI Wan 2.2 video model. Upload an image and describe what you want to see happening in your video. Within a couple seconds you’ll receive a short video. This space is powered by Hugging Face’s free compute resources which means no cost to you! Wan-AI is a group dedicated to video generation models. Check out their platform here: https://www.wan.video/

ZhuLinsen/daily_stock_analysis

It’s an autonomous Stock Analysis System powered by LLM. Every weekday, it will read the news and calculate indicators automatically, and push a decision summary report via Telegram / WeChat / Discord / Slack / Email for each stock. Each report includes score, trend, entry/exit points, risks, events, and action plan. Currently supported: A-shares, HK, US, JP, KR, TW. The system can use any OpenAI compatible engine, like DeepSeek, Claude, Gemini, GPT, Ollama on your computer. Preferably using GitHub Actions to host it for free and without a server, just need model keys, notification channels, etc. Also there’s a web UI for viewing reports, running manually, and talking to the agent about a stock. Main README in Chinese, Full Translation: English. Licensed under MIT.

TauricResearch/TradingAgents

TradingAgents is a multi-agent stock analysis system from TauricResearch, built to mirror the structure of a real trading firm. A Fundamentals Analyst reviews earnings reports and company financials. A Sentiment Analyst aggregates news headlines, StockTwits posts, and Reddit discussions into a single market mood read. A News Analyst monitors macroeconomic events and their impact on prices. A Technical Analyst reads chart patterns and indicators such as MACD and RSI. Two researchers — one taking a bullish position, one bearish — then debate the analysts' conclusions in a structured discussion. A risk management team evaluates portfolio exposure, and a Trader Agent makes the final decision. Everything runs in a single CLI session: specify the ticker, the analysis date, and the LLM provider, and the system runs the full multi-agent cycle. Version 0.5.0 adds backtesting over a ticker and date grid and SEC EDGAR fundamentals — for instance, a backtest run before Apple restated its 2010 filing reads the original 2008 figures, not the revised ones. Works with any market Yahoo Finance covers, including US, Hong Kong, London, Tokyo, and India.

Player-YN/BrowserKitten

BrowserKitten is a Chrome extension that enables you to create a programmable web agent with your browser. The idea is that you click on any element of the currently opened webpage, specify what you want to do, and the agent will perform the required action on your live browser (where you’re logged in). It works with Chrome 135+, has extra features on Chrome 138+ and can be installed as an unpacked extension without building or using a package manager. BrowserKitten follows a BYOK (Bring Your Own Key) model, meaning that you can use your own AI key and process data on your desired provider. It’s based on Chrome MV3, and it makes use of a Service Worker, Side Panel, and ToolLoopAgent Component for executing sequential action flows.

NVlabs/SoL-Pi

SoL-Pi is an independent extension of Pi’s code writing agent developed by NVIDIA Labs. This extension introduces 4 optional optimization features discovered during the course of autonomous research. Action Fusion allows users to invoke a file modification action and subsequent validation command using a single tool invocation. ObservationPack allows users to transform frequently occurring large tool outputs into persistent references to avoid redundant model context replays. Evidence-Preserving Reducer allows users to summarize lengthy diagnostic messages into concise summaries while preserving all quoted lines. Online Context Compact allows users to remove finished plan actions from current context based on economic cost and window pressure constraints. These modifications do not require changes to Pi’s codebase and are implemented using Pi’s publicly available extension interfaces. An accompanying arXiv paper (2609.20519) provides more details on how these features were derived through autonomous research.

nilbuild/page-mascot

Page Mascot is a React open source component that allows to put an animated mascot on a web page. This mascot follows the movement of the mouse with 9 different head positions and displays one of 9 expressions when clicked. These mascots use two 3x3 spritesheets to ensure smooth transitions and perfect alignment between all positions. The package includes 52 predefined mascots and 6 visual styles (color, ink, sketch, riso, paper, pixel). Created by Kamran Ahmed, author of roadmap.sh. To request a specific mascot based on your description or picture, install the agent skill and ask. It should work directly in Codex or require an OpenAI API key in Claude Code to generate images.

mcncarl/jianying-headless

Jianying Headless is a local automation tool for Jianying Pro on macOS(Jianying is desktop app of CapCut, widely used in China and oversea), you could provide editing plan as json format contain video segments, multitrack layout, speed, volume, piip, subtitle and title, then it will automatically create editable draft of your jianying project in isolated copy, render output mp4 file using jianying native engine, also support jianying Agent Skill so you could call editing tool in your agent workflow. This is not official Jianying SDK, just a preview demo released independently. Tested on Jianying Pro 11.5.0 macOS 26.5.1(Apple Silicon).

crewAIInc/crewAI

CrewAI is an open-source Python framework for building multi-agent workflows. It offers two main abstractions: Crews (role-based autonomous agent collaboration where each agent has a specific role and goal) and Flows (event-driven automations that combine precise control, single LLM calls, and native support for Crews). Agents can use tools, share memory, and hand off work to each other. The framework works with any LLM. CrewAI has been in active production use for three years with over 58k stars and more than 100k developers certified through community courses. A commercial control plane called AMP Suite is also available for enterprise observability and managed deployment.

Minimax-h3_Singularity is a community fine-tune of MiniMax AI's MiniMax-H3 video model, made by WarmBloodAban. Per the model card, work went into fixing face distortion in medium and long shots, unnatural skin shine, HDR color range, movement dynamics, and VFX effects. The author reports three days of fine-tuning the model's weights. It's fully compatible with ComfyUI and supports all the generation modes of the original MiniMax-H3. To install, ask your agent: load this version of MiniMax-H3 in place of the base checkpoint in ComfyUI, download the model, and open your current video workflow to compare quality with the original. You'll need a GPU with 16GB+ VRAM — the same as for the original MiniMax-H3; it won't run on a regular laptop.

thruwire/foreman

Foreman is a Python runtime that works in tandem with Codex coding agents and can determine if code is finished, if the agent is stuck, or if a human should intervene. Foreman constantly monitors a software development task’s progress while a coding agent is running. It checks the git diff, agent logs, and history of steps every 30 seconds. It asks a rapid classifier 9 binary questions about whether the code is implemented, enough tests exist, the agent is stuck, the code is going off course, or a human should intervene. Then it uses a deterministic Python policy to decide whether to continue, guide, halt, retry, verify, or escalate. This is an architectural experiment and is not intended for production use. Accuracy needs to be calibrated. Codex CLI and TypeSafe API Key required. Or, use foreman demo --repo . for an entirely offline version.

jaredpalmer/kev

This model processes a text only once and returns multiple fixed answers (questions) about that text: Is it a hot lead or just a visitor? Is it spam? Which team should handle this? It gives back a number instead of a text: Spam. 0.89. Works local on your Mac. Free. The small version is really fast.

awlevin/typesafe-computer-use

This tool drives your Mac from a plain-English command — it clicks, scrolls, and fills in forms for you. Instead of sending a screenshot to a big AI model on every step and waiting several seconds, it reads the screen itself and asks a small, cheap classifier what to click next; it only calls a big model when something needs to be typed into a field. To install it, ask your agent to set up awlevin/typesafe-computer-use from GitHub, following its README: add a TYPESAFE_API_KEY (and optionally ANTHROPIC_API_KEY if you want it to type text) to a .env file, then run a command like clicker "open browser, go to arsentev.ai" --act. You need a Mac running macOS 14 or newer and Python.

NeoHorse-1-9B is our post-trained model on top of Qwen3.5-9B optimized for use in agent flows. This model can call tools and program while following complex instructions. On ten benchmarks, it achieves a macro avg score of 69.04 vs 65.60 for Qwen3.5-9B. With a native context length of 262,144 tokens, you can fit an entire small project into a single session. We provide quantized GGUF weights for use with Ollama or LM Studio. At 4-bit quantization, this model requires around ~5.4 GB of memory and will run on a laptop with 16 GB of RAM. While this model isn’t suited for cutting edge agent flows, it is a good representation of what is possible with a model this size.

AI Notes is a web app developed by Lynote that makes docs, websites, audio/video content into structured notes with citations to original source fragments. AI Notes is a source-grounded, bilingual, local-first AI note-taking tool. It parses your source into chunks, creates notes, and cites its reference to the specific fragment of the source it’s referring to. It enables you to verify everything you read in one click. Offline and local. Multi-lingual (English, Chinese). MIT licensed.

miuuyy/codex-chatgpt-web

This is an unofficial desktop launcher for ChatGPT Web, which allows you to use ChatGPT Web models with OpenAI’s Codex CLI (a coding assistant CLI) without using up any of your Codex API quota. When you choose a ChatGPT Web model in Codex, the request will be made using your regular ChatGPT subscription instead of using up your Codex API quota. From the README: “Use the ChatGPT Web models available on your account, including Pro, from Codex’s native model picker—with ChatGPT Web’s separate usage limits, without spending your Work or Codex quota.” Depending on your subscription plan, you can use Luna and/or Think (Free/Go), Instant to High (accounts with reasoning controls), and Extra High and Pro (when available). It also has full harness mode which integrates ChatGPT with the current task’s files, terminal, and tools via MCP, keeping the conversation contextualized to the Codex task. It comes with its own browser and runtime, meaning you don’t need to install Chrome, Node, or Bun. Works on macOS 13+, Windows x64, and Linux x64.

hypit-ai/hypit

Hypit is an open-source standalone project developed by the hypit-ai team that enables AI coding agents such as Claude Code or Codex to author videos using a language. At the heart of this is the main format that allows all the components (media, captions, B-roll, and effects) to be associated with words rather than time, allowing an AI agent to modify the structure of the video as if it were code. By providing a viral video as input, the AI agent will analyze it and generate a workflow that can be cloned. Then you can iterate on it to get variations (e.g., with another narrator, another subject, different characters, etc.) or even generate a completely new video in the same format. This pipeline may optionally invoke various APIs to generate content (see the examples in the README), such as the Seedance video models, the GPT Image 2 from OpenAI, and the Google Video Intelligence API for caption face tracking. The framework is free to use, but you will have to pay for the underlying models you wish to use. In the examples of the README, the author generated a 20s video using Seedance and GPT Image models at a cost of $1.15. The project uses the DeepSeek Chat Completions API via automated GitHub issue processing to automate triaging of contributors issues (this does not fall under the video pipeline).

ComposioHQ/awesome-claude-skills

Awesome Claude Skills is a collection of production ready Claude Skills and Plugins that have been carefully curated and organised by Composio. According to the authors, this repo is ‘A comprehensive and curated list of 1000+ production ready and practical Claude Skills and Plugins for enhancing productivity across usecases on not just Claude.ai, Claude Code, but also across coding agents like Codex, Cursor, Gemini CLI, Antigravity and more’. A skill is a directory containing a SKILL.md file with some YAML metadata and plain text instructions about how to train Claude on a certain kind of task (e.g. generating SQL, processing PDFs, summarising spreadsheets, making gifs, analysing meetings). It also gives access to over 1000 integrations with other applications via the Composio MCP Gateway. These include services like Slack, Jira, Notion, Google Workspace, Trello and Github. Writing a skill is as simple as writing plain Markdown. It was created in October 2025 and already has over 75,000 stars.

Fission-AI/OpenSpec

OpenSpec is a spec-driven development (SDD) framework for AI coding assistants. Before implementing anything, the agent uses slash commands to explore the problem (/opsx:explore), create a folder with proposal.md, specs/, design.md, and tasks.md (/opsx:propose), implement from the plan (/opsx:apply), and verify the result (/opsx:verify). Works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, and 33+ more tools. Built by Fission AI, launched August 2025, and reached 68k GitHub stars in one year. The framework is explicitly designed to be 'fluid not rigid' — specs stay alive and evolve as the project does.

ukisai fine-tuned Alibaba's Qwen3.8-27B, a multimodal 27B model that takes both text and images as input, to cut down its wordy reasoning process that slows things down. They do this by penalizing reasoning-marker tokens and using transfer-learning techniques inspired by BottleCap AI's ThinkingCap. The result: 58.3% fewer reasoning tokens (median), noticeably faster responses, and even a score increase on LiveCodeBench from 76.76% to 81.55%. The model is now available as GGUF (a packaged model format) for llama.cpp, Ollama, and vLLM. It's distributed under the Swift Open License 1.0 — free for personal use or for companies with annual revenue under $1M; commercial users above that threshold need additional agreements. To install it, ask your agent: 'Install the Swift version of Qwen3.8-27B locally via Ollama: https://huggingface.co/ukisai/Swift-Qwen3.8-27B-GGUF — pick the quantized version for my computer, run it, and show me how to use it.'

Meta introduced Llama 3.1 8B Instruct (part of multilingual generative models with 8B, 70B and 405B params) in July 2024. This is currently the most downloaded open model on Hugging Face, with almost six million downloads per month. It’s available in English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. The model has 128,000 tokens context, which should be enough for any long document. It also supports tools and functions, making it usable for creating simple local agents. It is ~4.8GB in 4-bit and can run on a laptop with 16GB of RAM.

The first generation reasoning model from DeepSeek (DeepSeek-R1) was published in January 2025. It has a mixture-of-experts architecture with a total of 671 billion parameters, but only 37 billion effective parameters per token. Currently, it is the most popular reasoning model on Hugging Face (almost 741,000 downloads in the last month). It is able to reason about complicated problems in a step-by-step manner. According to DeepSeek, it performs on par with OpenAI-o1 on math, code, and reasoning tasks. Under an MIT license, this model can be used both privately and commercially. They have also released distillations of their model from 1.5B to 70B parameters. I think that the most convenient way to start using this model locally is to download the 7B or 14B distilled version, which can be run on your laptop.

Comfy-Org released a ComfyUI Node for the m-a-p YuE2-3B Music Generation Model. YuE2 is a Music Generation Model that takes as input a description of the mood and genre, and outputs music, including melody, drums, instruments, etc. It includes the SheetSage2 Audio Encoder, and the actual model. To install, just put the files into your ComfyUI checkpoints directory, and audio_encoders directory. This model is licensed under a cc-by-nc-4.0 license, meaning it cannot be used commercially. We see it used in the community for Ambient, Lo-Fi, and Pop music. This is a 3B model, so audio quality is good enough for content creation, but not for professional music.

hacksider/Deep-Live-Cam

Deep-Live-Cam is a real-time face-swapper: Upload your face, select your Webcam, hit Live and watch your Webcam output the selected face in real time! Multiple Face Mapping at once, Mouth Mask to keep lip syncing realistic & post-processing capabilities for videos. Deep-Live-Cam has been released in 2023, gained almost 97k GitHub stars within three years, making it one of the most starred tools for AI-based content creation! A built-in content moderation prevents the use for unethical purposes such as pornography and violence. Get ready-to-use installers for macOS Silicon, Windows and CPUs, or follow the open source Python installation instructions for self-hosting.

code-yeongyu/oh-my-openagent

Oh My OpenAgent is a plugin system for AI coding agents — Claude Code, OpenAI Codex, and Cursor — that runs up to eleven specialized agents simultaneously from a single prompt. The project launched in December 2025 and reached over 69,000 GitHub stars in nine months. Author is developer code-yeongyu. Three editions: Ultimate with full capabilities (11 agents, 54+ lifecycle hooks), Light for CLI compatibility, and Senpi, a beta standalone runtime. In Team Mode all agents are visible in a split tmux terminal, each in its own pane with live status. Built-in MCP support, LSP integration, and AST-Grep let agents work with real code structure, not just raw text. This plugin system is designed around graph-based agent orchestration — each agent is a node in a task graph.

Qwen3.6 is a multimodal model from Alibaba's Qwen team designed for agentic coding — it processes text, images, and video in a single session. The 27B version scores 77.2% on SWE-bench Verified, placing it among the most capable models that can actually run on consumer hardware. Context window is 262,144 tokens natively, meaning a full repository fits in one session. The new 'thinking preservation' feature lets the model retain its reasoning chain from previous messages, so an agent doesn't restart its logic on each tool call. Available on Ollama in two sizes: 27B at 18 GB and a 35B MoE variant with only 3B active parameters at inference time (23 GB).

K2-Horizon-7B is a language model from IFM (Institute of Foundation Models) with a native 512K-token context window — enough to fit an entire mid-sized codebase or a large repository in one prompt. It was designed for agentic use: tool calling, terminal tasks, browser use, and multi-step code work. On SWE-bench Verified, the standard benchmark for patching real-world code, it scores 70.6% — well above other models in the same weight class, which score 30–50%. Training data, recipe, and evaluation code are fully public. Quantized versions run via Ollama and LM Studio.

yi1108/printfilm

PRINTFILM is an open source AI studio for creating short videos and episodic comics. It contains an end-to-end pipeline, including story generation using any OpenAI compatible APIs(e.g., Kimi, DeepSeek, etc.), picture generation using Volcano Engine Seedream, video clip generation using Seedance, and voice generation using Doubao TTS or free edge-tts. It’s a multi-user system with user management, pay-as-you-go pricing, and external APIs. You can deploy this system on your own server or sell it to others. There is also an admin dashboard to manage different models and change text sources without changing the configuration file. You need to have Docker installed(for Postgresql and Redis), Python 3.12, Node.js 20+, and FFmpeg on the PATH.

Qiuner/birdview

Birdview is a developer tool that creates architecture maps for your codebases and generates them as standalone HTML files. No server needed! Opens right in your browser. Each module is identified by a unique ID and connected with real code (not a diagram you have to draw yourself). Share the generated map with your AI agent before letting it work on a coding task. This way, it knows the whole project context and which modules it can potentially change. First create an evidence-based snapshot of your architecture and then define the scope of changes. Birdview does not track the activities of your AI agents. Your agents let Birdview know which parts of the code they want to change and Birdview helps keep the context in sight. Light/dark mode, filtering relationships, comparing architectures. Node.js 18+ and MIT license.

mem0ai/mem0

Mem0 is an open-source intelligent memory layer for AI agents and assistants built by YC S24 startup (mem0ai). Mem0 enables agents to remember users’ preferences, history of tasks and contexts across sessions without the need for constant repetition. In April 2026, we introduced a new memory algorithm that can perform ADD-only extractions with a single LLM call (no chains of UPDATE/DELETE), do entity linking between different memories and do multi-signal retrievals based on semantic search, BM25 keyword search and entity search. With our managed platform, LoCoMo benchmark accuracy went up from 71.4 to 92.5 and on LongMemEval went up from 67.8 to 94.4 (with proprietary optimizations on managed platform, SDK performance is comparable, open source). Mem0 requires an LLM to run (by default, we use gpt-5-mini by OpenAI but any LLM API can be used) and is available as pip/npm library, Docker Compose server (self hosted) or as a cloud platform (app.mem0.ai). Mem0 can be integrated with Claude Code, Codex, Cursor, Windsurf using npx skills add.

infiniflow/ragflow

RAGFlow is an open-source RAG engine developed by infiniflow. Released in late 2023, with ~91,000 Stars. You can have a complete end-to-end solution for building AI assistants that answer questions based on your documents. Upload PDFs, Word files, or connect to Confluence, S3, Notion, Discord, Google Drive. Support document processing (MinerU/Docling) , chunking, embedding, and retrieval. Comes with a visual agentic workflow designer and agent templates. Provide answers using its own built-in chat UI or integrate with Telegram/Discord/Feishu/Line. Support any LLM through API. e.g., OpenAI/DeepSeek/Gemini. Added MCP support in mid-2025. Deploy in cloud: https://cloud.ragflow.io/. Self-hosted: CPU>=4cores RAM>=16G Docker.

HKUDS/DeepTutor

DeepTutor is an open source AI powered learning environment developed by HKUDS team at Hong Kong University since Dec 2025. It allows users to upload their pdf/books/youtube links to generate an interactive learning environment with quizzes/flashcards/progress tracking. The learning environment is supported by an AI tutor with 3 layers of memories (trace, summary, synthesis) so that the AI can keep track of your learning progress across different sessions. Also, it supports integrating with live ai agents such as Claude Code, Codex, DeepSeek, Kimi etc. so that users can interact with them in a chatting style. It supports various RAG engines such as LlamaIndex, PageIndex, GraphRAG, LightRAG. It also provides a markdown notes writing tool, obsidian vault importing tool, and immersive youtube watching experience with time-based tutoring. The entire environment can be deployed using just one docker image (ghcr.io/hkuds/deeptutor:latest) or pip install deeptutor. It’s still actively developed (v1.6.8 released on Sep 14, 2026).

browser-use/life-recorder

Life recorder is an application developed by the browser use group. It consists of two parts: An iOS app, running on iPhone, that captures ~1min of AAC audio. A python script, running on Mac, that receives these audio files from the iPhone via local lan connection, and uses whisper.cpp from Whisper to transcribe them locally. The transcriptions are saved in a single, ever-growing, markdown file. This file also contains timestamps at hourly intervals. Once the Mac has received the audio files, and completed transcription, they are deleted from the phone. There is no cloud storage or 3rd party transcription service used. The connection between the phone and the Mac is protected using a self-signed tls certificate, and a randomly generated 256 bit bearer token. In order to compile this application on iOS you will need xcode and a free apple developer account. We have provided a SKILL.md playbook, so that a coding agent (codex/etc.) can look at the code, compile and deploy the application to a connected iPhone, with minimal user intervention.

This model (by Nex-AGI) is described as ‘a next-generation family of agentic models built for long-horizon tasks in real-world environments’. It can surf the web, use terminal, run & test code, reason about images, and invoke external utilities. It comes in three different reasoning modes (fast, medium, deep) based on the difficulty of the task. It has a score of 71.2 on OSWorld-Verified and 43.8 on SWE-Bench Pro. It is available in quantized format for llama.cpp, Ollama, LM Studio

Face Swapper powered by FLUX.2 Klein 9B by Black Forest Labs. It’s a generative image model. Upload your source image (with the face you want) and target image (where you want to put the face). Wait while the model will process your pictures and give you the result. Zero GPU servers by Hugging Face, no installations/payments required for simple usage.

zhayujie/CowAgent

CowAgent (previously named chatgpt-on-wechat) is a self-hosted AI agent for 24/7 use. It has ability of task decomposition and planning with long term memory and session learning, built-in browser, file I/O, terminal, scheduler, web search, mcp intergration, install skills from skill hub or github, structured personal knowledge management, personal knowledge graph building and automatic long term memory learning with “Deep Dream” conversation. Three-level memory structure for auto learning from historical sessions and self-evolving. Supported all mainstream LLMs like Claude, GPT, Gemini, DeepSeek, Qwen, etc. Can be easily switched on web console. Integrated with Web, Telegram, WeChat, Feishu, DingTalk, Slack. Installer one line for Linux/macOS/Windows. Also available for docker and cloud.

viettranx/3dviz-pro-max

3dviz-pro-max is a standalone agent skill for Claude Code (plugin) and Codex (skill folder) that turns a one-sentence prompt into an interactive Three.js or Blender 3D scene. The key feature is a capture helper: after generating the code, the agent renders the scene, captures frames, and verifies what actually appears on screen — rather than describing the result from memory. The skill ships a 10-step workflow, 223 recipes across 24 visual directions (isometric, photorealistic, stylized, and more), 440 knowledge records, 22 ready-to-use kit blueprints, and 37 runnable study scenes with proven visual styles. Installs as a single plugin command in Claude Code or as a folder dropped into Codex.

Significant-Gravitas/AutoGPT

AutoGPT is an open-source AI agent platform that has grown from the original 2023 "auto-agent" experiment into a full hosted service. It offers four main surfaces: AutoPilot (describe a job in plain English and it becomes a working agent), a visual builder (drag, connect, and branch blocks for precise workflow control), a marketplace (proven community-built agents you can install and customize), and an agents dashboard (track every run, cost, and action). The hosted Platform manages infrastructure, model access, credentials, and updates — no self-hosting required, though that option exists. The project has 185,000+ stars and is cited by Andrej Karpathy and Replit CEO Amjad Masad.

ruvnet/ruflo

Ruflo is an agent meta-harness built specifically for Claude Code and Codex by ruvnet. The premise: 'Agent = Model + Harness' — the model writes, and the harness gives it tools, memory, loops, sandboxes, and controls so it can actually work. Ruflo adds 100+ specialized agents (coding, testing, security, docs, architecture), self-organizing swarm coordination, persistent cross-session memory, federated communication across machines, and security guardrails. The project is over a year old with 72k stars and reports 8.1M ecosystem downloads. It installs directly from the Claude Code plugin marketplace or via a single npx command. The full feature set includes 33 Claude Code plugins, 21 npm plugins, a web UI at flo.ruv.io, a GOAP A* goal planner, and support for Claude, GPT, Gemini, Cohere, and Ollama.

rizqinrr/viserys-agent

Viserys is a self-contained pack of 28 Markdown workflow skills for AI coding agents that spans the development process: DEFINE (clarify idea, identify requirements, spec-driven and constraint-driven development), PLAN (database schema design, tasks decomposed with acceptance criteria), BUILD (incrementally implement, TDD, context engineering, frontend UI, API design), VERIFY (browser testing using devtools, debugging), REVIEW (code quality, security hardening, performance optimization), SHIP (git workflow, CI/CD, migration, documentation, observability, launch). It comes with 4 reviewer personas: a staff-engineer level code reviewer, test engineer, security auditor, web performance auditor. Each skill is written in a simple Markdown file that any agent can understand without additional tools. There’s an out-of-the-box Viserys agent definition in the repository for OpenCode users so you can start using all 28 skills in seconds by pressing the Tab key.

letorig/video-generator-client

An async python wrapper and WebUI that aggregates four mainstream AI video synthesis services including Seedance (ByteDance), Kling (Kuaishou), MiniMax/Hailuo and Wan (Alibaba). Features text2video, img2video, polling live tasks, mp4 downloading, etc. It’s embedded in the python package, no need for Node.js or building the frontend. Supports the latest models: Seedance 2.5 supports 30 seconds of video with original audio. MiniMax H3 supports 4k@60fps up to 30 seconds of video. Kling 3.0 supports multishot with original audio. Wan 3.0 supports up to 1080p@30 seconds of video. MCP server for Agent integration.

agentverse-os/AgentVerse-OS

AgentVerse OS is a personal cloud operating system for developers working with AI agents, designed to run on a single Ubuntu server and be accessed through any browser. After a one-command install, you get a browser-based desktop with VS Code, Claude Code, Codex, and a terminal — all pre-configured and ready. Each project lives in its own isolated container with dedicated networking, so agents working on different projects don't interfere with each other. The built-in catalog includes 944 self-hosted apps from the Runtipi, Coolify, and Umbrel collections — databases, monitoring tools, media services, and productivity apps. All access is routed through Tailscale, so nothing is exposed to the open internet. The project is currently in alpha (v0.2) and supports single-user operation.

Installed MiniCPM5-2B from OpenBMB through Ollama — takes 1.5 GB quantized, runs on a laptop without a GPU. A 128K-token context window at this size is unusual: you can feed entire documents in one shot. Tool calling works out of the box: give it a task, the model decides what to run. Ran it on real tasks — solid for 2.5B parameters.

Nex-AGI released Nex-N2.5-Pro — an agentic model built specifically for long, multi-step tasks. Tried it via OpenRouter: hand it a chain of dozens of steps and it holds the thread all the way to the result without losing context. Open weights, Apache 2.0, but too large to run at home. What I liked: the authors are upfront about the goal — real tasks in real environments, not lab benchmarks.

We’ve applied routing-guided agentic post-training (a special kind of training technique to improve performance on agentic use cases such as tool use, task planning, and coding) to Alibaba’s Qwen3.5-4B, which achieved a +5.93 macro-average improvement on 10 benchmarks compared to Qwen3.5-4B baseline and showed remarkable improvements on agentic tasks (+9.79 on WorkBuddy Bench, +6.21 on QwenClawBench, +9.75 on HumanEval). The original context window size is 262144 tokens. The 4-bit version of the model (4.2B params) takes about 2.5 GB disk space. You can try out this model via GGUF repack by mradermacher here: https://huggingface.co/mradermacher/NeoHorse-1-4B-GGUF on LM Studio or Ollama.

We have developed a 35B sparse MoE model (Qwen3.5-35B) which runs on Apple Silicon with SSD offloading of experts: weights are kept on disk and loaded as needed into RAM, so only 3GB of active memory is used. A typical 35B model requires 64GB or more of RAM to fit its weights into memory. The full model weights are 19.6GB of disk space. It generates text at approximately 15 tokens/second, and is able to read input (prompt completion) at 140 tokens/second. You can control it via a custom CLI tool which you can install via pip, and use it either in chat mode or API server mode.

sansan0/TrendRadar

TrendRadar is an artificial intelligence-assisted online public opinion monitoring system and trend hotspot information collection, which integrates the latest hot topic information on multiple websites such as GitHub, HackerNews, Reddit, Weibo and other websites, and sorts out the latest hot topic information by keywords and ai filtering, and sends it to the back-end by various ways, including telegram, slack, email, bark, wechat, dingtalk, ntfy, etc. The version upgrade 4.1 officially supports mcp, which can be used as a real-time news source for AI agents. You can also add rss as your information source. Quick deployment within 30 seconds with docker one line code. Active maintenance for more than a year and a half. 62k+ Star.

koala73/worldmonitor

World Monitor is an open source real-time global intelligence dashboard that gathers news about geopolitics, finance, tech, commodities, and energy, then uses AI to generate short news briefs. The UI features a 3D Globe (globe.gl) + WebGL Flat Map (deck.gl), live panel feeds, and a Country Instability Index (CII) for 31 countries. It’s available as a web app (worldmonitor.app), as a native Tauri 2 Desktop App for macOS/Windows/Linux, or as a self hosted app using npm. The AI model can run locally (no API keys required) using Ollama, Groq, or OpenRouter, so pick your poison. We have multiple sites running off of the same codebase including tech.worldmonitor.app, finance.worldmonitor.app, commodity.worldmonitor.app, energy.worldmonitor.app, and happy.worldmonitor.app. For agents, we also have an MCP Server available at worldmonitor.app/mcp which lets you expose all of our monitoring tools to any MCP compatible agents such as Claude Code.

MiniCPM5-2B is a 2B parameter model by OpenBMB (part of their MiniCPM5 series), meant for local and on-device use. According to OpenBMB, it’s the “best open model of its size” in their tests and even rivals 4B models. It’s good at coding, math, long context, tools, and agency. The repo for the GGUF files for llama.cpp, Ollama and LM studio can be found here. The compressed Q4_K_M version is ~1.6GB and the 8-bit version is ~2.7GB. OpenBMB also released the training data for this model and the weights are available under the Apache 2.0 license.

firecrawl/firecrawl

Firecrawl is an open source web scraping service developed by the Firecrawl team. They also have a hosted API service. They crawl the internet, open web pages and return either clean Markdown text, structured JSON data, or screenshots rather than raw HTML code. This tool claims to crawl 96% of the internet, even if the page requires JavaScript to render. It also takes care of proxies and rate limiting. You can use this tool to simulate mouse clicks, scroll and typing. You can also extract information from PDF and DOCX documents, or use it as an MCP tool to connect to other AI agents.

eternityspring/reelbench-skills

reelbench-skills: a collection of skills developed by Wesley Wang for Claude Code & Codex. The first skill parses a given video into a shot table with time, shot type, camera motion, scene composition, edit pace. This is measured by ffmpeg and all subjective interpretations are delegated to the model. The second skill visualizes a video and its corresponding shot table simultaneously. Shots will be highlighted during playback. This tool will generate an offline HTML report with a player, shot table search, and keyframe display. This tool does not require any npm package, and it does not require an API key. There are some Chinese descriptions in the repo description, please ask your agents to translate the English version of the README if necessary.

Edge0-AI/Edge0

Edge0 is an open source inference engine developed by the team at Edge0 for running large mixture of experts models on macOS with Apple Silicon. Rather than keeping the entire model in memory, Edge0 streams required pieces from the SSD and anticipates future needs, leading to up to 59% faster generations according to its creators. The package comes with two pre-trained models: a 35 billion parameter model (23GB disk size, ~2.9GB max memory usage) and an 8B parameter model (~4.2GB disk size). These models are based on open source base models and feature adapters to regain lost quality due to compression. Currently, Edge0 is only available for macOS with M1 through M4 processors.

TimesFM 3.0 is a new release of Google Research’s pretrained time-series foundation model for forecasting. Given a time series (e.g., daily sales, website traffic, or inventory), TimesFM predicts future values and quantiles, giving a distribution of possible values instead of just a single point estimate. The model was trained on many time series datasets, including Wikipedia pageviews and Google Trends searches, as well as synthetic data. Official PyTorch weights are here. It has ~330M parameters. Model weights licensed under TimesFM Non-Commercial License v1.0.

MadsLorentzen/ai-job-search

This is a free and open-source job application framework built on top of Claude Code. You only need to enter your profile information once and then you can use some commands to help you apply for a job. Use /scrape to search online for job listings and see recommendations for jobs you will be a good fit for along with a fit score. Use /apply to assess a job, write a personalized resume and cover letter with a recommendation from a reviewer agent and then send the final product. Use /interview to assist you in interview prep. This tool was developed by geophysicist Mads Lorentzen who lost his job late 2025. He used this framework to apply for new jobs and found a job as an AI engineer in June 2026 after 69 personalized applications, 20 first round interviews, and 1 signed contract. This framework is currently language agnostic and country agnostic, you can also integrate this framework into other agents like Codex or Gemini CLI using AGENTS.md. This is approximately 42,000 GitHub stars, and is under a MIT license. This is an independent project and is not associated with Anthropic, the creators of Claude Code.

google/artemis

ARTEMIS is an open-source tool developed by Google for converting natural language to Android automation. This includes running end-to-end tests and flows between applications, such as opening Google Maps to navigate somewhere and then playing music on YouTube. There’s also an integrated MCP server that allows the phone to be controlled by AI code assistants like Claude Code, Codex, Antigravity, and Windsurf. This allows the agent to gather Logcat information and take screenshots. There’s a single-click script to set up all the required software (ADB, scrcpy, FFmpeg), start a local web server for device connection and screen mirroring, and replay executions. ARTEMIS claims to achieve over 99% success rate on the AndroidWorld benchmark. It’s licensed under Apache-2.0 and reached 3,000 GitHub stars in its first month.

Mistral Medium 3.5 is Mistral AI's first 'merged' flagship: a 128-billion-parameter dense model that combines instruction-following, multi-step reasoning, and code generation in a single set of open weights. Previous Mistral models were specialized — Devstral for code, Mistral Large for dialogue. Medium 3.5 does all three in one request, plus it understands images via a custom-trained vision encoder. Context window is 256,000 tokens. On SWE-Bench Verified (real GitHub bug-fixing tasks) it scores 77.6%; on the agentic τ³-Telecom benchmark — 91.4%. Reasoning effort is configurable: it can switch off for fast replies and engage fully for complex problems. Weights are open under a modified MIT license; self-hosting requires approximately four GPUs.

petergyang/no-ai-slop

No AI Slop is a writing skill that detects and removes common AI-generated text patterns while preserving the author's authentic voice. It identifies 20+ recurring patterns such as binary contrasts ('It's not X, it's Y'), throat-clearing openers, weasel attribution phrases ('experts agree'), fake-profound statements, and buzzword-heavy empty analysis. Three modes: edit mode (cleans the text and lists exactly what changed), detection mode (finds patterns without modifying the text), and satire mode (intentionally generates clichéd AI writing, for humor). Available as a skill in Claude Code and Codex via a single npx command, as a plugin in ChatGPT, or by pasting the install prompt directly into your AI chat. Install takes one command.

bytedance/deer-flow

DeerFlow is a open source engine developed by Bytedance(Tiktok). It allows multiple AI agents working together to complete a task simultaneously rather than sequentially. DeerFlow 2.0(released Feb 2026) has been refactored to allow sub-agents to collaborate on a task, such as searching online, writing and executing sandboxed code, and verifying the results. They can also share a persistent long-term memory across sessions. DeerFlow integrates with OpenAI, Anthropic, Gemini, DeepSeek, Qwen, and other OpenAI compatible APIs such as OpenRouter. You can easily get started by asking your coding agent to clone the repository from github.com/bytedance/deer-flow, follow the instructions in Install.md, select Docker(a tool to package applications for running anywhere), specify your preferred LLM provider and API keys, and access the web UI at localhost(your local host) via a web browser. This will take you just a few minutes to set up.

Krea 2 Turbo is a 12 billion parameter text to image diffusion model developed by Krea.ai and made available as open weights under the Krea 2 Community License. It allows you to create images from textual descriptions in a freeform natural language style, with applications ranging from creativity and commerce to research. It’s currently available as part of Draw Things and DiffusionBee (desktop apps for local image synthesis, macOS only), Google Colab and Kaggle notebooks, and over 100 community spaces in the Hugging Face hub. At 4-bits, the model requires around 7.7 GB of memory, which makes it suitable for a laptop with 16 GB of RAM.

SAM 3 is our general-purpose foundation model for segmenting objects in images and videos. Given textual prompts describing the desired object to segment, the model produces a precise segmentation mask. We also support visual prompts (e.g., points, bounding boxes, sketches of segmentation masks) and their combination with text. Furthermore, the model can automatically track objects through time without frame-by-frame annotations. On the SA-CO benchmark (which includes 270K unique visual concepts, i.e., more than 50x compared to prior benchmarks), SAM 3 achieves 75–80% human performance. It consists of 0.9B parameters and is ~1.7 GB in full precision.

Vincentwei1021/anything2explainer

anything2explainer is a skill for Claude Code and Codex that produces full-featured animated explainer videos from any subject matter. After installing this skill by cloning the repo to your skills directory and specifying your chosen topic, the agent pipeline will perform research, generate narration, synthesize TTS with word boundaries, build a storyboard with individual shots, and spawn parallel build agents that render each frame as a Remotion (React) component. The final product is a 1280×720 H.264 MP4 video with audio narration, subtitles, chapter cards and progress bar. Generating a 2–3 minute long video takes roughly 1 hour of wall-clock time, while a high-quality 3–5 minute reference video takes about 2 hours. Both English and Chinese languages are supported out of the box. At four key moments, you will be asked for input.

unslothai/unsloth

“Unsloth: The first desktop app to run and train models.” Unsloth Studio (run with unsloth studio) provides a web UI for browsing models, downloading and running LLMs, diffusion, TTS, and embedding models locally. Datasets can be built via a drag-and-drop UI, with support for PDF, CSV, DOCX file types for model fine-tuning without code. Available for Windows, macOS, and Linux. Supports NVIDIA, AMD, Intel GPUs, CPUs, Vulkan. Optional integration with ChatGPT / other cloud services. Benchmarks of model performance for Google (Gemma), OpenAI (gpt-oss), Alibaba (Qwen), and Meta (Llama). Fine-tuning 2x faster, 70% less VRAM than traditional pipelines. MoE model preparation 12x faster, 35% less VRAM. 75 000+ stars on GitHub.

Comfy-Org/ComfyUI

ComfyUI is a visual node-graph editor for running AI generation models locally. You assemble pipelines by connecting blocks — a model loader, a text encoder, a sampler, a VAE decoder, a save node — and the data flow is visible before you hit run. It natively supports over 30 model families for image, video, audio, 3D, and text generation, including Stable Diffusion, Flux, Wan, LTX-Video, HunyuanVideo, and Qwen3. VRAM is managed automatically with smart offloading and partial graph re-execution, so changing one node does not restart the entire pipeline. Custom nodes extend the graph with thousands of community-built operations. Workflows are saved as JSON files, shareable and reusable. It runs fully offline — no telemetry, no mandatory cloud.

Nemotron 3 Nano Omni is a multimodal model from NVIDIA that processes video, audio, images, and text in one inference pass. The 31-billion-parameter architecture activates only a fraction of weights per token, reducing memory pressure despite the overall parameter count. It understands video clips up to two minutes long, transcribes speech with word-level timestamps, reads documents with embedded charts and tables via OCR, and supports GUI automation for agentic tasks. The context window spans 256,000 tokens — enough for a lengthy document or a multi-turn session with multiple file inputs. NVIDIA built the model as part of the Nemotron family, drawing on Qwen3-VL-30B-A3B from Alibaba and other Qwen variants. The Ollama release makes it available locally; the compressed download weighs around 28 GB.

MiniCPM-V 4.6 is a pocket-sized multimodal model from OpenBMB: roughly 1 billion parameters, 1.6 GB at full precision. It is the lightest model in the MiniCPM-V line — version 4.5 carried 8 billion parameters. It reads single images, multiple images at once, and video clips, with a 256,000-token context window. The size drop comes from a different backbone: SigLIP2-400M plus Qwen3.5-0.8B in place of Qwen3-8B. Separately, mixed visual token compression cuts encoding FLOPs by more than 50%. There is a trade-off, and OpenBMB state it plainly: the model reaches Qwen3.5 2B-level capability on many benchmarks, while the larger 4.5 stays more accurate. Officially supports iOS, Android, and HarmonyOS, and runs via Ollama with a single command.

upstash/context7

Context7 provides up to date and versioned documentation for libraries on demand for your AI code generating agent. Without Context7, your AI agent would be making code based on outdated training data that could be months or years behind, resulting in it using deprecated APIs or antiquated styles of programming. With Context7, when your agent asks questions about libraries, we fetch the latest docs from their source and insert them into the agent’s context window. Context7 can be used both as a CLI skill (no MCP needed!) and as an MCP server (for native MCP calls). Get a free API key from context7.com to raise your rate limit! Compatible Agents: Claude Code, Cursor, OpenCode.

langgenius/dify

Dify is an open source LLM app builder. Upload your PDFs, Word docs, spreadsheets to the visual canvas, choose a model (Claude, GPT-4, Llama, DeepSeek, or any other OpenAI compatible endpoint), and create an AI chatbot/agent without coding. There are over 50 prebuilt agent tools such as Google Search, DALL-E, Stable Diffusion, WolframAlpha, code execution, etc. Each app has a built-in REST API so that you can easily integrate it with Telegram bot or put a widget on your website. Deploy to Dify Cloud (free sandbox plan provides 200 GPT-4 calls) or host on your own machine.

HKUDS/nanobot

Nanobot is an ultra light-weight, open-source, self-hosted personal AI agent framework implemented in Python by Hong Kong University of Data Science. It provides WebUI to access it via browsers, a terminal client and chat app (all share the same agent). Your agent has long term memory, tooling integration via MCP, cross AI model routing, task subdelegation to subagents, schedule automation and OpenAI compatible APIs for third party applications. Single curl command to install on MacOS/Linux, one line powershell code to install on windows. No hassle. Just launch the WebUI and talk to it. Support all model providers like OpenAI, Anthropic Claude or local models using Ollama. Published 7 months ago. Already getting close to 48,000 stars.

bilawalsidhu/gods-eye-view

God’s Eye View is an Open Source 3D World Globe Simulator showing Realtime Aircraft Transponder Data, Ship Beacon Data, Satellite Data, Earthquake Data, Traffic Data and Public Cameras from Open Sources. It features a Voice Agent powered by OpenAI’s Realtime API which allows you to interact with the World Globe using Voice Commands like “Go to Tokyo”, “How many flights are over Texas?”, “Night Vision Mode”, “Draw a border around the State of Texas”. The Voice Agent provides 28 tools including Camera Tools, World Annotation Tools, Live Layer Queries and Console Commands. There is also an AI generated Heads Up Display (HUD) which shows 5 word Intelligence Descriptions of the Scene which are updated in Realtime as you move around. This Project became Viral on Youtube (5M+ Views) and Top of Github Trending in August 2026. It was only released two and a half months ago but its already very close to 20,000 Stars. You can install this easily in 1 Click using Pinokio or Run it Locally and get Keyless Access to Flight Data, Ship Data, Satellite Data, Earthquake Data, Camera Data. Its also using Fully Photorealistic 3D Rendering using Google Photorealistic 3D Tiles which can be accessed for Free using a Cesium ion Community Account or using a Google Maps API Key.

achimala/dream-loop

A Dream Loop is a creative agent skill that allows you to develop games, apps and visual scenes iteratively within a closed loop. It works by having an AI first dream up a high quality target image which represents the end goal and then using this to build your project. This process is accompanied by an AI critic giving feedback on the difference between the live image and the target image until the critic is happy. Once this occurs, another option is to have the AI dream up again to generate a more ideal target image based on the current project state. Can be used with any agent that has an image generator, vision input and subagent capabilities (currently tested with GPT-6 Astra on Codex, but it should work with good models like Claude Fable 5 etc.) Install by running a single command `npx skills add achimala/dream-loop`

Z-Image-Turbo is an open text-to-image model developed by Tongyi-MAI(Alibaba AI group). It’s a compact version of Z-Image model with the same image quality but only requires 8 inference steps (compared to the standard 50). You could also test it out online at Hugging Face Space (without any GPU or registration): just input the description and specify the image size (512–2048 px) and the optional random seed to reproduce the result. This model supports English/Chinese text rendering in the image. And the weights are open source.

SenseNova U1.5-8B-MoT is a multimodal model developed by SenseNova (a Chinese artificial intelligence company). It provides image generation and editing capabilities in one tool. It can perform object substitution, local editing, image captioning correction, and partial scene renovation while keeping other details unchanged. It has the ability of clear text generation. Not only can it generate clear English text, but also Chinese text. This is something that most image generation models currently lack. It supports output resolutions up to 4K and is able to execute multi-step complex instructions based on images.

yanliudesign/mono-color-skill

A Claude agent skill by designer Yan Liu that transforms a theme, phrase, photograph, or article idea into an original editorial image in one-ink or two-ink print style. The skill follows a five-step process: read input, choose layout, assign ink plates, compose the page, then generate. Default settings use a two-ink design with 70–85% dominant ink and 15–30% accent on a neutral substrate. Eight single-ink options are available — Cobalt, Botanical Green, Terracotta Orange, and others — plus nine two-ink recipes. Visual rules enforce that 25–55% of the composition stays empty, with halftone, risograph grain, or photocopy texture. The skill vets results for ink count and originality before delivering. Released three weeks ago and already approaching 3,000 stars.

wshobson/agents

A plugin marketplace that provides 94 domain-specific plugins (202 agents, 183 skills, 105 commands, 16 multi-agent orchestrators) from a single source that easily integrates with Anthropic Claude Code, OpenAI Codex CLI, Cursor, OpenCode, Antigravity CLI, and GitHub Copilot. Each plugin focuses on a particular domain (Python, front-end, SEO, security, ML, infra, docs, etc.). When you install a plugin, only its parts get added to the context; not the entire marketplace. You can install skills individually too using gh skill install wshobson/agents or npx skills add wshobson/agents without cloning the repository. This marketplace has been running for ~1 year now and has earned almost 40,000 stars. It’s one of the most starred Claude Code plugin repositories.

shanraisshan/claude-code-best-practice

A community-driven guide to best practices for Claude Code, featuring 83 tips on prompting, agents, CLAUDE.md, hooks, MCP servers, workflows, memory, and much more. This includes curated collections of skills, examples of real-world workflows used in large projects, and an exhaustive index of all the new features in Claude Code. This covers popular new features such as Ultrareview, Devcontainers, Auto Mode, Computer Use, and the Agent SDK. Continuously updated by community members with feedback from the Anthropic team.

jtydhr88/screenwriting-skills

The skills include information about story structure, premise/theme, characters and conflicts, dialogues, scenes, formats/adaptations and industry knowhow. Descriptions of the skills are available in both English and Chinese, while their bodies are in Chinese, which Claude understands and uses no matter your prompt language. The skills are installed through the plugin marketplace of Claude Code (just one command to install all of the skills).

inclusionAI/Choruz

Choruz is a local first collaboration tool for humans and AI agents like Claude Code, Codex, Pi, Grok, OpenCode, or any webhook enabled agent. Every agent has its own running CLI in its own workspace with access to files, terminals and task boards. Communicate with your agents using direct messages, groups, threads and @mentions like in a chat application. Let your agents work together by automatically delegating tasks. Built-in browser based terminal, file manager/editor, ssh access to remote systems and (optional) bridges to Slack and Telegram. Developer preview, interfaces and data formats are subject to change.

github/spec-kit

Spec Kit is an open source toolkit from GitHub for spec-driven development using any AI coding agent. Rather than communicating features in an informal way, you can define your desired outcome and why in a structured spec. Then the CLI (specify-cli) will walk you through the process using slash commands (/speckit-specify, /speckit-plan, /speckit-tasks, /speckit-implement, /speckit-converge) to help you set up project principles, write a spec, create a technical plan, define tasks, implement, and finally converge on the result until it meets the spec. The following optional extensions are also available: bug fixes (assess → fix → test) or ideation (intake → research → define → shape → decide). Spec Kit is compatible with GitHub Copilot, Claude Code, and Codex. This project has achieved over 134,000 stars in its first year.

donvito/codex-astra-luna-orchestrator

A configuration package for Codex that sets up a multi-agent topology: GPT-6 Astra acts as root orchestrator and final reviewer, while GPT-5.6 Luna handles execution subagent work. Includes five agent roles—explorer, worker, researcher, tester, and reviewer—plus a token-usage analysis tool and install scripts for macOS, Linux, and Windows. Adapts automatically to Pro or Plus plans during setup, keeping orchestrated sessions within Plus rate-limit windows when needed. The Astra orchestrator skill is invoked with $astra-orchestrator.

MiniMax M3 is MiniMax's flagship model for coding and agentic tasks, now accessible through Ollama's cloud platform. The context window reaches up to 1 million tokens (guaranteed minimum 512K) — enough to load an entire medium-sized codebase, long conversation history, or extensive documentation into a single request without truncation. The model is natively multimodal: it processes text and images in one unified pipeline, so you can attach a screenshot alongside text and it understands both together. On BrowseComp — a benchmark measuring how well a model autonomously finds and synthesizes information from the web — MiniMax M3 scores 83.5, compared to 79.3 for Claude Opus 4.7. This is a cloud-only model: it runs on MiniMax's servers and is available through Ollama at $0.60 per million input tokens and $2.40 per million output tokens. The weights are publicly released, but running them locally requires data-center-scale hardware.

DeepSeek just added vision to their V4 Flash model — I tested it through their API and it genuinely understands screenshots, charts, and images alongside text. What surprised me most is how well it handles agentic tasks with visual input: sent it a screenshot of a broken UI and it pinpointed the problem without me describing anything. The benchmark numbers on agentic tasks are strong. You can't run 305B locally, but it's available through DeepSeek's API right now.

Downloaded Spark-X2.5-4B from XHToken (SparkLLM Team) and was genuinely surprised: a 4B model that runs on any laptop with 8 GB RAM, yet supports coding, translation, tool use, and structured generation. The 1 million token context window at this size is unusual — most 4B models top out way lower. Over 200 languages. Set it up in LM Studio in about five minutes with no friction.

GLM-5.3-Flash from Zhipu (Z.ai) is their first natively multimodal model in the GLM-5 series — 321B total parameters but only 18B active at inference time (MoE architecture), which makes the API cost much lower than the parameter count suggests. The team claims it approaches Claude Opus 4.8 on coding and agentic benchmarks at one-tenth the price. I plugged it into an agent pipeline via the Z.ai API as a cheaper replacement for heavy reasoning steps. Context is 300K. English and Chinese only — not a universal language model.

MiniMax, the creator of mainstream multimodal models, launched the H3 video generation model. It can automatically generate a video based on an image or a piece of text, along with a sound. With the help of ModelTC, MiniMax H3’s generation speed has been improved. This version only needs several times fewer generation steps to get the same result. And it comes with pre-installed ComfyUI workflows, so there is no need to manually set up a workflow. This model supports both image-to-video and text-to-video generation. In the first month, it was downloaded over 1.26 million times. You can try it online without installing at ModelTC’s official studio x2v.light-ai.top.

yang0/handraw-style

A gallery of 216 numbered hand drawn illustration styles for AI image generation. The gallery consists of 7 parts: international editorial illustration (001–035), storybook illustration (036–054), modern graphic design (055–082), japanese contemporary illustration (083–123), chinese contemporary illustration (124–154), web & media aesthetics (155–200), chinese illustration supplements (201–216). You can use the internal html gallery to select a style by number, then input ‘number + theme’ to receive bilingual prompts and send a reference image to an image generator if needed.

tinyhumansai/openhuman

OpenHuman is an open source personal AI desktop application for Mac, Windows & Linux. Written in Rust, and bundled in a simple installer with no terminal or config files. Automatically syncs your emails, calendar, documents & chats to local compressed Markdown files in a SQLite-based Memory Tree upon first use, and re-syncs every 20 minutes. Integrates with more than 100 OAuth providers, 5000 MCP servers and 90,000 Skills, seamlessly dispatching each job to the optimal model. TokenJuice compresses data sent to the model, reducing tokens used by up to 80%. Privacy Mode enforced at the Rust level allows you to turn off all inference and ensure it never leaves your device. The OpenHuman app can manage concurrent agents, orchestrate durable workflows with a visual interface, contact you through 17 different chat channels (Telegram, WhatsApp, Slack, email client, etc), and even generate images & videos out of the box.

tigerless-labs/agent-memory

Agent Memory is a local-first, agent agnostic, long term memory runtime for AI agents. Memories are saved as plaintext Markdown files on your computer. Greppable, git trackable, portable. There’s also an SQLite index for a rebuildable cache. It’s safe to delete. Uses conversation state to write memories, not agent timing. Use a Manage pass outside of conversations to age/archive memories based on value, can only suggest deletions (never perform them). 52.9% LongMemEval-S accuracy (vs. 35.8% MemCore, 5.8% no memory). Linux, Windows, macOS (Apple Silicon) support.

sickn33/agentic-awesome-skills

The Agentic Awesome Skills (AAS Core) is a catalog of 2113 pre-written SKILL.md playbooks — reusable instruction files for AI agents to read and run. For example, Claude Code, Codex, Cursor, Google Gemini CLI, GitHub Copilot and other AI agents can analyze your project and select particular skills from the local AAS catalog. AI Agent makes the final decision — AAS Core's read-only compose_stack tool verifies the selection in memory, and aas CLI stores it as aas-stack.json and displays unmodifiable plan before the action. Skill is about Git, testing, documenting, SEO, security, content creation or hundreds of other tasks. Started in 2024, AAS Core is a community effort with 46,000 stars in the AI coding ecosystem. It's free and MIT licensed.

Rion-Wu-tech/wechat-intelligence-hub

WeChat Intelligence Hub is a Codex skill to turn your local WeChat message history (Tencent’s chat app) into structured personal intel. It parses unanswered contacts, promises, sales leads, reconnections from group chats, creates intel reports in time ranges on demand. It consists of 2 parts — WeChat CLI reader (wechat-cli) and Intel Hub (wechat-intelligence-hub), packaged together. The reader queries the local WeChat database, which is accessible offline only. To query all data, you need to authorize local access (database key). There is a demo mode using fake data. It can be installed with a single prompt to Codex. Project README is mostly Chinese.

punkpeye/awesome-mcp-servers

Directory of Model Context Protocol (MCP) servers — an open standard for secure interaction between AI agents and local and remote data sources, developed by Anthropic. The catalog includes production-grade and experimental servers for files, databases, and external systems such as OpenAI API, Apple Shortcuts, Apple Reminders, Google Drive, Slack, Notion, GitHub, etc. It is compatible with Claude Code, Cursor, Claude Desktop, and other MCP agents. The implementations are available in Python, TypeScript, Go, Rust, and others.

Leonxlnx/unlazy

Unlazy is a completion discipline framework for AI agents that enforces structured verification through runnable gates and approval records. It uses an 'acceptance ledger first' methodology: write the expected outcomes before execution, then verify results against those criteria. Gate contracts are Markdown files combining a shell command (CHECK:), expected output (EXPECT:), and evidence tracking. Approval records are stored locally with bindings to specific gate definitions. An optional Stop hook for Claude Code blocks the agent from finishing until all gates pass. Supports parallel execution via --jobs flag (1-64 concurrent). Reached 3,000 stars in its first month after publishing — one of the fastest-growing skills in the Claude Code ecosystem.

kydlikebtc/awesome-grokbot

awesome-grokbot is a curated, daily-checked catalog of 624 live public Grok Bot configuration shares. Grok Bot by xAI gives AI agents their own always-on cloud computer with a browser and files; this repo is the searchable index of those configurations. Eight categories cover coding assistance (64 agents), research and briefings (110), content and publishing (91), inbox and calendar (30), finance and ops (45), personal admin (149), and team handoffs (98). The catalog includes bots that work with ad platforms such as Meta Ads and Google Ads, CRM lead imports, and outreach. Every link is automatically verified daily — expired entries are moved to a separate retired.json file. A companion search site lets you filter by category and search instantly without installing anything.

EverettFish/holo-card-studio

Holo Card Studio is a Codex skill that generates interactive 3D holographic trading cards from a single text description. The agent draws four image layers — background, line art, subject, and text — stacks them with depth parallax, adds a rainbow holographic sheen that shifts with viewing angle, a Voronoi star effect, and a gold border. The output is a Three.js browser page where the card can be rotated and flipped, plus a card.blend Blender project with Chinese-labeled material nodes you can adjust. All parameters live in card-config.json — rarity, number, attribute text — making batch production simple. No cloud, no uploads: Blender and Node.js on your machine do all the work.

Qwen3-TTS is the TTS model developed by Alibaba’s Qwen team. An online demo for this model is now open on Hugging Face Spaces, which supports 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian) and their local variations. It allows users to give textual commands for specifying pitch, sentiment, and tempo to synthesize a variety of tones (e.g., happy, relaxed, exciting). It has low end-to-end latency (~97ms), meaning that you will receive almost immediate feedback. It also uses ZeroGPU, so there’s no need to register or download anything.

vinzdg/codenotch

Codenotch is a macOS app that puts a small notch on one edge of your screen (top, bottom, left, or right). That notch shows a ring for each coding tool you use — Claude Code, Cursor, Codex, Antigravity — with how much of your usage limit is left, pulled straight from those apps' own data, no extra sign-in needed. If you run a separate work account for Claude Code, you get a second ring with its own limits. While a session is busy, a small arc spins inside its ring; if it's waiting on your input, the ring blinks instead. To install the Mac version from source: ask your agent to clone https://github.com/vinzdg/codenotch, run `brew install xcodegen`, then `make run`, following the README — you'll need a Mac with Xcode. There's also a separate Windows build with its own installer, but that agent step is specifically for building the Mac version.

okf-memory/okf-agent-memory

According to the developers, OKF Agent Memory is ‘a Domain-Neutral, Git-Native Persistent Project Memory for AI Agents based on the Open Knowledge Format (OKF) v0.2’. The Open Knowledge Format is a memory format created by Google. Without persistent memory, an agent forgets all architectural details and session state when a context window expires. OKF Agent Memory saves project memory as simple Markdown files with YAML front matter in a ‘knowledge/’ directory in your repo which get loaded automatically when you start a new session. It embeds an MCP server so you can run Claude Code or any other MCP compatible agent without the need for external database servers or services. It implements a Progressive Disclosure mechanism so that only relevant context information gets presented to the agent at any point in time. According to the developers, this results in an 80% reduction in token usage for long projects. It has an in-memory BM25 search index with a sub-300 microsecond response time. It works with local and cloud-based LLMs like LM Studio and Ollama.

NousResearch/hermes-agent

Hermes Agent is ‘The self-improving AI agent built by Nous Research.’ It works 24/7 on low cost cloud services ($5/month VPS, essentially free on serverless platforms like Modal and Daytona [hibernation mode]) and takes tasks from Telegram, Discord, Slack or terminal. Unique features include learning from each task performed and developing reusable skills to speed up future similar tasks, as well as building a persistent user identity. Compatible with any LLM backend (Claude, OpenAI, Nous Portal, self-hosted). 7 execution environments: Local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox. Open source (MIT), easy install (1 command on Linux, macOS, WSL2).

headroomlabs-ai/headroom

Headroom is a context compression layer developed by Headroom Labs. It sits between your AI agent and LLM and compresses tool output, logs, RAG fragments, files, history etc. before sending to the model. It compresses JSON 60–95% and code ~20%, but the answers are the same as if they were not compressed. It can be attached to your existing agent in a single command (headroom wrap claude), acts as a transparent proxy listening on port 8787 without modifying any code, is available as a Python/TypeScript library, and can be run as an MCP server. Works with Claude Code (Anthropic), Codex (OpenAI), Cursor, Grok, Copilot, Cline, Goose, OpenHands, and more. Shared memory store for multiple agents allows consistent context between Claude, Codex, Gemini, Grok, etc. headroom learn analyzes previous failed sessions and stores corrections in CLAUDE.md to prevent future errors.

The French robot company Pollen Robotics has published a web simulator of their Microduck humanoid robot using their pollen-robotics namespace on Hugging Face. This uses real reinforcement learning policies learned on physical Microduck robots, not pre-recorded videos. The control policies are implemented with a neural network that is run using onnxruntime-web in the browser, there is no need to upload your data to some server, it just runs locally in your browser tab. There are two different kinds of locomotion (wheels & legs) and 6 predefined behaviours (running, sitting, somersault, ball kicking, roller skating, squatting). If you open the web app in your browser you’ll see multiple visitors as transparent ghosts running around in the same scene at the same time.

VoxCPM2 is a speech synthesis model developed by OpenBMB, the creators of MiniCPM model series. It’s a 30 language model with 48kHz sound quality and trained on more than 2 million hours of multi-lingual speech data at 2B parameter. You can play around with the web demo here on Hugging face. There are three different modes available. Design a new voice from given text. (No Audio required) Clone a controlled voice from your own audio sample. Advanced mode which includes complete style control feature. Just record a small clip and type any text you want to be read, it will come out in that voice. Or if you don’t have audio recording you can describe the voice by typing its age, accent, tone etc, it will generate for you. No installation required and no need of a high end GPU.

rtk-ai/rtk

RTK (Rust Token Killer) is a command-line proxy, written in Rust, that sits between your terminal and your AI agent and compresses command output before it reaches the model's context. It applies four strategies — smart filtering, grouping, truncation, and deduplication — to 100+ supported commands: file operations (ls, cat, grep), git, test runners (pytest, cargo test, jest), build tools (TypeScript, ESLint), package managers, AWS CLI, Docker, Kubernetes. It plugs into Claude Code, GitHub Copilot, Cursor, Gemini CLI, Windsurf, and 12 other tools via hooks or plugins. Install with brew install rtk on Mac/Linux, or winget install rtk-ai.rtk on Windows — it's the simplest route there, no PATH setup needed. It's Apache 2.0 licensed, with 78,000 GitHub stars.

op7418/guizang-yingzao-skill

Yingzao is an agent skill for Claude Code and Codex that transforms photos of buildings, streets, cafés, artifacts, and local food into art-directed editorial posters. Unlike a photo filter, the skill performs a full design pass: the agent reads the photo's composition, selects a matching visual reference from a built-in recipe library, designs a Chinese display-type system and spatial layout, then sends the original photo, reference, and typographic guide to an image-generation model in a single call. Subject extraction, color field, and typography all exist in the same visual world — not layered on top. After delivery you can optionally extend the poster into a 9-frame video storyboard with a ready-to-use prompt for a video model. Requires Python 3.10+ with Pillow, NumPy, OpenCV, and fontTools, plus an agent that supports Skills and image generation.

deeplethe/utopia

Utopia is an open source enterprise knowledge graph built by DeepLethe in Rust. It parses documents, creates a bitemporal knowledge graph (meaning each fact has a “valid time” and “stated time”), and allows you to query the knowledge graph using an AI agent. It supports PDF, DOCX, PPTX, XLSX, CSV, Markdown, HTML, Notion, GitHub, Jira, WebDAV, RSS, and S3. It uses full text search (Tantivy), vector search (pgvector), graph traversal (RRF Fusion), and streaming results with inline citations. It has multi-user access control, an append only audit log, an ontology workbench, and forward chaining reasoning. It was released just one month ago but already has over 4000 GitHub stars. It’s currently at version 0.1 so the database schema will change with forward only migrations, so back it up first!

ahujasid/camera-to-blender

Use your smartphone camera to capture an object and see its 3D representation in your current Blender session within 60 seconds! (via Tripo3D’s AI based 3D generation) camera-to-blender connects your smartphone camera with Blender using AI based 3D generation. A small Python server is running on your computer. Open the web UI on your smartphone, look at any physical object and snap one picture. Optionally remove the background via Gemini API and forward the resulting image to Tripo3D for 3D mesh generation. In ~30–60 seconds you will receive the final result via WebSocket and automatically imported into your current Blender session. Alternatively you can use your laptop webcam. Small Blender add-on for receiving models.

affaan-m/ECC

ECC is a plugin for AI agents such as Claude Code, Codex, Cursor, etc. It includes 68 custom agents for planning, code reviews, security, and fix builds; 293 skills (prebuilt workflows) for TDD (test driven development), research, docs, etc.; 94 commands; and a memory system that remembers previous interactions. After installation, your AI agent will follow the ECC process of Plan, Test, Build, Review, Check, Remember when interacting with you rather than just diving right into coding. It also includes AgentShield, a tool that checks your hooks, MCP configs, and agent files for any potential hidden prompt injections or leaks. Install by typing /plugin install ecc@ecc in Claude Code, or by running npx ecc-universal setup in your terminal.

aaif-goose/goose

goose is a general purpose open source AI agent with a native desktop app (macOS, Windows, Linux) and CLI and embeddable API to write, research, and automate with Claude, OpenAI, Gemini, or a local model. goose is a native desktop app and CLI with embeddable API. It supports 15+ AI providers like Anthropic Claude, OpenAI, Google Gemini, Meta (via Meta’s Model API), Ollama, OpenRouter, Azure, and more using your API keys or subscriptions. goose can interact with 70+ MCP (Model Context Protocol) compatible extensions such as File I/O, Web, Databases, Email, etc. It is written in Rust for performance and cross platform compatibility. It is hosted by the Agentic AI Foundation (AAIF) under the Linux Foundation umbrella and is a fork of the original goose project started by Block. Download the native desktop app without setup or install the CLI with a one-liner.

The Hy4 preview version is a next-generation MoE flagship model developed by the Tencent Hy Team. It has 770 billion parameters and activates 49 billion parameters per token at inference time. This allows it to achieve the performance of a large model without having to compute all of its weights at once. It also has a 1M token context window, which is approximately equivalent to 750,000 words. This allows the model to process entire codebases or books at once. It also supports tasks like software engineering (such as understanding, planning and debugging long horizon development tasks), office work (including converting context into various types of documents, spreadsheets and financial models), game development (creating playable prototypes from textual descriptions) and scientific reasoning (in AI, physics and mathematics). It is licensed under the Apache 2.0 license. To run it, you will need 8 tensor parallel GPUs. It is a server side model.

Ornith-1.5 is an agentic coding model developed by ornith-ai using iterative self-improvement. It was trained by solving real codebase problems, getting feedback, and iteratively improving. Quantized to 4-bits, it’s 22 GB and the fastest coder in its category, equivalent to Claude Opus 4.6 medium on real codebase tasks. It has the highest quality multi-turn conversations of any local model. SWE-bench Verified: 79%, Terminal-Bench: 67–68%, MCP-Atlas (real agentic tasks w/ tools): 70.2%. Context window: 262K tokens (~1M with YaRN). Native tool calls, MCP, and OpenAI API compatible. peculiar-ragdoll has wrapped the original ornith-ai weights in GGUF for local use with Ollama, llama.cpp, etc.

Breeze TTS 2 is an open-weight text-to-speech model from BreezeBlue, ranking #1 among open-weight models on the Artificial Analysis TTS leaderboard and claiming to outperform several closed proprietary systems. It has 3.5 billion parameters and supports three use modes: Voice Clone (reproduce any voice from a short audio reference), Voice Design (describe a voice in natural language — age, accent, style — and the model generates one), and Voice Direction (clone a voice with emotional steering). The model also supports Vocal Events: expressive sounds like laughs, coughs, and sighs can be embedded directly in the synthesis transcript. Latency to first audio is under 40ms, making it suitable for real-time conversational use. Supported languages are English and Chinese. Requires a CUDA-capable NVIDIA GPU (12 GB for standard inference, 24 GB for fast mode), Linux, and Python 3.10+.

OpenHands/OpenHands

OpenHands Agent Canvas is a self hosted control center for developers to manage and run multiple AI agents. Rather than connecting to each agent through its own terminal or window, OpenHands allows you to aggregate them (or any agent supporting the Agent-Client Protocol) into a single web interface running on your computer. You can then chat with agents, build automations that respond to Github events or Slack messages, schedule when agents run, and easily switch between running agents locally, remotely, or in the cloud while maintaining state across sessions. Each backend can run agents in a sandboxed environment, and you can also host an agent server that you and your team can share and collaborate with (e.g. for an automated code review workflow). We’ve been open source for 18 months and have grown to 86,000 stars. We offer a paid SaaS solution in addition to our free, self hosted offering.

nexu-io/html-anything

html-anything is a local-first web application developed by nexu-io team. It can convert various kinds of data such as Markdown, CSV, Excel, JSON, plain text, etc., into a single HTML file with publishing quality using your own coding agent’s CLI. When starting the application, it will automatically detect Claude Code, Cursor Agent, OpenAI Codex, Gemini CLI, GitHub Copilot CLI, OpenCode, Qwen Coder, Aider. It also includes 75 skill templates which can be composed to generate HTML files for different purposes (e.g. magazine articles, keynote speeches, resumes, posters, Xiaohongshu & tweet cards, web mockups, data reports, Hyperframes video scripts). Your agent will stream the generated HTML content to the web application and you can export it with just one click. For example, you can copy the inlined CSS code and paste it on WeChat without further formatting; Or download a 2x PNG image directly to X or Instagram composer; Or save the generated HTML file locally as .html. You don’t need an additional API key, it will reuse your current API session (claude login/cursor login/codex auth). It has been released for 4 months, and now has reached 8 k★. Licensed under Apache 2.0.

n8n-io/n8n

n8n is an open-source workflow automation tool that has been around for 7 years with more than 200 k stars. It offers a visual workflow editor combined with the ability to add custom JavaScript or Python code where necessary. There are 1,500+ integrations and 9,000+ community workflow templates to visually connect your services, e.g. start a workflow on incoming emails or new entries in your CRM system or trigger it via a webhook. Use conditional logic, transformers and AI agents to decide the next action based on the input data. Built-in AI capabilities allow you to use any OpenAI compatible model (e.g. Anthropic, Google or even your own locally running Ollama models) inside a workflow step. It can be self-hosted by running just one Docker command, but also offers the n8n cloud. Source code is always available, but we also offer additional enterprise features under a commercial license.

mlc-ai/web-llm

WebLLM is an open-source, high-performance in-browser LLM inference engine from MLC AI, powered by WebGPU for hardware-accelerated computation. Language models run entirely inside the visitor's browser — no server is required, no data is sent anywhere. It is fully compatible with the OpenAI API, so existing code using openai.chat.completions.create() requires minimal changes to run locally. Supported model families include Llama 3, Phi 3, Gemma, Mistral, and Qwen. It installs as an npm package (@mlc-ai/web-llm) and runs computation in a Web Worker or Service Worker to keep the UI responsive. A live demo at chat.webllm.ai lets you try it without any installation.

GLM-5.3-Flash is a natively multimodal model in ZhipuAI’s GLM-5 series, which unlike other GLM models, handles both text and images in a single API call. It follows a Mixture-of-Experts approach with a total of 320 billion parameters, and 18 billion being used per call. It can handle up to 300,000 tokens of context, which is more than enough for multiple long documents. It outperforms the previous GLM-5.2 model on all benchmarks. Ideal for agentic purposes, code generation and image-based document analysis. This GGUF is available for Ollama, but it needs to be run on high-end servers with hundreds of gigabytes of RAM.

PhiloLabs/fable51-worlds

PhiloLabs dispatched autonomous Claude Fable 5.1 agent swarms to reconstruct real city districts entirely from open geographic data and public reference imagery. The pipeline runs in four stages: parallel research agents gather OpenStreetMap geometry, USGS elevation data, transit specs, and a per-fact storefront census; offline scripts generate optimised 3D asset kits; a pure Three.js runtime assembles terrain, streets, crowds and traffic from JSON specs; and Playwright-driven QA agents screenshot fixed viewpoints and diff them against free-licensed photographs to catch errors. Two worlds are finished: Union Square, San Francisco (453 OSM buildings, 75 hand-authored façades, 129 named storefronts, 220 pedestrians on a 1,398-node nav graph, Powell Street cable cars, and two explorable interiors: Apple Union Square and Nintendo SAN FRANCISCO) and Higashiyama, Kyoto (266 buildings, 2.3 km continuously walkable route climbing 76 m from Gion to Kiyomizu-dera, rendered as a hand-painted anime background). No game engine, no proprietary 3D tiles — every element is generated from code.

omdsh-dev/DSH-better-sidebar

DSH Better Sidebar is an open-source plugin for DeepSeek Harness (DSH) that replaces the minimal default sidebar with a full development workbench. It adds a file explorer with a CodeMirror editor, an embedded multi-tab browser in a sandboxed iframe, a real terminal (xterm.js + node-pty) with reconnect replay, a Git panel with visual two-way diff and commit/revert controls, a Codex-style side chat thread that runs on the main session's full context, and a sub-agent task view with live output and kill controls. Layout is dual-pane — right sidebar plus bottom panel — with tabs draggable between panes or tearable into floating windows. The plugin exposes a `betterSidebar` API so any third-party plugin can register its own tabs; the ecosystem already has 28+ extensions. Released a month ago and already has 3 200+ stars.

lnkiai/m3e-canvas

M3E Canvas is a visual UI editor that helps you create Material 3 Expressive screens using drag-and-drop components such as buttons, cards, nav bars, dialogs, text input fields, etc. It also enables you to link the screens together using tap/swipe animations (slide, fade, expand) and describe the navigation flow by tapping through the screens. It doesn’t require any backend database — data is stored in localStorage and the app is fully client-side (no installation required). You can generate prompts in English, Japanese and Chinese. Check out this live version of the app: lnkiai.github.io/m3e-canvas/. Feel free to use this tool without cloning it.

codejunkie99/fable-orchestrator

Fable Orchestrator is a local-first routing skill for Codex that adds Claude Fable 5.1 as a dedicated planning and adjudication layer. Claude never writes code or owns the workspace — it builds a plan, decides which model handles which work, and adjudicates when decisions arise. GPT-5.6 Luna handles standard implementation; DeepSeek V4 Flash takes repeated loops and high-throughput iteration. Every implementation node must use one of those two routes — if neither is reachable, the workflow reports the blocker instead of improvising a solution. Install runs from the repository root with ./install.sh --copy, landing the three skill files in ~/.codex/skills/fable. Invoke with '$fable <objective>' from any Codex task.

anthropics/commerce-agents

Commerce Agents is Anthropic's reference implementation of two types of AI agents for stores: a Shopping Agent, which lives on a store website, and a Merchant Agent, which helps merchants manage their business. A Shopping Agent helps users shop by searching, comparing products, adding items to a cart, answering questions about orders and policies, and remembering user preferences across sessions. It does all the work of getting the cart ready for checkout, but the actual checkout process happens outside of the agent (the store is responsible for placing an order and charging a credit card). A Merchant Agent helps merchants understand their sales data, update product listings, adjust pricing, and run promotions. All actions taken by the Merchant Agent are delayed and must be approved by a human before they take effect. There are four demonstration applications included: Retail, Travel, Telecom, and Entertainment. Also included is a Claude Code Plugin for generating custom agents based on user descriptions. Ask your agent to ‘Deploy the demo from https://github.com/anthropics/commerce-agents, read the README, launch the retail vertical with python scripts/run_demo.py retail, and ask me for the ANTHROPIC_API_KEY’. Requires Python 3.11 or higher, Node 22, and a Claude API key. No GPU needed.

MiniMax H3 is MiniMax's video model that generates a clip with a synchronized audio track in a single pass — sound is produced alongside the video and matched to the on-screen action, not added manually afterward. Previously reaching this model in a browser required two separate services: one to encode the text prompt into embeddings and another to run the actual video generation, with a handoff delay and a separate queue between them. This community-built space by mrfakename combines everything on one page with no intermediate steps. Generation speed is roughly 1.5× faster than the previous split deployment: a 56-frame clip takes about 23 seconds versus 35. The space runs on cloud GPU at no cost — no registration or installation needed.

Qwen3.8-27B is a 27B parameter model developed by Alibaba. It is one of the best open models for reasoning, math, and coding. The ISTA-DASLab (Austrian Institute of Technology) team has published a GGUF version of this model using non-uniform quantization. This means that each tensor is quantized individually based on its sensitivity to maximize the model quality while keeping the model size small. This version of the model comes with the integrated vision projector, meaning it can handle images, screenshots, diagrams, and photos. The model size is between 8.4 GB and 11.8 GB depending on the quantization. This should fit on any laptop with at least 16 GB RAM.

subsy/skill-cabinet

Skill Cabinet is a local web UI to scan through all skill directories on your system (including .claude, .codex, .cursor, .agents, Hermes profiles, and paths like ~/.*/skills) and display all discovered skills in one location. Alongside the skill body, you’ll see the YAML metadata, where it was sourced (on GitHub), and if it’s a file, symlink, or non-existent. Duplicate skills found across agents are displayed, along with high-risk patterns in skill bodies. Skills can be quarantined, removing them from all running agent drawers while leaving them on disk (where they can be restored). Multiple skills can also be deleted in bulk. Run with npx skill-cabinet (requires Node 20+) to automatically open it on port 3781.

stablyai/orca

Orca is a desktop application for Mac, Windows, and Linux that helps you manage and coordinate many AI code agents (Claude Code, Codex, Pi, OpenCode) working on different tasks at the same time, in separate git worktrees. Fan out one or more prompts to your agents and see the results side by side, then merge the winning solution. The mobile app (iOS and Android) lets you check the status of your agents and give them additional directions from your phone. Use Design Mode to click on something in a running Chromium browser and send the HTML, CSS, and screenshot right to your agent’s prompt so you can easily make changes to UIs. Orca was released six months ago and has gained ~60k stars.

Ryze-AI-Adgent/open-seo-mcp-skills

Open-source SEO and GEO skills for Claude (MIT license) which use the Ryze MCP connector. Eight skills extract your own data from these accounts (Google Search Console, to get actual positions and traffic clicks, GA4 to get overall traffic and AI referrals from ChatGPT, Perplexity, Claude, and Google Ads for keyword volume and cost per click). It also uses already integrated DataForSEO service for competitive research, keywords and SERP data. To get started, you’ll need to install 2 things: Ryze MCP connector in Claude settings and Claude Code plugins. All data is extracted without data markup, directly from your accounts.

Player-YN/PawWork_ZhuaZhua

PawWork is a Chrome MV3 extension that interacts with any page you have open. Enter Paw Mode, select an element of the page (data table, text, image, or hyperlink) and tell the AI what to do using the sidepanel. It will return live spreadsheets (Univer), drawings (tldraw), editable documents, or entire html websites based on the element(s) you have selected. If your task is complicated, it will offer a plan for approval first. The resulting files will be saved locally to the IndexedDB and OPFS storage. It won’t upload anything to the web. The code is run in an isolated WASM sandbox. You will need to provide the model key yourself. We support OpenAI, Anthropic, etc.

open-webui/open-webui

Open WebUI is a self hosted AI interface. It has a rich set of features to interact with AI models. Connect to local models using Ollama or any OpenAI compatible API (e.g., OpenRouter, LMStudio, GroqCloud, Mistral, vLLM, …). Upload files & images. Use voice input. Perform RAG (search web, documents). Manage your models. View fine grained user roles & permissions. Extend functionality using plugins/tools. Install using Docker, Pip, Kubernetes. 150,000+ stars in 3 years. De-facto standard for self hosted AI interfaces.

harry0703/MoneyPrinterTurbo

MoneyPrinterTurbo is an open-source, all-in-one AI short video generator. You provide a topic or keyword; it generates a script using an AI language model, matches HD stock footage from Pexels, Pixabay, or generated AI video (MiniMax H3, Seedance, Wan), adds an AI voiceover and synchronized subtitles, and produces a finished video. Output formats include 9:16 portrait for TikTok/Shorts, 16:9 landscape, and 1:1 square. Finished videos can be published directly to TikTok, Instagram Reels, and YouTube Shorts. It works with a wide range of AI providers for scripting: OpenAI, Anthropic Claude, Google Gemini, Kimi/Moonshot AI, DeepSeek, xAI Grok, and over a dozen more, as well as local Ollama models. The tool runs on Windows, macOS, and Linux via Python 3.11+. 120k+ GitHub stars.

MiniCPM-V 4.5 is a multimodal language model from OpenBMB (a research group at Tsinghua University), built on Qwen3-8B as its base. It understands single images, multiple images at once, and high-FPS video — all in one model without separate pipelines. Video compression works efficiently: six frames are packed into just 64 tokens instead of the typical 1,536, so long clips fit within the context without truncation. The model handles images up to 1.8 million pixels at any aspect ratio. On OCR benchmarks it outperforms GPT-4o-latest and Gemini 2.0 Pro. Context window is 40,000 tokens, and the download is 6.1 GB. Available via Ollama: `ollama run minicpm-v4.5`.

Hugging Face Space with Wan-AI’s Wan2.2 image-to-video model (14B params) deployed using ZeroGPU in the browser. Upload an image and provide some textual description about how you want the movement to be, and it will generate a small animation video for you! No need to install anything, no GPU required. This space utilizes AOT (ahead-of-time) compilation to achieve significantly better inference speed compared to regular deployment of this model. You can control the number of inference steps and the length of the generated video.

PhoneLLM Alpha 1 is a 32B open-weights model built by Pipecat for voice agent applications. Voice doesn’t work well with most language models because of its slow first token latency, multi-turn context loss, and inconsistent tool invocation during conversation. PhoneLLM overcomes these challenges with low latency, robust multi-turn dialog, and accurate tool invocation. It is said to be competitive with GPT Terra 5.6 while being 94% cheaper and 1.3 seconds faster (P95 time to first token). Licensed under the BSD-2 License.

thedotmack/claude-mem

Claude-Mem is a persistent memory compression system for Claude Code. It automatically logs tool use experiences from past sessions, creates semantic summaries, and inserts them into new sessions. This allows Claude to remember information about your project across multiple restarts. You can use cmem.ai (a free trial cloud service), your own OpenRouter, Gemini, or Anthropic API key, or OpenCode, and Anthropic CLI. It comes with a live web UI, skill command for natural language history search, and <private> tag to prevent saving sensitive information. This project now has more than 92,000 stars!

Leonxlnx/taste-skill

Taste Skill is a transferable design guideline for AI coding agents to avoid creating boring user interfaces. It’s an ‘Anti-Slop Frontend Framework’, which is essentially a list of rules that an agent can read and follow to make sure they don’t fall into common traps. The framework has 3 knobs to dial in how much variation there should be in the layout (centered & clean or asymmetric & modern), animation (hover effects only or scroll/magnetic), and visual weight (sparse or dense dashboards). Included in the repo are different versions of the skill such as the default v2, a notion/linear inspired minimalistic version, an industrial brutalist version, an expressive high motion version, a redesign skill to improve an existing codebase, and some image generation skills to create a visual board first and then send the frames over to your AI coding agent to implement. This works for React, Vue, Svelte, and any other major coding agent.

JuliusBrussee/caveman

Caveman is a skill for Claude Code and 30+ AI coding agents that teaches your AI to avoid unnecessary words like preambles, summaries, and explanations. In our benchmark tests on 10 different tasks, we found that it reduces the number of output tokens by an average of 65%. For example, it reduced 15,704 tokens down to just 121 tokens (a 130x compression ratio). Caveman works either as a transparent proxy or as a Claude Code plugin. We’re upfront about its limitations: it only affects the number of output tokens, and it introduces about 1–1.5k additional input tokens per turn, which means that for shorter conversations it may actually increase the total number of tokens. Its primary advantage is that it makes interactions quicker and easier to read.

GLM-5.3 is an open-weight language model developed by Zhipu AI(Z.ai) for the GLM series. It achieves the state of the art performance on all coding benchmarks such as Terminal Bench 3.0 and DeepSWE and beats all previous open-weight language models. All its improvement comes from the post-training stage instead of model architectures. We focus on its ability to serve for agents and manage large codebases. It has a context size of one million tokens, and it is available in both English and Chinese through vLLM, SGLang and OpenAI-compatible APIs.

Plugged it into an OpenAI compatible API, and wow, the improvement over older Qwens is clear. Qwen announced a new 180B multimodal model which takes both text, images, and videos as inputs with a built-in 262K token context length (expandable to 1M). GPQA Diamond achieved 91.7%, SWE-bench Pro achieved 62.5% — I plugged it into my agent and suddenly tasks that used to stump him are working perfectly. Note that this is still an experimental architecture release, and the license is qwen-community instead of Apache so make sure to check your legal department.

I played around with the Kimi API, which gives you 1 million tokens of context so you can throw in an entire codebase and let it do its magic. Moonshot AI just might have published the biggest open weight multimodal model ever: 2.78 trillion parameters, handles text, images, and videos all at once. It’s accessible via an OpenAI compatible API, so your agents don’t need any code changes. Integration with HuggingChat is already available for free web use.

I tried it out on CPU, and it seems pretty quick (<0.5 GB memory usage, looks good). Audio8 published a small TTS model with zero-shot voice cloning abilities: simply input some text and 3–5 seconds of reference audio, and it will output speech in the desired voice. Supports multiple languages. Free for projects with less than $2M in annual revenue, perfect for most independent developers.

Unsloth packaged Alibaba's Qwen3.8-Flash-Next as quantized GGUF files for local deployment. The original model, built by the Qwen team, is positioned as an architectural preview of what Qwen4 will bring — unlike ordinary multimodal models that simply describe images, it reasons over visuals as part of an extended conversation. Native context window is 262K tokens, extensible to 1M. Handles agentic reasoning, software engineering tasks, office automation, and image-based question answering. Runs via Ollama, llama.cpp, or LM Studio — pick the quantization level that fits your hardware.

saurabhkumar8112/cyclomatic-complexity-skill

This is a Claude Code plugin that’s optimized for AI-generated code. It evaluates the complexity of every function and locates the worst hotspots for automatic refactoring, simplifying complex conditionals with straightforward logic. At the end of each iteration, this skill creates a table of before-and-after code complexity scores to help you understand which functions have been improved. Supports all languages that complexity analysis tools support and will follow your project’s linter rules if configured.

openclaw/openclaw

OpenClaw is a free and open source local AI Assistant. It communicates in channels you already know and love like WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, etc. The Gateway acts as the main service in charge of managing sessions, tools and connecting channels. You can pick any model you want and configure channels and tools to suit your needs. Optional apps provide additional functionality, like voice interaction, canvas, camera, screen and local action triggers. Built by the OpenClaw Foundation and the Community. Available on Mac, Linux and Windows. Installs with a single command using curl or powershell.

mouredev/hello-sdd

The code for the course on Spec-Driven Development (SDD) with AI agents from Spanish dev teacher mouredev. The basic concept is to first write a structured spec before asking an AI coding agent to do something, so that it has a plan to follow rather than winging it. We provide three reusable template files: AGENTS.md (info the agent should know about the project, like Claude.md), spec.md (with numbered functional requirements in EARS format, edge cases, etc., out-of-scope, and acceptance criteria), and prompts.md (with pre-written prompts for all phases of SDD: constitution, spec, clarification, planning, tasks, implementation, validation, change). An example project worked through using SDD, end to end. We also provide a reusable spec generator skill for Claude Code, at .claude/skills/spec-generator/, which asks you questions about your project and then automatically generates a spec.md based on our template. Finally, we provide a link to the opencode version of the spec generator skill at .opencode/skill/.

jub0t/Concat

The free, open source alternative to CapCut with no ads, no sign ups, no subscriptions, no pay walls. Built with Rust for all your media processing and React for the editing interface. Work locally with this offline video editor! It has AI-generated auto-subtitles with Whisper and its own text to speech feature. It supports multiple tracks, cutting, splitting, merging, transitions, time stretching, voice effects, styled text, and templates. Its available for macOS (unsigned binary — run one command to remove quarantine from README), Windows, and Linux. While it is still in Alpha (version 0.2.0) the functionality is there but it will be a little rough around the edges.

hkqr/my-free-code

My Free Code is a local gateway for routing Claude Code & other code agents to AI providers. Your code agents communicate with the gateway using the usual Anthropic/OpenAI protocols, which then routes them to the corresponding AI providers such as Anthropic, DeepSeek, Groq, xAI, NVIDIA NIM, OpenRouter, etc. The tier routing (e.g., Fable, Opus, Sonnet, Haiku) is done via env vars (e.g., MODEL_SONNET=deepseek/deepseek-chat). When a provider is down, it will fallback to the next provider in line without repeating the previous responses. It also comes with a web ui (/admin), a token count api, model discovery, auth, & provider health checks. Built with python 3.10+, run at http://127.0.0.1:8082.

google-gemini/gemini-cli

A free, open-sourced AI agent developed by Google, for your terminal. With a free tier of 60 requests/minute, 1,000 requests/day via personal Google account (no payment needed), a 1M token context (meaning you can keep an entire codebase in your memory all the time, without breaking it down), built-in features like Google search, file manipulation, shell commands, and web request, and the ability to extend the agent with your own integrations through MCP (Model Context Protocol). Completely open-source under Apache 2.0.

browser-use/browser-use

The Browser Use Python library lets you fully control a web browser like a human. This means that you can open websites, click elements, type text, fill forms, and scrape structured data. You can specify tasks in natural language and let your AI agent take care of the rest. You can easily install this library as a skill for Claude Code, Codex, Cursor, and other agents. All you need is one prompt. Once installed, you can direct your AI agent to perform tasks in a web browser such as those in the README: fill out a job application based on a resume, get a list of followers in CSV format, look up and compare products, and track prices. You can use this library with any LLM by calling its API (e.g., Anthropic, Google, OpenAI, Browser Use Cloud). This library has 111k stars and is maintained by browser-use.com.

stemdeckapp/stemdeck

StemDeck is a free, open-source desktop application that splits audio into up to six stems — vocals, drums, bass, guitar, piano, and other — entirely on your machine using Meta's Demucs AI model. Drop in an MP3, WAV, FLAC, OGG, MP4, or M4A file, or paste a YouTube URL, and StemDeck handles the rest. A DAW-style waveform editor lets you mute, solo, and adjust levels for each stem, loop a region, and export individual tracks or a custom mix. On first launch the app downloads the Demucs model (~170 MB) and Python runtime, then works fully offline afterward. macOS and Windows builds are available as ready-to-run installers — no Python or system dependencies required.

XiaoDuoYa/codex-with-chatgpt

This skill enables ChatGPT Web UI by OpenAI to be a planning and reviewing assistant during programming, and let Codex take charge of execution. No need to get API Key from ChatGPT, it connects with ChatGPT official Web UI using MCP OAuth Bridge securely and read-only mode. Also, there is no need to upload your repo to ChatGPT, ChatGPT will only read part of your code on your workspace when needed. It can auto upgrade everyday via github. Auto install dependencies, build and setup init script by pasting it to Codex. Only need manual login to chatgpt or input 2fa code. Both Chinese and English documents available.

tt-a1i/simplify-codebase

A skill for Codex/Claude Code to identify and safely eliminate accidental complexity from code bases. Deleting lines of code is just an effect, but the real value is eliminating facts/states/contracts/concepts that don’t need to be maintained any more. This skill will generate a proof report before deleting anything (how much maintenance cost it introduces, which consumers (e.g. production/test/dynamic/external) will be affected, what behaviors will be lost, and the simplest test case to detect incorrect deletions). There are 2 working modes: Survey (read-only analysis of potential candidates with ranked evidence) and Change (apply verified simplification changes with batch-level verification). This skill respects public interfaces/dynamic imports/data formats/migrations/architecture rules as important evidences. If there’s an actual consumer or boundary to resolve, this skill suggests keeping things instead of blindly deleting them. Main readme (Chinese), English readme (README.en.md).

OnlyTerp/opengrok

OpenGrok is a Python package which provides a model selector for Grok Bot. Each provider mapping has a corresponding wire capture stored in this repo. We follow an ‘evidence or it doesn’t ship’ philosophy. The install command will scan for an existing Grok Bot install and either reuse the bindings or prompt you with 3 questions. It then generates a config file. You can easily try out a binding using the model selector and clicking ‘save’. There is also a doctor utility which takes a baseline of your system during install and keeps track of files and services. Post Grok Bot update it will tell you exactly what changed and what still needs applying. Hot reload provider mappings. Cross platform (Windows/MacOS/Linux). Zero dependencies other than python.

oboroge0/hayamimi

Hayamimi (which means ‘quick ear’ in Japanese) is a python based tool that provides realtime multilingual speech to text transcription using CPU only, without using any cloud API or GPU. It uses quantized (INT8) ONNX models to redirect utterances to appropriate models based on their language. Each of Japanese, Chinese, Korean, Cantonese, English, 24+ EU languages has their own specialized model, while others are redirected to Meta’s Omnilingual ASR. Achieved 5.8% character error rate on real broadcast Japanese audio, which is less than half compared to Whisper large-v3-turbo (13.8%), at realtime 10–50x. Subtitles are partially displayed during your speech (~0.5 sec interval), finalized subtitles (~100ms delay) are displayed when you finish. Live subtitles are displayed on web dashboard. Speaker labels (e.g., S1, S2) and translations are available too. OBS Overlay for streaming caption is also supported. Low memory usage (< 2GB). Python 3.10 or higher, ffmpeg, ~3.1GB of downloadable models (1.1GB for minimal japanese+english installation)

Nanako0129/sepia

Sepia is a portable agent skill for Claude Code, Codex, Grok Build, and Antigravity that repairs AI-generated text at the layer research shows actually gives it away. A 2026 study (StoryScope, 61,608 stories) found that a classifier using narrative-structure features alone detects AI fiction at 93.2% accuracy — swapping word choice barely moves that number. Sepia targets the structural tells: narrator explaining the theme, single-track causal plots, emotions rendered only as bodily sensation, resolved endings, linear time. For fiction it runs a three-pass protocol: narrative architecture first, then discourse flow, then surface style. For professional documents — release notes, PR replies, postmortems, tickets, technical articles — each type gets a dedicated thin rule file on top of a shared checklist. Four operations: write, review (diagnose only), refactor (minimal edits), recreate (full rewrite). Installs via the Claude Code or Codex plugin marketplace in one command.

leopard627/fire-your-seo-agency

This is like an SEO agency but an AI Agent instead of Human Contractors. It analyzes your website’s search engine optimization (SEO) across 5 different channels(Google, AI Overviews(Google, Bing), ChatGPT/Perplexity, LLM brand knowledge(LLMs), Naver). And gives you a scorecard with solid data and evidence for each channel instead of vague suggestions. Then it solves your technical SEO problems, implement llms.txt and JSON-LD, and create web pages based on “one question = one page” rule. Finally, it sets up another measurement and shows you the results with numbers. I tried this approach on a real Korean stock research website and got 1.54 million search impressions within 30 days without any advertisement cost.

KKKKhazix/sun-style-writing

A .skill file for Claude Code, Codex, Kiro and similar agents that teaches 10 storytelling techniques pulled from one well-known Chinese-language text. The techniques: show-don't-tell description (action instead of judgment), contrast (numbers next to numbers), repetition with variation, plot hooks and callbacks, object narration (letting objects speak instead of people), imagery, time-montage, silence (stopping where the character stops), anticlimax, and circular structure. After installing, the agent writes more concretely and stops replacing real scenes with generic emotional phrases. To install, ask your agent: "Install the skill from https://github.com/KKKKhazix/sun-style-writing — find the folder with the .skill file in the repo and install it, then write a test paragraph about any life event in this style so I can see the difference."

kacperkapusciak/goldie

Goldie generates App Store screenshots & preview videos from iOS simulator. It takes screens with the argent testing framework, puts each screen inside a device bezel (iPhone 17 Pro or others), places a background and a heading, stitches all of them together into a preview video and validates the final output against the App Store guidelines. Supports SwiftUI, UIKit, Flutter, React Native, Kotlin Multiplatform apps. In the studio, select fonts (Merriweather, DM Sans), layouts (editorial, showcase, magazine) and background colors. Key feature: use Claude Code or Cursor to invoke Goldie with an easy text prompt, so that the agent will execute it and deliver all the files to you automatically — no terminal needed! Screenshots are 1320 x 2868 px (for 6.9” screens), preview videos are 886 x 1920 H.264, 15–30 sec long.

WAN 2.2 is Alibaba's open-source video generation model — one of the most capable models for converting images and text into short videos. Saravutw built a browser-based Hugging Face Space around it with Lightning optimization: generation runs in just 4–8 steps instead of the standard 50, making it dramatically faster. You upload a photo, describe the motion you want, and the model produces a short video clip. No account, no installation, no local GPU required — it runs on Hugging Face's Zero GPU free tier. The space has 247 likes and 5 community discussions, making it one of the most active WAN 2.2 browser demos.

OmniVoice is a voice cloning & text-to-speech platform developed by k2-fsa supporting more than 600 languages. Upload a sample recording of someone’s voice, write some text and get synthetic speech generated in their voice and language using the same. Built using HuggingFace Zero (GPU free tier) and deployed through Gradio, there’s no installation required. This model also has the ability to control the pace and length of the generated speech. We use a blend of Qwen3-0.6B for text encoding, OpenAI Whisper for reference transcribing and our very own OmniVoice model for synthesis, amounting to roughly 1.6B parameters in total.

Qwen3.8-27B is a strong multimodal model developed by Alibaba, but it tends to take many steps to reason through each response. This makes it great for analytic situations, but a bit too much for general conversation. DavidAU finetuned this model with Cold Fusion + GAIN to reduce its reasoning steps by 50%–90% without sacrificing accuracy. MTP variations can be used to further boost the token rate to over 90 tokens/sec. It is distributed as GGUF and runs locally on LM Studio, Ollama, and llama.cpp. A context size of 256,000 tokens (up to 1 million). You can adjust the number of steps taken during each request based on your desired level of speed vs. accuracy.

totec448-spec/chat-on-steroids

Chat On Steroids is a Windows, macOS and Linux desktop application to transform ChatGPT into a working agent, able to access your local files, execute commands on terminal and (on Windows) manage your screen, mouse and clipboard. The app hosts a local MCP server and connects ChatGPT to your machine using OpenAI’s Secure MCP Tunnel. You interact with ChatGPT as usual in the browser but now with this app running in the system tray as a permission management layer. You can specify which directories you want to give access to, allow or disallow commands and enable/disable capabilities. The app comes with a built-in Chrome extension to record sessions, implement Compact & Resume and leverage an experimental multi-agent goal loop (OpenRouter).

StarTrail-org/PixelRAG

PixelRAG is an open source visual retrieval system developed by Berkeley SkyLab and BAIR. Rather than extracting text from webpages, it turns the webpage (along with other document types such as PDFs and images) into a set of screenshot tiles that can be queried visually. This allows a vision model to answer questions about the webpage’s visual layout, tables, charts, and infographics, which are often lost when the webpage is parsed into text. PixelRAG comes with a pre-built hosted index of 8.28M Wikipedia articles that you can query for free on pixelrag.ai without needing an API key. If you’re using Claude Code, there’s also a pixelbrowse skill available in the repository that provides a /screenshot command. When you run this command, Claude will render the webpage using Playwright, then interpret it visually as an image. This means Claude will be able to read charts and tables like humans do! There is no need for an MCP server or backend, nor any GPUs; the skill simply invokes a local installation of the pixelshot command.

guillaumemeyer/watermarks-remover

watermarks-remover is an agent skill and local Python service for removing AI watermarks from text and your data. It works by removing non-deterministic (statistical) and deterministic (Unicode character/exotic space/bidirectional control) watermarks from your text/files. Supported AI watermarks include those used by Anthropic (Claude), Google (Gemini/SynthID-Text), OpenAI, and open-source LLMs. It also removes C2PA, EXIF, XMP, and document metadata from various file types such as PNG, JPEG, WebP, PDF, DOCX, XLSX, PPTX, EPUB, HTML, Markdown, MP4, WAV, FLAC, and more. It can be installed as a Claude Code plugin through the marketplace and uses a local service via an HTTP endpoint (no Python required on the agent host). It also automatically hooks into PostToolUse with a hook for each Write, Edit, and MultiEdit command.

amosblomqvist/learn

Pi is an open-source AI coding agent CLI (pi.dev) that supports Claude, OpenAI, and Google models. The amosblomqvist/learn repository is a ready-made .pi configuration directory that turns Pi into a structured teaching agent. It includes a teaching philosophy skill, a quiz extension that poses graded questions after each lesson block with instant feedback, a visualization extension that generates SVG and Mermaid diagrams when a concept is clearer as a picture, and a researcher subagent that cross-checks facts before they appear in the lesson. Installation is one git clone into any project folder — the repository becomes the .pi config and Pi picks it up automatically. Created and shared by a developer as his personal learning system, with the full configuration open to edit.

Laguna XS 2.1 is a coding-focused model from Poolside AI, a company that builds AI specifically for software engineering. The architecture is mixture-of-experts with 33B total parameters and only 3B active at inference time, which keeps memory usage manageable despite the scale. SWE-bench Verified score of 70.9% reflects solid real-world coding task performance. The model reasons natively between tool calls and supports a 262,144-token context window — enough to hold an entire project in one session. Poolside designed it specifically to run on a local machine. The Q4 quantized version weighs around 20 GB.

SenteLabsAI/OpenExecutive

OpenExecutive is an open-source AI system developed by sentelabs.ai that provides a virtual executive team for your company. Eight specialist agents work behind the scenes — Chief Strategy Officer, CFO, CHRO, General Counsel, COO, CMO, CPO, and Board Communications Director — each drawing on a built-in MBA knowledge base and documents you upload about your company. You see one coherent executive voice; the internal agent architecture is never exposed. The system maintains episodic memory of past decisions across sessions, so the next conversation starts knowing what was recommended before. A built-in scheduler can proactively surface follow-ups and time-sensitive actions. Requests go to the Anthropic Claude API; requires an API key. Runs locally via Docker; production deployment on Fly.io uses Google OAuth with an email allow-list to restrict access.

Ornith 1.5 35B A3B is an ensemble of experts model for code development by ornith-ai. Although it has 35B parameters, it only uses 3B parameters at inference time and so is actually much faster than a fully dense 35B model. Ornith 1.5 35B A3B achieves 79% on SWE-bench Verified, which is the first time a local model has surpassed this score. With its 262,144 token context length, Ornith 1.5 35B A3B can handle all your code at once in a single session. Out of the box, it supports tool invocations and integrates seamlessly into a code developing agency.

MiniMax H3 Turbo, developed by MiniMaxAI, a top Chinese AI company, is a video generation model that can generate short videos based on text or images. Now, the official web demo has been released by MiniMaxAI on Hugging Face Spaces. It’s hosted by MiniMax’s servers so that no local GPU and installation is needed. There are several styles for you to customize the appearance of the generated videos. If you need a quick preview about the performance of H3 Turbo without installing ComfyUI and server resources, this is your choice.

Tencent/WeMM-Embedding

The WeChat Vision team at Tencent has introduced a series of universal multimodal embedding models: WeMM-Embedding. These models unify text, image, video, visual document and interleaved multimodal information into a single vector space for cross-modal search. For instance, they allow users to retrieve relevant video snippets or documents based on text queries, or find images similar to user-uploaded images. There are three variants of WeMM-Embedding, each with 2B, 4B and 9B parameters. All the model weights are released on Hugging Face platform, named as tencent/ . You can easily use these models with standard transformers or sentence-transformers. It also supports production deployment with vLLM and SGLang. With a Matryoshka embedding mechanism, it enables you to project a larger embedding vector (64–4096) to a smaller one without re-training, by sacrificing some performance for saving storage and inference time.

s0xDk/refactoring-ui-skill

A Claude Code skill that encodes the concrete design rules from Refactoring UI by Adam Wathan and Steve Schoger — the designers behind Tailwind CSS. It gives Claude fixed scales for spacing, typography, colors, and shadows instead of arbitrary values. Claude learns to build hierarchy through font weight and color rather than size alone, and can diagnose vague UI complaints like 'looks cheap' or 'feels off' into specific mechanical fixes. Every rule was cross-checked page by page against the book. Includes a symptom-to-fix reference table, color palette building guides, depth and light simulation techniques, and a complete contrast-verified CSS token starter set.

ChenShuo2004/cs-board

cs-board is a free local AI video production application that runs on Windows. It is developed by Chinese group ‘有温度’(You Wen Du). With cs-board, you can upload a 10–30 second sample clip to create a clone voice, copy a Chinese script, select one of the 12 template (ranging from simple whiteboard style to vintage collage or cyberpunk), then the program will segment the script into scenes, generate images, animate the images using hand-drawing effect, synchronize the subtitle, and finally generate an MP4 video file. All data related to the project, including API keys and final video files, are stored locally. Tasks can be shared between different machines on the same network. Clone voice is done using the local IndexTTS 2.5 server, while image and text generation are performed via OpenLux API.

alchaincyf/huashu-excel

huashu-excel is a Claude Code/Cursor/Codex skill which can be installed on any agents that support skills format. This skill will equip your agent with an entire workflow of excel analysis, including a preliminary structured diagnosis of the excel file to identify issues such as header error, merged cells, subtotal rows in the body, duplicate entries, numeric values presented as text, and different date formatting. Then the skill will clean the data and compare the calculated sum with the subtotals in the excel file. Any difference will be reported to you with clear explanation and the final result will be withheld until you understand why. For example, when analyzing this Chinese sales data, using the traditional pd.read_excel() method leads to a total that is 161% higher than the actual value without any warning or error. But with huashu-excel, all the three issues (subtotal row mistakenly included as one store, double counting of regional subtotal, duplicate entry) are identified in the diagnosis phase. This skill requires only openpyxl as external package.

North Mini Code 1.0 is the first developer-focused model from CohereLabs (Cohere), built specifically for agentic software engineering. Architecture: 30B parameters with 3B active per step (MoE), 256K-token context, up to 64K output tokens. The model supports tool use, terminal commands, multi-agent coordination, and interleaved thinking — reasoning out loud before committing to an answer. On SWE-Bench Verified it scores 67.6, which is competitive for a model that runs on a single machine. Throughput is 2.8x higher than similarly sized coding models. Run locally with: ollama run north-mini-code-1.0.

NVIDIA Nemotron 3.5 Lightning is an open 30B mixture-of-experts model with only 3B parameters active at each inference step. It was built for always-on agents: personal assistants managing calendars and email, automation workflows, and similar long-running tasks. The context window is 1 million tokens. According to NVIDIA, the model delivers 4x higher throughput and 30% lower task completion time compared to similarly sized models. An MLX variant (23 GB) is available for Apple Silicon. Run locally with: ollama run nemotron-3.5-lightning.

Meta Muse Glimmer-30B is an open multimodal model distilled from Meta's internal flagship Muse Spark, tuned for local deployment on consumer hardware with 24–32 GB VRAM. It was built for always-on autonomous agents: tool use, code writing, failure recovery when tool calls go wrong, and multi-step reasoning across text and images. According to Meta's official benchmarks, it scores 94.7% on AIME 2026 (math olympiad), 76.0% on SWE-Bench Verified (real-world coding), and 75.5% on MCP Atlas (agent evaluation). The context window is 128,000 tokens. Licensed under Apache 2.0. Run locally via: ollama run muse-glimmer.

Wan 2.2 is an open-sourced video generation model developed by the Wan-AI team. Given an input image and a textual description of the target motion, Wan 2.2 can generate 720P videos. In our r3gm Hugging Face space, we run Wan 2.2-14B under an optimized inference pipeline (fp8 quantization and AOT compilation) on free ZeroGPU resources with zero sign-ups required. Compared to Wan 2.1, Wan 2.2 is a huge improvement because it was trained on 83% more video data and has a new mixture-of-experts architecture that decomposes the generation task into different expert blocks to achieve smooth and coherent motion.

Ornith-1.5-9B is a compact language model from ornith-ai built for agentic coding and step-by-step reasoning. The Ornith series is designed around end-to-end self-improvement — the models learn from their own outputs during training. This GGUF release packages the 9B variant as a single file readable by llama.cpp, Ollama, and compatible tools. The original version was released in a server-only format; the GGUF build brings the same model to consumer hardware without any environment setup.

wide-trace/open-higgsfield

Free and open-source alternative to the Higgsfield AI web app for generating images and videos with a unified interface. Gathers 40 AI models (12 for images + 28 for videos including Flux, Kling, Veo) in one place instead of using multiple services. All models are accessible with the same prompt bar, each model has its own parameters. Up to 4 generations are allowed per prompt. The last 60 prompts are kept in the browser storage. Gallery divided into 4 parts: images, videos, assets, favorites. Available online at openhiggsfield.ai. No installation required. Needs generation service API key.

cloudflare/computer

Cloudflare Computer is an SDK provided by Cloudflare that allows AI agents to have their own virtual computer. This virtual computer has a durable filesystem (SQLite Durable Object), as well as a code execution environment. There are three different ways to run this code (full linux container with FUSE, just-bash in a Dynamic Worker, or a JS module), each of which can access the same durable filesystem via one API call. In addition, the MCP example in this repo provides a way for agents to interact with the workspace directly (examples/mcp). This means that you can use your Claude Code or other agent directly within the workspace without writing any code. This is currently labeled as preview and not intended for production, but there are working examples and demos for all the environments in the examples directory, including a tutorial for how to write a markdown document and convert it to a PDF using pandoc in a container. Cloudflare Computer has received 8700 stars in less than 3 months.

bam-bam-2/solo-skills

Solo Skills is a collection of 26 AI Agents skills developed by a Korean solopreneur who automated 49 things in his one man company. These skills are written in SKILL.md which can be used by Claude Code, Codex, OpenCode. They are executed as procedure scripts for your AI agent. These skills include making demo video without screen recorders (iframe capture + ffmpeg), making pdf from html (Paged.js), posting threads(Threads API), setting up discord alerts(launchd), and working on other machines(remote). There is a special skill called claude-codex-fallback that lets you use Claude as default but fallback to Codex when it hits rate limit. Some skills are using Korean services (Naver Mail, Threads API) but most of them are universal. You just need 2 commands to use these skills (clone this repo && copy skills folder).

proliferate-ai/proliferate

Proliferate is an open-source AI IDE (AGPL-3.0) for macOS that runs multiple coding agents — Claude Code, Codex, OpenCode, Cursor, and Grok — in parallel, each in an isolated git worktree with its own branch, terminal, and conversation history. Tasks don't interfere with each other's files. MCP servers, skills, Computer Use, and Browser Use are configured once and shared across all agents automatically. Proliferate supports subagents (agents that delegate scoped work to child agents and pick the results back up) and scheduled workflows — recurring or event-driven agent runs like nightly code review passes or dependency bumps. The full control plane is self-hostable via Docker Compose, AWS CloudFormation, GCP, Azure, or Kubernetes, documented in the deployment guides in the repo.

FLUX.2 Klein Multi-LoRA is a web image generation space using Black Forest Labs’s FLUX.2 Klein, a small version of their second-generation image model. FLUX.2 Klein has 4–9 billion parameters, while FLUX.2-dev has 32 billion parameters, therefore, FLUX.2 Klein is smaller and faster than FLUX.2-dev. A unique thing about this space is that it supports multi-LoRA style fusion. You can mix and match multiple styles using LoRAs simultaneously, and adjust the weight of each LoRA freely. This makes it easy to use style fusion for content creation if you don’t have many tools. This space is powered by Zero GPU from Hugging Face, you don’t need to install anything locally. Also tagged with mcp-server, which means you can interact with this space through agents!

Free AI Detector is an online tool developed by Lynote which analyses your text and gives an idea about its probability of getting detected as AI generated content. Enter any text, and get the score based on how “human” or “machine like” the content is. No signup, no download, just a simple web app hosted on Hugging Face Spaces, no need to install anything locally. Launched on 5th Aug, it has got 330 likes and 8 discussions within 3 weeks. Usually, these AI text detection tools come with subscription plans, but here’s a tool without any subscription.

rome-os/rome

Rome is an open-source agentic OS (MIT) where agents are not only informational but also create persistent functionalities. Simply describe a task in plain language to create a Rome App — a tailored interface with agents, persistent data, schedules, and hooks. Everything from email triage to price tracking to code reviews to morning briefings becomes an autonomous app that can run while you’re offline. Browse and download other apps from the Community App Store or share your own. Use the self-evolutionary loop — the more tasks you describe, the more functionality you’ll get out of your Rome instance. Install Rome as a Docker container with one line or check out Rome Cloud (preview) — no installation required.

localai-org/kimodo.cpp

The LocalAI community has ported the NVIDIA Kimodo text-to-motion model to C++ / GGML so you can use text to animate 3d models on your computer! The Kimodo model uses a text prompt (or LLM2Vec embedding) to generate 3d human motion data (in SMPL-X format: root translations + joint rotations). Under the hood, it’s powered by a Llama 3 text encoder and the Kimodo-SMPLX-RP-v1 diffusion model, both available as out-of-the-box GGUF weights from the LocalAI-io organization on HuggingFace (download with a single line of code). Works with CPU or Vulkan compatible GPU. Compile with CMake. Run a browser demo server locally at localhost:8094 and test out prompts. Licensed under Apache-2.0. Model weights for non-commercial research use only.

kunchenguid/backpass

backpass is a command line interface (CLI) that helps you iterate on your AGENTS.md file (or CLAUDE.md) based on your agent usage. It parses local agent transcript files from seven different agent harnesses (Claude, Codex, Pi, OpenCode, Grok, Cursor CLI, Hermes) without any API keys or uploading. Each suggested change includes direct quotes from actual sessions to support it, and each new rule must be supported by at least two unique sessions (no single session can support a change). Each execution will suggest up to five changes (called a “gradient step”), which are small, noisy, and generally safe. All secrets are redacted before being passed into any model calls. Changes are only written via the backpass apply command, which displays all proposed changes along with their supporting evidence and allows you to manually approve or deny each change. backpass requires Node.js version 22.5 or higher and the acpx utility on your PATH.

ApodexAI/FrontierAgent

Apodex AI’s FrontierAgent is an open-source multi-agent runtime and terminal TUI designed for long-horizon tasks such as research, file analysis, and document writing. In addition to the one-shot ReAct workflow mode (where a single stateful agent performs the entire task), there’s also the Agent Team mode, which lets you break down the task into parallel subtasks and delegate them to sub-agents, then aggregate their outputs and compose the final result. You can see real-time progress of individual tasks, currently running tools, and past actions. Mutative actions will prompt for diff approval and allow you to save checkpoints along the way, making your session resumable. All file I/O is sandboxed so that input files are immutable, your workspace is at /workspace, and all persistent outputs are placed at /outputs. FrontierAgent supports connecting to any OpenAI compatible API server (e.g., Apodex, OpenAI, or your own local model server). Open sourced under Apache 2.0.

MiniMax Music3 — 2.4B parameter model capable of generating complete songs (up to 5 minutes long) including lyrics, instrumentation, and structure (verses, choruses, bridges, outros). Input is a set of verses/choruses and a description of the desired song (genre, BPM, instruments, mood), the model will create a story around them. This model uses a combination of global (8B) LLM for high-level planning and local (0.6B) LLM for low level details, alongside Flow Matching to produce 32 kHz stereo WAV. However, this model doesn’t seem to have an explicit license, which means you probably shouldn’t use it commercially.

Deploy MiniMax H3 — a 33B transformer by MiniMaxAI which converts text/images to videos with matching stereo audio upto 2K resolution. This model has been divided into 3 parts: context pre-processing, 768p base generation & 2K upscaling(H3-Regenerate-2K is currently closed-sourced). Used this model on image+scene description tasks, got ~4–15 sec of videos @24fps. This is the first open-weighted model that I’ve come across which supports stereo audio natively(no need of any audio processing pipeline)

Run DeepSeek-V4-Flash-0731 using vLLM on a server environment. A 304B MoE model released under the MIT license by DeepSeek. ‘Flash’ refers to DSpark speculative decoding to achieve fast inferencing while only activating a fraction of the total parameters. On Terminal Bench 2.1, it scored 82.7 compared to DeepSeek-V4-Pro Preview’s 72.1. On DeepSWE, it increased from 12.8 to 54.4 and on AutomationBench from 12.8 to 25.1. Low/high/max reasoning_effort settings allow you to control how deep/fast you want the model to reason about your task.

We’ve made Qwen3.8-27B (Alibaba’s 27B multimodal model) available as Dynamic 3.0 GGUF quantizations for local use with Ollama and llama.cpp. This model has image understanding capabilities, natively supports a 262K context length (can be scaled up to 1M tokens with RoPE), and is tuned for agency and tool invocation. With Dynamic 3.0 you can select from 12 different quantization levels (from IQ1, ~6GB to BF16, ~55GB) to find the right tradeoff between performance and hardware requirements. By keeping key weights less compressed, we believe we’ve found a way to surpass other quantizations at similar sizes in terms of accuracy.

Omni Image Editor is a browser-based AI image editing application published by selfit-camera on Hugging Face Spaces. It runs entirely in the browser without installation and supports four core operations: generating images from a text prompt, editing existing photos (removing unwanted elements, changing details), upscaling resolution, and removing watermarks. The space has gathered over 2,400 likes on Hugging Face, reflecting strong community adoption. No account or GPU is required on the user's side.

Ornith-1.5-9B is a 9.7 billion parameter dense language model developed by ornith-ai on top of Qwen3.5 and Gemma4, further pretrained and trained with RL. What sets this model apart is its self-improvement training paradigm where the model was able to generate tasks for itself, solve them, and improve without needing static datasets from humans. This model can take both text and image inputs, call tools, and operate within agentic workflows like those powered by MCP. When used in an agentic workflow using OpenHands harness, it achieves a score of 70.6 on SWE-bench Verified which is quite good for a 9B model. You can find a quantized GGUF version of this model on Hugging Face, which runs with ~6GB of memory at 4-bit.

TRELLIS.2 is the next iteration in our series of 3D generation models. It’s a Hugging Face Space that runs without installation. Given an image, it generates a 3D model via Trellis2ImageTo3DPipeline. There are 6 different rendering options available to visualize the generated output (Normal, Clay render, Base color, Forest HDRI, Sunset HDRI, Courtyard HDRI). Our first TRELLIS model was one of the most forked 3D generation models on Hugging Face and there are now more than 100 community spaces based on it. Now, TRELLIS.2 brings even better quality 3D generation!

Anime illustration text to image model fine-tuned on circlestone-labs/Anima on NVIDIA’s Cosmos-Predict2-2B, has been trained on 1.7 million anime images with a July 2026 cutoff. It has 2.9 billion parameters and was created by increasing the 28 layers of the original architecture to 40 layers through interleaved insertion and setting the output projections to zero. It was trained on a mix of captions from both Gemini and Claude (including both tag and natural language descriptions but no score-based tags), and can be loaded into ComfyUI as a single file diffusion checkpoint (.safetensors). License: Non-commercial.

Comfy-Org packaged MiniMax's H3 Turbo video generation model specifically for ComfyUI users, with everything pre-configured. The package includes four quantized model variants (bf16, int8_convrot, fp8_scaled, nvfp4_awq) and pruned options to accommodate different GPU budgets. Three turbo LoRAs enable fast 4-step and 8-step generation instead of the usual 50. Ten specialized embeddings — including fire_breath, dark_magic, bullet_time, and others — can be mixed in to push specific visual styles. Pre-built workflow files cover all three generation modes: text-to-video, image-to-video, and reference-to-video. The int8_convrot variant is recommended with PyTorch cu130.

zhu1090093659/dsh-web

DSH Web is a plugin bundle for DeepSeek Harness Web that extends the base interface through the official DSH profile mechanism, without modifying upstream code. It ships a complete set of modular plugins: a multi-column task board with real cron execution (tasks run on schedule even with the browser closed, with optional sleep-prevention on Windows, macOS, and Linux), an SSH panel with terminal, file transfer, tunnels, and cluster management, mobile remote control via QR code with SSE live sync, image understanding via a describe_image tool, a right-side panel with file explorer, editor, Git history graph, and browser, plus 19 skin themes with a live-preview editor. All plugins install independently; the full bundle installs with one npm command. A community marketplace at dsh-market.com distributes additional skins, pets, and plugins.

s1dashu/ip-as-logo-skill

I added this skill to my agent setup on a whim — it's literally one SKILL.md file, and yet it reliably generates company-ready IP mascots: bold rounded silhouettes, solid background color, strict lower-corner composition. I pointed it at a few project names I had lying around and was genuinely surprised by how consistent the output was. What impressed me most is the explicit complexity limits baked into the prompt — no fussy gradients, no noise, just lovable and clean.

MengTo/threeui

I've followed MengTo's Design+Code since the early days, so when he open-sourced ThreeUI Community I dug in immediately. It's the same application shell, navigation, and live WebGL renderer as the paid product — just without the Pro and Beta components. You get a full browser-searchable catalog of Three.js UI components with shader previews and a variant picker, all running in the browser without login. I spent an embarrassing amount of time just clicking through the shader demos.

duty1g/x64dbg-mcp-server

I'm not a full-time reverse engineer, but this one caught me instantly — it's a native x64dbg plugin, built in Zig with zero dependencies, that exposes 71 MCP tools over HTTP. You connect Claude or any MCP-compatible assistant and literally ask it to set a breakpoint, dump registers, or read memory. It supports both 32-bit and 64-bit, has 22 event callbacks for debugger state, and the single-binary output means there's nothing to install beyond dropping the DLL. This is what AI-assisted binary analysis should look like.

DenisSergeevitch/desktop-fly

I installed this mostly as a joke and ended up reading about the FlyWire connectome project for an hour. It's a macOS app with a 3D fruit fly that wanders your desktop using a live spiking neural network simulation — ~668 neurons and ~19,000 real synaptic connections from the actual FlyWire dataset. The fly walks, grooms itself, sleeps, and escapes your cursor, all driven by real neuroscience. You can open a brain visualization panel showing all 23,210 neuron positions and stimulate individual regions to watch behavior change.

iAmCorey/Wake

Once you start running Claude Code, Cursor, and Copilot CLI in parallel across a few projects, finding old sessions becomes a chore. Wake indexes all of them locally — reads Claude Code, Cursor, and other agent transcripts — and lets you full-text search with SQLite FTS5 trigram indexing, which means it handles CJK and 30+ other languages without configuration. Built in Rust with GPUI, it's read-only and makes zero network requests. I pointed it at my project folders and had two months of sessions indexed in under a minute.

simonlin1212/a-stock-data

A-Stock Data is a standalone Skill file(structured Markdown + Python code) for Claude Code and Codex, which integrates Chinese A-share market information from 19 data sources into 54 callable endpoints without authentication. It includes 11 layers, including real-time stock quotes and order list information provided by mootdx and Tencent, research report information and PDF download provided by Eastmoney, smart money flow and industry hotspots provided by Tonghuashun, limit up pool, board of directors, ETF option Greek number and implied volatility information, interactive questions and answers between investors provided by cninfo, and macro economic index data such as PMI provided by the People’s Bank of China and the National Bureau of Statistics. There are also corresponding backup mechanisms for primary APIs that are subject to request quotas. This project was started for three and a half months and got more than 9 000 stars. The main readme is written in Chinese. You can also view the English version (README_en.md).

open-gsd/gsd-core

GSD Core is a context engineering and spec-based development framework for AI code agents like Claude Code, Codex, OpenCode, Cursor, Windsurf, Copilot, Antigravity CLI and more. It prevents context rot, the silent degradation in quality from agents as they fill their context windows. GSD Core is a formalized 5 phase process of Discuss, Plan, Execute, Verify and Ship where all heavy lifting and research occurs in isolated subagents with a clean 200k context window. Files like STATE.md and CONTEXT.md provide structured data between sessions to prevent losing context. GSD Core has been out for 3 months and has over 8 600 stars. It can be installed using npm and has /gsd-new-project and /gsd-onboard commands. You can have several plans open at the same time, executing in waves. This project also has a GEMINI.md context file for Antigravity CLI, the official replacement of Google’s Gemini CLI (Google sunset Gemini CLI in June 2026).

kgoedecke/doop

An open source alternative to Paper.design, the multiplayer design canvas. Work together with other humans and AI agents in real time. Connect your AI agents to the built-in MCP server and stream the design from HTML to a frame at a time — watching the design come to life in your browser. Invite Claude Code with a single command: `claude mcp add --transport http doop http://localhost:4300/mcp`. Every frame of the canvas is a sandboxed iframe containing real HTML code, which the AI agents progressively fill out, check a screenshot of, and adjust if necessary. Collaborate on a project with many humans and AI agents at the same time — each with their own live cursors and indicators for editing in each frame, and a shared activity feed. Use the design memory feature to pin select frames as “exemplar” frames which will then be extracted to style rules that all the AI agents must follow. Set up self-hosted, no configuration required (using docker compose up or npm run dev with a built in Postgres database, no external services).

furkankly/zoetrope

Zoetrope parses Claude Code’s JSONL transcripts (~/.claude/projects/) and presents the flow graph of a Claude Code session as an animated flow graph. Main agent on top, with agents & work groups below, tools are presented as live chips. Tool calls are shown along with the name of the tool, the number of times it was called, and the total time spent in the tool. You can visualize Claude Code sessions live or rewind & fast forward sessions. The minimap shows up when the graph exceeds the bounds of the screen. Same functionality is made available via WebAssembly at zoetrope.furkankly.dev. Simply drag & drop a Claude Code transcript to get the same flow graph in your browser. All processing happens locally.

anywhere-labs/dsh-desktop

DSH Desktop is a open-source cross-platform native desktop GUI frontend for DeepSeek Harness without any dependencies such as Node.js or command line interface, it allows you to easily start and use DeepSeek Harness like a normal software, supports both Windows x64 and macOS Universal versions, downloads and includes the local Web UI, Host service and plugin system at a specific version, also provides features such as system tray, automatic startup with system, automatic update, etc. Install it by downloading the installer from GitHub Releases or dshdesktop.cn, then double click the installer and you can find DeepSeek Harness running as a tray icon with all the DSH plugins. As the community open source project(not related to DeepSeek), we have developed some extra plugins based on the official way of DSH plugin development, which makes sure our extra plugins will still work even after the upstream DeepSeek Harness was updated. In just 11 days, the project has already achieved nearly 20 000 stars.

Qwen3.8-2.4T-A95B is the most powerful model in the Qwen family, and the first time Qwen has released open weights at the quality level of their flagship paid model, Qwen-Max. It's a MoE (mixture-of-experts: only part of the model runs per request) architecture — 2.4 trillion parameters total, 95 billion active per token, with 92 layers and 512 experts (10 routed + 1 shared). It combines Gated DeltaNet (a linear-attention mechanism) with classic Gated Attention to handle long inputs efficiently. Native context is 262,144 tokens, extendable to 1,010,000. It's built for agent tasks, code, and research chains: SWE-bench Pro 67.7, GPQA Diamond 92.6, Terminal Bench 2.1 86.6. Text only; 'thinking' mode (step-by-step reasoning before the final answer) is always on and can't be turned off. To try it, ask your agent: 'Find me an API provider with access to the model Qwen/Qwen3.8-2.4T-A95B: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B — show me how to make a test request to this model through Together AI, Fireworks, or another provider that supports it.'

Ornith-1.5-35B-A3B is a Mixture of Experts model trained by Deep Reinforce (Ornith Team) for coding and agency. It has 35 billion parameters, but only uses approximately 3 billion at each token, so it runs faster than an equivalently sized dense model. It achieves 79% on SWE-bench Verified, compared to 73.4% for Qwen3.6-35B and 52% for Gemma 4-31B, and 70.2% on the MCP-Atlas agentic benchmark, compared to 62.8% for Qwen3.6-35B. It has a context length of 262,144 tokens, which can be increased to ~1M using RoPE scaling. It can make function calls through XML compatible with the OpenAI API, and use <think> blocks for chain of thought reasoning. Licensed under the MIT License. There is a prebuilt GGUF version here: ornith-ai/Ornith-1.5-35B-A3B-GGUF

kulkas2pintu/wan555 is a Hugging Face Space running Wan 2.2 14B — a video generation model from Alibaba. Forked from r3gm/wan2-2-fp8da-aoti-preview, it runs on Hugging Face's free Zero tier. Supports text-to-video (T2V) and image-to-video (I2V) generation. No account, installation, or payment is required — just open the link in a browser. Queue times may occur due to shared free hardware, but the tool is entirely free to use.

Ling-3.0-tiny is a hybrid MoE reasoning model developed by inclusionAI, which consists of 7.9B total parameters, but only activates 1.3B parameters per token. With its KDA-MLA architecture and 128-expert sparse MoE (8 routed + 1 shared expert active per token), it supports fast inference. It can run at 86–90 tokens per second on MacBook M4 Pro in FP8 mode, using ~8.34 GB RAM under an 8K context window. With YaRN scaling, the context window size can be scaled up to 256K tokens. When enable_thinking=true, temperature=1.0, top_p=0.95 is set, it will perform step-by-step reasoning first and then output the final answer. You can access this model through SGLang(Pre-built docker image), a special vLLM branch or an open pull request of Ollama for apple silicon.

The Agent Memory Leaderboard from agentmemories.ai is the first public leaderboard for long term memory in AI agents. We benchmark on 4 key areas of long term memory: retrieval, multi-session dialogue, personalization, and temporal reasoning (i.e. remembering the correct order of events). We have two categories: open source academic models, and production APIs. Our initial launch was in August 2026, but we already have entries from academic researchers and commercial providers. Runs in a Hugging Face Space, no install required!

unicity-aos/aos-ce

AOS Community Edition is an open-source agent operating system from Unicity AOS. It gives AI agents — Codex, Claude, Grok — a managed, inspectable runtime: a CLI (`aos`), HTTP API, and a set of 21 first-party capsules (modular user-space building blocks). Agents connect through a built-in MCP server (`aos mcp serve`); on macOS and Windows AOS shows its own approval dialog for sensitive operations, on Linux it uses Pinentry. Every release ships checksums, Sigstore bundles, and GitHub build-provenance attestations. The Community Edition also includes Forge, construction tooling that lets an agent inspect the running system, find capability gaps, and build new least-privilege capsules. AOS launched in mid-July 2026 and reached 8,500+ stars in six weeks. Licensed MIT/Apache 2.0.

t8y2/dbx

DBX is a cross platform database client implemented in Rust, only ~20 MB in size (no Java, no Python, no chromium). Supports 90+ databases like MySQL, Postgres, SQLite, Redis, MongoDB, DuckDB, ClickHouse, Elasticsearch, MS SQL Server, Oracle, etc. Built-in AI SQL helper understands natural language requests and generates SQL, explain queries, optimize them, or verify generated queries with built-in security checks before running. Supports Claude, OpenAI, local AI models through Ollama. Independent MCP server (npx @dbx-app/mcp-server) allows other MCP clients like Claude Code + Cursor + Windsurf + … to connect to your DBX connections with configurable access levels (read-only, safe write, or full access). Desktop application for macOS, Windows, Linux; Docker image for shared use; Web version. Released late April 2026, 16,000+ stars in 4 months.

nateherkai/scroll-craft

Scroll-Craft is a Claude Code plugin that builds scroll-driven websites to a professional design standard. It picks from 8 mutually exclusive page formats — filmic one-shot, continuous world, chaptered editorial, typographic poster, split stage, and others — where each format forbids the moves the others use, so two builds can't converge accidentally. Every build must include a bespoke 'signature move': a unique interaction that exists on that site alone; a fingerprint gate compares each new build against your history on 6 dimensions and the agent must change the plan (not the record) if it fails. After building, a headless browser walks the entire page at every scroll position: it measures real contrast on composited frames (per text line, at the brightest video frame that ever passes under it), detects dead scroll zones where nothing changes on screen, and flags content that never reaches full opacity. The skill enforces strict typography rules (two font families max, 45–75ch measure), colour roles (six roles, no pure black), and depth tools (five techniques, not just drop shadows). Building from your own photos and video costs nothing; AI-generated assets require an optional KIE_AI_API_KEY.

jaredrhod/fullstack-agent

fullstack-agent is a setup wizard to install Claude Code’s personal AI assistant from 4 different open source projects: ai-memory-vault (persistent memory, stores all of Claude’s knowledge about you and your work in unlimited size text files), backtalk (voice response, press button, say something, hear response within ~1 second), ai-visualizer (full screen animations that change based on listen, think, speak), barehands (hand gesture control, works with a webcam without a headset). Everything is optional and can be installed individually. The setup wizard will ask you what you need. If something goes wrong, just tell Claude what happened and he will fix it for you with the included troubleshooting manuals. Updates are done with git pull and do not affect your personal information or agent identity. Available for macOS (Apple), Linux and Windows. On macOS, Apple Command Line Tools are included, macOS will prompt you to install them if you run it for the first time.

ifixai-ai/iFixAi

iFixAi is a standalone AI agent evaluation system that focuses on non-technical aspects such as deception, fabrication, manipulation, unpredictability, opacity etc. It can be used by any AI service providers including OpenAI, Anthropic, Gemini, Azure, Bedrock, OpenRouter, or custom AI agents via HTTP interface. It provides a grade from A to F along with detailed breakdown on each pillar. You can evaluate any AI agent whether it’s from OpenAI, Anthropic, Gemini, Azure, Bedrock, OpenRouter, or even your own AI agent. It can also be used as a Claude Code plugin/Codex plugin/CLI wizard/One Click Skill for Cursor/VS Code/Windsurf/Cline/… When it comes to citable grade, we need a second opinion from another independent AI service provider (not the one we are evaluating). It takes less than 2 minutes to run all the tests. Approximately $12–18 if we use claude-sonnet-4.6 as the judge (OpenRouter price in mid-2026). Licensed under Apache 2.0.

freestylefly/awesome-gpt-image-2

Awesome GPT-Image-2 is a structured prompt library for GPT Image 2. Created by independent Chinese developer 苍何 (freestylefly). Awesome GPT-Image-2 has 532 reversed cases and 20+ high-quality templates in 11 categories: UI & Interfaces, Charts & Infographics, Posters & Typography, Products & E-commerce, Brand & Logos, Architecture, Photography & Realism, Illustration & Art, Characters & People, Scenes & Storytelling, and Documents & Publishing. Inspired by “Prompt as Code”, it decompose prompts into atomic pieces like subject, lighting, materials, layout, and visual details for agents, scripts, and automation use cases. Most of the template documents are in Chinese, but a preview website (gpt-image2.canghe.ai) for browsing, copying, and filtering prompts are in English. A skill for Claude Code/Codex is also available on npm for selecting styles, templates, and scene tags based on the same data set. Awesome GPT-Image-2 was released in late April 2026. It got 12500 stars in 4 months.

FlashML-org/FreeToken

FreeToken is a device-native inference server that makes it possible to run large open-weight MoE (Mixture of Experts) models on edge devices. It was just released a month ago and has already got 3300 stars! Developed by Berkeley/MIT FlashML team (Matei Zaharia, co-author of Apache Spark, Ray, Song Han, Ion Stoica) and published on arXiv (2608.16157), it leverages bandwidth adaptive CPU–GPU co-execution, global LRU expert cache and semantic KV cache checkpoints to load models layer by layer to GPU and host memory, caching hot experts. With this approach, one can deploy 290B+ models such as DeepSeek-V4-Flash, Qwen3.6-35B-A3B or GLM-5.2 to NVIDIA RTX 30/40/50 gaming GPUs locally and avoid paying per token and uploading your data to the cloud. Supports Anthropic API out-of-the-box, so one can connect it to Claude Code and other tools like Codex and OpenCode. Desktop apps for Windows and Linux with UI are available, CLI (pip).

cclank/lanshu-create-ai-presenter-video

Lanshu Create AI Presenter Video is a Codex skill for producing presenter (digital human) videos. You provide a topic or finished script and an authorized portrait photo — the skill writes or refines the copy, synthesizes or clones a voice, generates a digital human video, syncs lip movement to the audio track, adds captions and keyword animations, edits the timeline, and runs a quality check. The workflow fixes the full voiceover first, then uses it as the timing backbone for everything else — reducing lip drift and cut mismatches. Default output is a 9:16 1080×1920 30 fps video, 45–75 seconds. Before any paid step the agent discloses cost and waits for approval. The project is provider-neutral: it selects whichever video, TTS, and lip-sync services are available in the current Codex environment.

amagine-ai/Amagine3D

Amagine3D is an open source AI platform to generate editable parametric CAD 3D models. Describe your requirements in natural language, provide references and key dimensions and let the AI agent create geometry in a self-iterative fashion by creating design candidates, checking them in a physical environment and saving validated geometry as new baseline. The outcome is not just an image but a fully editable parametric CAD model using build123d library and Open CASCADE Technology kernel (used in FreeCAD, CATIA etc.). You can continue editing parameters in workbench and geometry will update itself. Export in STEP format for compatibility with other CAD platforms, STL for 3D printing and 3MF (color aware) for multi-material 3D Printing. Node.js >=20.19 and Python >=3.10 to <=3.13 are required.

Qwen Image Edit 2511 LoRAs Fast is a Hugging Face Space by prithivMLmods that runs the Qwen image editing model directly in a browser interface. You upload a photo and type an instruction describing what to change — remove background, switch to night lighting, add objects, change style — and the model generates the edited version in seconds. Optimized with LoRA adapters for faster inference without quality loss. No account or installation needed; runs on Hugging Face's free Zero GPU infrastructure. The space has accumulated over 2,600 likes, making it one of the most popular image editing tools on the platform.

IndexTTS-2.5 is an open source TTS model developed by the Index Team of Bilibili that can synthesize a voice with only one reference audio sample to capture the tone and color of the voice. It provides 8 emotion control modes (such as happiness, sadness, anger, etc.) controlled by prompts, speech rate control and phonetic accent control, and is applicable to Chinese, English, Japanese, Spanish and Arabic, and does not support Russian. It needs to use Python 3.10–3.11 and a GPU card with about 6GB of memory, Clone the GitHub code and execute uv sync to install.

DeepSeek-V4-Pro-0813 is the official release of DeepSeek V4 Pro from DeepSeek AI, superseding the earlier preview version. The model has 1.7 trillion parameters, MIT license, and is specifically engineered for agentic use cases: long-horizon planning, terminal operations, and multi-step automated pipelines. On Terminal Bench 2.1 it scores 87.9 versus 82.7 for V4-Flash — a consistent improvement across benchmarks. The model includes a built-in speculative decoding module called DSpark for faster inference. Weights are open on Hugging Face; running it at home is not feasible — it requires data center GPU infrastructure.

Zeejay0/gathered-scenes-zine-skill

Gathered Scenes Zine is a collection of two Codex skills for AI-powered photo-to-art generation using GPT Image 2. The first skill (Gathered Scenes, `$scenes-gathered-zine-v1-3`) keeps the original photo as a visual anchor in a collage that adds abstract shapes derived from the photo, a single high-saturation colour, and visible torn-paper fibre edges. The second skill (Scene Distillation, `$scene-distillation-zine-v1-3`) does not include the photo in the result — it extracts the semantic core, emotional tension, and visual metaphor, then creates an entirely new illustration in an editorial minimal style. Skills are installed to `~/.codex/skills/` and invoked inline: `$scenes-gathered-zine-v1-3 [description]`. Repository includes a before/after case archive showing exactly what was preserved from each original scene. Released roughly six weeks ago and reached 4,000+ stars. Requires OpenAI Codex with GPT Image 2 access.

truefoundry/trueforge

TrueForge is an open source agent wrapper developed by TrueFoundry that wraps around any LLM to make it into an agent. It was released about a month ago and now has 3,400 stars. It provides model calling functionality, tool invocation, isolated code execution (using Daytona), session management, and human approval (where the agent stops and asks for permission before performing important actions while keeping track of all its actions in an audit trail). It works with OpenAI, Anthropic, Google Gemini, and other OpenAI compatible endpoints. It supports skills (instruction sets stored in Git) which can be added to the agent without modifying the underlying runtime and can be updated dynamically. There are three ways to interact with it: through a chat UI, via an HTTP API for integrating it into your applications, or using embeddable UI elements to build your own UIs.

PrimeIntellect-ai/prime-agent

Prime Agent is a coding and research agent developed by Prime Intellect, open sourced just 4 months ago and already with almost 18,000 stars. It’s an IPython persistent environment which can run code, manipulate files, create and interact with parallel subagents, all without going through you. The main feature of the agent is self improvement — every time you use the /refine command, it analyzes its previous performance and improves its own additional prompts and skills for the next time around. You can even save commonly used flows as reusable skills. Finally, the long running auto mode with periodic heartbeats allows the agent to keep working for hours while you sleep, on complicated research tasks. Available for Mac OS X and Linux. At first launch, you can either select a subscription plan or use your own OpenAI/Anthropic/other API key.

openai/codex-security

Codex Security is the official OpenAI CLI and TypeScript SDK for AI-driven code security. Scan your code with Codex to detect security issues, verify them, and generate patch proposals. Enable automatic patching of “high” and “critical” issues with --patch --patch-severity high, commit modified files and create a draft Pull Request on GitHub with --create-pr. Scan all components in your monorepo independently and aggregate issues by root cause with scan-components. Use multiple subagents for deep scan mode (up to 96 hours), or run on OpenAI, Amazon Bedrock, OpenRouter, or any other OpenAI-compatible platform (e.g. Anthropic via OpenRouter). Launched two months ago and already over 10,000 stars! Works with Node.js >= 22 & Python >= 3.10, requires a ChatGPT account with access to Codex Security or an OPENAI_API_KEY.

microsoft/SkillOpt

SkillOpt from Microsoft Research (released July 2026, already 16,000 stars) treats natural-language skill documents (SKILL.md files) as trainable parameters for LLM agents. Without modifying model weights, it runs an iterative refinement loop: the optimizer scores agent rollouts against benchmarks and makes bounded edits — add, delete, or replace fragments — accepting only changes that pass a held-out validation test. The result is a deployable best_skill.md artifact of 300–2,000 tokens that plugs directly into Claude Code, Codex, or any harness that reads skill files. Microsoft's own benchmarks show +19.1-point accuracy lift inside Claude Code, +24.8 inside Codex, and +23.5 in direct GPT-5.5 chat, across six benchmarks and seven models. Supports OpenAI, Azure, Claude, Qwen, and MiniMax backends; includes SkillOpt-Sleep mode for nightly offline self-evolution and a WebUI dashboard for monitoring.

lidge-jun/opencodex

OpenCodex is a local proxy that routes Claude Code, Codex CLI, Claude Desktop, Grok Build requests to any LLM backend providers(40+ supported by default, Any OpenAI compatible endpoint). It can be configured via web UI(localhost:10100). Supported backends include Anthropic, Google Gemini, xAI, Kimi, Azure OpenAI, Ollama, DeepSeek, Tencent Cloud… The provider/model format is supported in the model picker. Other functions such as: using a pool of ChatGPT /Codex accounts with session affinity/load balancing(keep active connections sticking to the same account), combos(a magic virtual model id which can fallback or round robin between multiple backends). OAuth supported backends include xAI, Anthropic, Kimi. Fully revert your settings via ocx stop.

cobusgreyling/loop-engineering

Loop Engineering is an open source library of patterns, templates, and CLI tools for creating systems that orchestrate AI code agents like Claude Code, Codex, Cursor, etc. It’s based on the concept of “building a loop” instead of manually prompting the agent each time, your system periodically prompts the agent using persistent state, budgets, and skill files. Using the loop init CLI, you can set up a new loop within 5 minutes by generating STATE.md, LOOP.md files, budget and constraint files and get a Loop Ready Score. Use the loop doctor CLI to see the top 3 things you can do next. Other CLI tools include loop-audit (to get your Loop Ready Score), loop-cost (to estimate your token spend), loop-sync (to check for drift), loop-context (to manage memory), loop-worktree (to create isolated git worktrees), loop-gate (to enforce path denylists), and an MCP server (to lookup patterns). As Boris Cherny (Head of Claude Code at Anthropic) puts it: “I no longer prompt Claude. I have loops that prompt Claude and figure out what to do.” Introduced just 2 months ago, it has more than 10,500+ stars.

Ariescar/anyCreature

anyCreature is an open source pipeline for describing creatures in a single sentence, and getting back a game-ready GLB with rigging, skinning, vertex colours, and ambient occlusion baked, generated using AI. It’s a 5 stage (00_START to 04_SHIP) pipeline, described in cards, which are loaded by an executing AI session. There are no runtime dependencies, and can be run with just a JSON spec file as input, and a skinned GLB as output (`node engine/cli.js spec.json out.glb`). All output is kept strictly on track by independent ‘blind reader’ quality checks, where the AI session must produce a 24 pixel silhouette which is recognisable by a ‘blind reader’ AI which has no context on what should come next. There are also no self-checks, so if a silhouette doesn’t work for the context, it gets re-worked until it does. Node.js 18+, Python 3.9+.

antirez/ds4

DwarfStar is a native inference engine developed by Salvatore Sanfilippo (Redis creator) for DeepSeek V4 Flash / DeepSeek V4 PRO / GLM 5.2. It is NOT a generic GGUF loader. It supports Apple Metal (Mac, main target), NVIDIA CUDA (multi-gpu) and AMD ROCm. It uses 2-bit quantization for routed MoE experts only (no shared experts, projections, routing), preserving quality while keeping the model size small. Built-in HTTP server allows micro-batching for multiple users. You can connect two MacBook M5 Max or M3 Ultra laptops with RDMA to do tensor parallelism. SSD streaming is possible if you don’t have enough RAM. See the benchmarks in README. For example, on MacBook Pro M5 Max 128GB it achieves 39 t/s. Beta.

Yuan1z0825/nature-skills

Nature Skills is a collection of 19 reusable AI research skills for Claude Code, Codex, and other compatible agent frameworks (Node.js 18+). These skills include paper writing (Nature-like writing style, abstract/introduction writing, peer review response), paper reading (source-linked markdown reader with equation/figure support, paper to presentation in Chinese), and general research (literature search across multiple sources with cross checking with Nature/CNS journals, statistical analysis, data availability statement writing, patent filing based on papers, reference checking). The skills can be installed globally using npx, and will be available in all sessions without any setup. Each skill represents a reproducible process rather than a set of prompts.

virgiliojr94/book-to-skill

Book to Skill creates a Claude Code skill, GitHub Copilot CLI, Amp, or Codex skill out of any technical book, directory of documents, or collection of files. Rather than reading an entire book into context at once, it identifies relevant frameworks, heuristics, anti-patterns, and chapter files which are then read into context on demand (when asked about that specific subject). Authors: 24–51x fewer tokens required to achieve same results as blasting the entire book into context. Input formats: PDF, EPUB, DOCX, MD, HTML, RTF, MOBI. Contains: SKILL.md (~4,000 tokens) (an index), chapter files (~1,000 tokens/file), glossary, patterns reference, cheatsheet. Can be used for non-books too: architecture docs, style guides, research paper sets, RFCs, compliance standards, etc. Anything you can use as a reference that you’d like to memorize.

MiniMax-AI/MiniMax-H3

Official Repo of MiniMax. There are 9 skills for the H3 video model(Hailuo). The important one: h3-prompt-writing skill, is a simple markdown skill without calling any APIs, thus compatible with Claude Code, Claude Agent SDK, Cursor, Windsurf, Codex, LangChain, all the agents which can read SKILL.md. There are 2 first party prompt guidelines in the repo: base-en.txt for keyframe mode and text mode; ref-en.txt for Ref2VA mode. Other 8 skills are product ads, 3D animation, stop motion, brand promotion, subtitle videos, etc. These are some special video generation skills designed for MiniMax Hub platform. No portability. Vendor: MiniMax; Online API of this model.

lencx/Minke

Minke is a desktop app (macOS, Windows, Linux) for DeepSeek Harness developed by indie developer lencx. It unifies agent chat, syntax highlighted file diffing, file manager, browser and plugin manager into one window without requiring tab switching. Minke can connect to DeepSeek using its cloud API or to local models using LM Studio + Ollama (or any other OpenAI compatible API). Minke can also be accessed from mobile devices using Tailscale Serve. On iPhone or Android, Minke is opened as a PWA where you can carry on agent tasks and remote terminal usage. Data is stored locally. Minke supports installing plugins from its user interface which is connected to a built-in Github catalogue.

langchain-ai/openwiki

An agent powered by LangChain AI that builds and maintains wiki documentation for a codebase or knowledge base. The agent reads your source files and generates linked Markdown docs into an openwiki/ directory. Each claim will be grounded to certain lines of code, so your wiki is always up-to-date with changes in your code. It outputs docs in Google Open Knowledge Format (OKF v0.2) to work with other OKF-aware tools. OpenWiki works with 13 different model providers including OpenAI, Anthropic, Gemini (Google Vertex AI), Amazon Bedrock, and more. You can even integrate OpenWiki to Claude Code & Codex with just 1 line of code (openwiki integrations install claude). Your agent will take care of the documentation for you using the same models you already have. Also includes 9 built-in connectors such as Notion, Slack, Gmail, X, Hacker News, LangSmith, and more. Your wiki will auto update itself using Github Actions whenever you make changes to your code. Launched in June 2026. Already has 15k stars.

drumih/turbo-fieldfare

Optimized Swift + Metal inference runtime for Google’s open Gemma 4 26B-A4B model. It’s designed specifically for this model (not a generic engine like MLX or llama.cpp), and only caches the shared 1.35GB core and FP16 KV cache in memory while streaming in experts from SSD as needed. This enables running the full 14.3 GB model with just ~2 GB of memory, making it accessible even on low-end Apple Silicon Macs. A fully documented set of 103 experimental results are included for various kernels, caching, I/O, and decodes. Includes a Mac native app, CLI, model streamer, and a local OpenAI server. Requires macOS 26+.

BigPizzaV3/CodexPlusPlus

Codex++ is an external launcher and manager of OpenAI Codex desktop app, written in Rust 1.85+ and Tauri 2.x. It connects with the official OpenAI Codex desktop app by using Chromium DevTools Protocol, without modifying app.asar or installing patch files under installation folder. This program provides four types of providers (official account, official account & API, pure API and aggregated provider with fallback & load balance support to multi-provider) to manage auth info respectively. It also has a powerful session manager with multi-session scan, multi-session delete, multi-session markdown output, session tokens history. Also it includes some UI enhancer features like plugin center open, Chinese UI switcher, remember conversation width & scroll position, Stepwise suggest next step with custom model, Upstream worktree create, Zed Remote support, user script support. Installer is available for Windows x64 (signed), macOS Intel (x64, signed) and macOS Apple Silicon (arm64, signed). Most README is written in Chinese, but you can find the English version here.

ayghri/i-have-adhd

i-have-adhd is a skill for Claude Code and Codex that makes your agent’s responses more concise by only focusing on the steps to take first and then explaining later. The agent always responds in accordance with 10 rules: provide an explicit instruction or command to do next, enumerate steps if necessary, finish with a single action item, avoid distractions, repeat back context on every turn, never use preambles like “Great question!”, recaps, or closers like “Hope this helps!”, limit lists to 5 items, refer to time in minutes instead of “soon” or “briefly”, don’t apologize for mistakes, everything is configurable. Clone the repo, modify SKILL.md, install your fork. Published in May 2026, now with 23k stars — trust me, we need it.

alibaba/open-code-review

Open Code Review (OCR) is Alibaba Group’s internal AI-enabled code review CLI, open-sourced after 2 years of internal adoption by tens of thousands of developers. OCR takes a Git diff as input, uses an agent with tool usage capabilities to ship changed files to a configurable LLM, and generates structured code review feedback with line-level granularity. OCR is a hybrid system that combines the advantages of deterministic engineering (e.g. file selection, grouping, and rule matching) and LLM agents (e.g. dynamic context and comment generation). On 50 popular open-source repositories + 200 real pull requests + 10 programming languages, OCR uses ~1/9 tokens of general-purpose agents like Claude Code while achieving higher precision; recall is low to avoid introducing noise. OCR supports plugins for Claude Code, Codex, Cursor, CI/CD support for GitHub Actions and GitLab CI, compatibility with OpenAI, Anthropic, and arbitrary providers.

Spielewoy/autoprompt-skill

Autoprompt is a skill for coding agents that improves scores by +14.61 points (+45% fewer task failures) on the independent Terminal-Bench 2.1 benchmark. DeepSeek V4 Flash 0731 is one of the reference models on the leaderboard. Use an interactive installer (npm install -g autoprompt-skill && autoprompt) to install the skill automatically based on your coding agent (supports Claude Code, Codex, OpenCode, Kilo Code, VS Code Copilot, Prime Agent).

pathwaycom/arc-task-gen

Pathway’s ARC Task Gen creates new ARC-AGI-1-like reasoning tasks. It is distribution matched to the publicly available evaluation data. This repo goes along with the paper on our new model architecture BDH-CQ, which incorporates iterative reasoning and in-context learning through evolving recurrent memory in a continuous latent space. We find that a 150M parameter BDH-CQ model achieves 29.5% pass@2 on the ARC-AGI-1 evaluation set at $0.0007/task, which is 11x more efficient than GPT-5.6 Luna Low. The generated tasks are in the same ARC JSON format as the originals, so they are compatible with existing evaluation frameworks.

odysseus-dev/odysseus

Odysseus is an open source, self hosted, AI workspace, bringing chat with any language model, AI agents with tools and MCP integration, deepweb research, a document editor, emails (IMAP/SMTP) with AI triage, notes, tasks, and calendar. The Cookbook shows you what model to use depending on your hardware. Compare compares 2 model outputs side by side. Everything stays on your computer. AGPL-3.0 license.

nexu-io/open-design

OpenDesign is an open-source version of Claude Design, a closed product by Anthropic. It’s a native desktop application for macOS and Windows, launched in late April 2026, with almost 90 000 GitHub stars within four months. It is a design studio for AI coding agents. You describe what you want, the agent creates it on the fly inside the app, with a preview, and you keep iterating until you’re satisfied. The agent can create web and mobile prototypes, live dashboards, slides decks, images, HyperFrame motion graphics and videos. It supports many coding agents including Claude Code, OpenAI Codex, Cursor, OpenCode, GitHub Copilot, DeepSeek Harness, Kiro, Kimi, Hermes and more than 20 others. It also supports BYOK (Bring Your Own Key) for any OpenAI compatible endpoint. It supports DESIGN.md design systems so you can set up your brand colours, fonts and components, and all the designs created by the agent will follow the design system.

missuo/herdrm

herdrm is a native macOS app that provides a graphical user interface (GUI) for herdr.dev, a background service that runs coding agent terminals. It displays all of your local and remote herdr instances in a sidebar, including Claude, Codex, Gemini, Grok, OpenCode and more. You can see the live status of each agent, such as blocked, spinning, or checked off. Click on an agent and a live terminal will connect to the agent’s PTY. It also sends system notifications when an agent is finished on any device or needs interaction.

JustVugg/colibri

Colibrì is a pure C inference engine (no engine dependencies, single file core!) for cutting edge MoE models on your personal computer. It merges storage, RAM and VRAM into one level of memory: the dense shared core (~17B params, ~9.9 GB at int4) remains in RAM, and 19,456 routed experts (~19 MB each at int4, ~370 GB on disk) are streamed as needed with a per layer LRU cache. Models supported: Kimi K3 (2.8T) by Moonshot AI, GLM-5.2 (744B), Inkling (975B), DeepSeek V4 Flash (284B), Qwen3.6 (35B-A3B), OLMoE (7B). Built-in OpenAI compatible server allows clients speaking the OpenAI API (e.g., Claude Code by Anthropic) to connect to a locally hosted model. Backends: CPU, CUDA, Metal, NUMA. Web dashboard featuring realtime stats and live visualization of the entire set of 19,456 experts.

img2threejs/img2threejs

img2threejs is an AI agent skill for Claude Code, Codex and OpenCode, published in mid-July 2026 and reaching 12 400 GitHub stars in five weeks. It takes one reference image of an object and produces a TypeScript THREE.Group factory that rebuilds the object from primitives, procedural shaders and generated geometry — animation-ready, with pivots, sockets and colliders baked in. The pipeline runs eight stages (blockout → structural → form → material → surface → lighting → interaction → optimization), each gated by a vision review before advancing. Helper scripts are pure Python 3.10+ standard library — no pip install, no extra dependencies. The model spends tokens only on visual judgment and code; JSON validation, comparison sheets and manifest generation all run in deterministic scripts. Supports general hard-surface objects, characters (with facial landmark fitting and per-region confidence reporting), and CS2-class weapons (dedicated component contracts and family-specific review gates).

fufankeji/deepseek-harness-studio

DeepSeek Harness Studio is an Electron-based Desktop App developed by 赋范空间 (BeyonData), which is a fully visualized client of DeepSeek Harness with a plugin marketplace, a preset gallery, and visual agent running environment. It supports MacOS arm64 and Windows x64, open source and alpha version.

CopilotKit/openbot

OpenBot is a free and open source agent platform developed by CopilotKit. It provides each AI agent their own self contained computer with their own browser, files, and access to only the tools you provide. Supports any agent built using LangGraph, Mastra, CrewAI, or a custom endpoint via the open AG-UI protocol. All tool calls are routed through a gateway which makes decisions, logs, and then executes or denies (and explains why). Currently in alpha development.

browser-use/macos-harness

Here’s a new project from browser-use — a tiny Python process that exposes 6 low level primitives for interacting with macOS to a LLM: take screenshots of windows, send keyboard/mouse actions to a specific application PID (via Apple Accessibility), access AppleScript, a full browser via CDP, and the local filesystem. During interaction, the model learns and writes the necessary code along the way. No prior tooling needed for any specific application. Only available on macOS, MIT license, anonymous data collection.

bowenliang123/dsh-context

The dsh-context is a DeepSeek Harness plugin which provides the /context slash command and the Context tab. The tab presents the current contents of the model’s context window (system prompt, tool schemas, user messages, injected context, assistant messages) in a 6 color stacked bar chart relative to the model’s context window size. In the History view, each bar represents one model request, allowing you to see how the model’s context window fills up over time and when it was compacted/pruned.

vibeinging/deepseek-harness-desktop-app

DeepSeek Harness Desktop App is a local AI desktop environment based on DeepSeek’s official AI agent runtime dsh. It provides an AI coding assisting workspace that contains DSH sessions, project management, file management, embedded multi-tab browser, Git worktree isolation for parallel AI coding, Canvas for editable documents with history, and Office artifacts creation(such as DOCX/XLSX/PPTX/PDF) in a single desktop window. It has both Chinese UI and English translation version. At this moment, it is only tested in macOS Apple silicon. The Windows 64-bit will be done soon.

SigmanticAI/apex-inference-chip

APEX is an open source RTL implementation of a complete transformer decoder layer, developed by Sigmantic AI using their autonomous verification agents. It runs Qwen2.5-0.5B at 0.56 tok/s on actual FPGA hardware, which is 140x faster than when first deployed. All blocks are bit-exact-verified against a NumPy golden model prior to being released.

myYangyunfan/dsh_desktop

dsh_desktop is a stand-alone Windows/macOS desktop client application of DeepSeek Harness (dsh). It packs the whole dsh runtime (with all officially supported plugins) into a single binary file. It comes with an embedded Node.js runtime, so you just need to download and double-click, without using terminal, npm, or installing Node.js. It also bundles the dsh-openclaw-bridge plugin to connect dsh sessions to Tencent WeChat and QQ chatting platforms by integrating OpenClaw and Tencent’s official bot SDKs. It also provides a balance widget to show the realtime cost of each session, Windows system tray notifications upon agent task finishings, desktop pet (a little whale), and a watchdog to recover from crashes. It supports Chinese UI language. It comes in both portable exe and installer package.

Electricitysheep/dsh-handbook

A 15-chapter community handbook for DeepSeek Harness (dsh)—DeepSeek's official AI agent runtime, open-sourced on August 13, 2026. Written primarily in Chinese with an English translation, it covers installation, Web and headless modes, plugin development, performance tuning, caching strategy, and an ecosystem analysis of 1,804 plugin repositories and 780 community discussion threads. Every command is tested and verified on real hardware.

cinderline/northcinder

NorthCinder is an open-source MCP server (0.1.2) that you install on your machine to be used with your AI app. NorthCinder enables your AI agent to access actual online stores (Shopify, WooCommerce, eBay, Etsy, Amazon) and provides a shortlist of results sorted according to your preferences along with an explanation in a machine readable format. The ranking algorithm does not take into account any seller payments and each response contains:

NanmiCoder/dsh-agent-teams

dsh-agent-teams is a DeepSeek Harness plugin by NanmiCoder that enables captain-led multi-agent coordination inside a single DSH session. The current session becomes a captain that spawns durable sub-agents with distinct roles, breaks the goal into dependency-aware tasks, routes them through a shared scheduler, tracks state on disk, and consolidates results. It includes ten coordination tools, direct messaging between members, automatic retry for stranded tasks, and a live Web UI with a real-time task graph.

jdshfhds/northcinder

NorthCinder is an open source, self hosted MCP server for shopping agents. It can interface with shops like Shopify, WooCommerce, eBay, Etsy, Amazon and more. It will rank items based purely on what’s specified in your brief (price, shipping time, features, seller reputation) and flag sponsored offers below organic ones. It’s transparent about how it ranks, providing machine readable explanations for each result and keeping an audit log of its decisions locally.

hairyf/deepseek-harness-desktop

A community-built Tauri desktop app that packages DeepSeek AI's official agent platform (deepseek-ai/deepseek-harness, 158 K stars) into a 5 MB native installer for Windows, macOS, and Linux. On first launch it bootstraps a bundled Node.js runtime and downloads the Harness core automatically; every subsequent launch syncs to the latest upstream release. All sessions, profiles and settings stay on your machine at 127.0.0.1:3080. It went viral because it removes the single biggest barrier to DeepSeek Harness: a complex developer setup.

decionis/agent-safe-pipeline

Agent-Safe Pipeline is a TypeScript reference implementation built by Decionis that adds an independent authorization layer to AI agents. The agent captures its proposed action as an immutable intent, the Decionis policy service evaluates it and returns ALLOW, ESCALATE, or BLOCK, and a SafeExecutor runs the action only after a valid single-use grant. For escalations, a Presence service collects verified human approval before Decionis re-evaluates. The repo includes runnable examples: refunds, GitHub deploys, procurement, and MCP tool gating.

ysr666/dsh-vision-router

DSH Vision Router adds 14 visual tools to text-only agents in DeepSeek Harness: image Q&A, object grounding with pixel coordinates, object detection, cropping, SVG tracing, pixel diff between screenshots, long-screenshot OCR, foreground extraction, dominant color analysis, and HTML screenshots. When an image is pasted in chat, the plugin automatically routes the request through a built-in free Gemini 3.7 Flash vision chain — no configuration, no Python required, and image turns work exactly like ordinary tool-calling turns.

yjh051108/dsh-routing-suite

dsh-routing-suite bundles two components for the DeepSeek Harness agent framework: a runtime management layer and a task-aware thinking mode preset. The preset reads the nature of each incoming task and routes it to the appropriate behavior: planning mode for structured analysis, reactive mode for fast execution, and deep thinking for complex problems. The community developer tested the router across 23 scenarios and measured 96% routing accuracy and 100% task convergence. The suite crossed 3,000 stars within days of release, riding the wave of interest in the DeepSeek Harness beta.

yetone/cumora

Cumora is an open source collaboration tool that lets AI agents participate equally with humans. They have roles and memories, take ownership of tasks from a shared Kanban board, avoid stepping on each others’ toes when coordinating, and can even exchange emails. There are two brain modes: Cumora Cloud (agents live in managed, per-agent pods on the OpenAI Responses API) or BYOA (you connect your own Mac or VPS to the server with your local Claude Code or Codex CLI, so we don’t have access to your provider keys).

Hisn00w/ASu-skills

ASu-skills is a plugin set for Codex (OpenAI code agent) and DeepSeek Harness that helps with the process of campus job recruitment in China. With it you get access to these five skills: finding and contributing to real open source projects by making small pull requests (for portfolio), rewriting your experience and projects to fit a certain job position, generating an editable HTML résumé with PDF export, cloning the author’s high density tech resume template, and parsing emails and screenshots into a structured job application tracker. This toolkit has been starred almost 1000 times within 5 days during the fall job recruitment period in China.

dsh-market/dsh-market

dsh-market is a DeepSeek Harness plugin which provides an App Store like UI within DeepSeek Harness Settings. With this enabled you can search and view a curated list of 800+ plugins with categories, ratings and AppStore like screenshots. There’s even a separate section for themes/skins. You can update, backup and restore your list of plugins as JSON file and enable/disable them on the fly without having to restart. Plugins can only be installed from our curated awesome-dsh-plugin registry to prevent installation of random packages.

dataelement/dsh-desktop

DSH Desktop is a community desktop app built by the dataelement team that packages the local DeepSeek Harness web experience inside an Electron shell for macOS (Apple Silicon and Intel) and Windows. It auto-starts a local Harness instance on a random loopback port, manages the process lifecycle, and persists profiles, plugins, and sessions across upgrades — all without the user touching a terminal. macOS builds are code-signed and notarized by Apple. The project ships preset packages as portable .dshpreset files and comes pre-loaded with the dsh-market plugin store.

Anionex/dsh-vision-toolkit

DSH Vision Toolkit is a plugin of DeepSeek Harness that brings 10 visual features such as image Q&A, pixel-level object grounding, object detection, cropping, tracing, pixel difference between screenshots, long-screenshot OCR, foreground extraction, dominant color extraction, HTML screenshots into DeepSeek Harness for text-based models. DSH Vision Toolkit comes with an integrated free Gemini 3.7 Flash vision engine(300 images/day/machine without registration) or allow you to use your custom OpenAI or Anthropic compatible model.

xiaobright/dsh-anchored-standard

This is a community project that is not associated with nor endorsed by DeepSeek. This is a DeepSeek Harness agent framework project called Anchored Standard. The problem being solved is that DeepSeek V4 Pro has a strong bias toward the initial tool catalog presented. If we start with the full Standard catalog, then it starts adding hedges and being less direct. By using the exact same 2 tools as the Minimal preset, it anchors the initial response to be more direct and focused on completing the task. Once the first tool is called, it automatically switches to show the complete Standard catalog. In Project2 evaluations, it got scores of 98–99 compared to 91–92 for the Standard preset.

vercel-labs/eve-software-factory-template

Foreman is an open-source AI software factory template powered by the eve framework. It integrates with GitHub and Linear and processes all labeled tasks through a series of 4 agents: a Classifier to sort requests and ask follow-up questions when needed, an Analyst to generate an implementation plan based on a live clone of the repo and define acceptance criteria, an Implementer to carry out the plan in an isolated environment, execute project-specific tests, and commit changes to a branch, and finally, a Reviewer who examines only the diff (not the Implementer’s thought process) and provides a judgement along with supporting details. When not running, Foreman keeps a live factory memory, a collection of notes about the repo that all agents start with. Deploying Foreman is as simple as clicking the Vercel button to automatically set up GitHub & Linear integrations.

elie222/rakazo

Rakazo is an open source replacement for Grok Bot by Inbox Zero Inc. (the makers of the Inbox Zero email management app) which can be installed on your local computer or server. Each bot has its own private memory, sandboxed computer environment (Docker locally or E2B remotely), long term memory, scheduled routines, and subagents which can be instantiated in a single turn. The entire product can run without any need for Rakazo’s cloud services. This is a completely self hosted solution. It offers a Pi catalog of models to choose from (currently supports OpenRouter API keys, ChatGPT Plus/Pro with device code auth, GitHub Copilot, and SuperGrok. Not yet available in UI is Claude Pro.) Web, Electron desktop, and Expo mobile. Built with Cursor and Grok 4.6.

ccch1mneyyy/dsh-TUI

This is a plugin written by the community for DeepSeek Harness, an open-source AI coding agent developed by DeepSeek. This plugin brings a full screen terminal experience to DeepSeek Harness, with a status line of the live AI coding agent, a thought stream, a context progress bar with TPS, and session time-rewind functionality by double tapping the Esc key, just like how Claude Code works. This plugin was selected as one of the best from the community by DeepSeek Harness official WeChat public account, so it gained lots of stars quickly.

anywhere-labs/deepseek-harness-desktop

Anywhere Labs has developed a desktop application as a wrapper of the official DeepSeek Harness agent framework(DeepSeek open sourced the Harness agent framework on August 13th 2026 with MIT License). It can automatically start and manage the local Harness service, integrate with system tray, and provide an installer to download. You don’t need to install Node.js or use CLI to use it. Main readme and UI are Chinese. In the future, we will support plugin center, mobile remote control (iOS/Android), and integrate with some IM softwares(e.g. WeChat, Feishu, Discord…).

alchaincyf/deepseek-harness-orange-book

A Chinese guide for the DeepSeek Harness agent framework, written in 24 hours after DeepSeek open source it on August 13th, 2026. Author @AlchainHust (a.k.a. HuaShu), an AI native independent programmer who never typed a line of code, has documented all the system prompts, the 129 lines of the boot config dump, three raw session logs(normal mode, PTC mode, creation mode), a video of the AI self-creating a new tool in 19 steps(from 32 to 33 tools), how much it truly costs to run a PTC session and 35 optional packages in the repo not included by default. A small english description can be found in the readme, the rest of the content are written in Chinese.

zhu1090093659/dsh-web-ui

BSD-3-Clause set of plugins and skins for DeepSeek Harness (AI Code Assistant, DeepSeek AI) created by independent developer zhu1090093659. The set contains task board with cron jobs, git graph, side panel with files/diff, real-time tokens and TPS statistics, remote control on a mobile device with a pairing by QR code, SSH management, 8 skins including Windows XP Luna and Trading Terminal with a live ticker.

i3T4AN/KADATH

KADATH is an open source evolutionary multi-agent runtime written by single developer i3t4an on top of Hugging Face smolagents. It uses a biologically inspired approach to evolve solutions to tasks, where a population of agents compete over several epochs to solve your task, each epoch being evaluated independently with respect to a frozen benchmark, and the kernel then applying Darwinian principles of natural selection (preserving the top 30%, allowing the mid-tier agents to mutate according to the successful mutations of the top agents, and eliminating the bottom tier). Over a reported 10 epoch run, the maximum fitness increased from 18 to 91, demonstrating that the system improved over time through competition between agents, without the need to tune individual agents.

deepseek-ai/deepseek-harness

DeepSeek AI (creators of the r1 and v3 models) has open sourced their agent harness (called DeepSeek Harness or dsh). It’s built on one core principle: all of the components (models, tools, web ui) are plugins that can be installed and uninstalled at will. At its core is a component framework called Cordis that supports hot reloading of components. Currently in developer preview, all you need to do is run a single npx command to get started.

milind-soni/OpenMausBot

OpenMausBot is a desktop macOS app built by Milind Soni, inspired by Grok Bot (xAI), but completely independent and unrelated to xAI. Each contact in this Telegram-like side bar is an actual AI agent (Claude, Codex, or Grok) that runs locally on your mac using your personal credentials. The agents can be provided with a cloud Linux desktop that they autonomously control and are able to integrate with over 500 apps (Gmail, Slack, Github, Notion, etc.) using Composio. They also show an inline approval card for every single command they execute in a terminal or file they edit.

dmmulroy/anti-slop

A Oxlint plugin written by developer Dillon Mulroy that includes 15 rules intended to prevent many common TypeScript and JavaScript patterns made by AI coding agents. Rules include no-module-mocking, no-unknown-returns, no-chained-type-assertions, and other "short-cuts" taken by AI coding agents when writing 'working code' instead of evidence based code. This project is intended to be vendored (copied directly into your project) instead of using a hard dependency on npm so that you can customize the rules as needed.

0xnyn/airship

Airship is a free and open-source CLI (airship.design) that provides a visual design layer over top of any running dev server without requiring you to modify any project files, build scripts, or dependencies. When you click an element on the screen, Airship detects where in your codebase the element was generated in the dev build and sends that information along with the user’s request and a few frames to Claude Code, Codex, or OpenCode to make the edit. Airship supports desktop and mobile frames that can be edited simultaneously.

UditAkhourii/neuroarxiv

This is an open source Claude Code skill created by independent developer Udit Akhouri. It takes your problem description, parses it into arXiv categories, uses HTTP to request arXiv papers from export.arxiv.org, then reads each paper in isolation (using one LLM call per paper so they don’t contaminate each other) to score them and reach a consensus on a recommended, cited solution. In their eval, the creator found that it was able to refuse to use a retracted solution that appears in the same list, something that regular search and cold-reasoning weren’t able to achieve.

tanishqkancharla/calldiff

This command line interface (CLI) tool diff function call stacks in git commits. It shows which callees have been added/removed/relocated in relation to an entry point. Uses Tree-sitter (AST-based) and works with 22 programming languages. This helps with an agentic code review where git diff lines are obscuring the structure of the changes. You can register it as an MCP server, so it is used by AI agents. The output is also available in JSON format.

shadcn-ui/chatbot-template

The shadcn-ui team is the creator of one of the most popular UI component libraries. They have created a new chatbot template using the Vercel AI SDK, which comes complete with stream chat, provider search, and a human-in-the-loop questionnaire. This template deploys to Vercel in one click, and doesn’t require any API keys when deploying to Vercel because Vercel takes care of authenticating for you.

gvzdv/claudish-to-english

This is a Claude Code plugin that displays all responses from the AI in plain English using a local LLM with Ollama. Your original message and transcript remains unchanged. This plugin fails open, meaning that if Ollama is not installed or times out, then Claude’s original responses will be displayed. (You can also optionally set up another hook to convert Claude’s markdown responses.)

antirez/h3.c

h3-metal is a native C inference engine for Apple Silicon that allows you to run MiniMax H3, a 33 billion parameter video synthesis model developed by Chinese AI company MiniMax, completely on-device using Apple’s Metal GPU framework. Developed by Salvatore Sanfilippo (creator of Redis), h3-metal allows you to generate videos from text, conditionally generate the first/last frame, and reference-based video synthesis, without having to make any cloud API calls.

KKKKhazix/human-writing

I installed this skill into my agent and the difference was immediate — outputs stopped sounding like a summary machine. The project takes a clever approach: instead of banning specific phrases, it detects underlying structural patterns — 'report voice', uniform sentence lengths, over-explaining the obvious. Version 1.1.0 moved to pattern-based detection rather than character-string bans. The lite version fits in ~2,000 characters, so it works directly in a chat window without any setup.

ShawnPana/phone-harness

Paired my iPhone with macOS Mirroring and ran this in about fifteen minutes — no jailbreak, no developer account required. The core idea is elegant: ~500 lines of Python capture the mirrored window as screenshots, Vision OCR finds tap targets, then HID-level events send input back to the device. The README describes it as 'the poor man's DOM' which is exactly right. It's stateless — no background daemon, each run is a clean slate.

xoreaxeaxeax/asm-hall-of-shame

I spent an embarrassing amount of time reading this. xoreaxeaxeax — the same author behind ropeme and sandsifter — catalogues x86 instructions ranked by worst-case execution latency, the exact opposite of what performance guides normally teach. The top entry is fxrstor64 with PCIe starvation: ~198 billion cycles per instruction. They managed to make rdrand take 5,579 cycles by depleting the hardware entropy pool — a detail I never would have thought to look for.

MengTo/kage

MengTo (Design+Code) dropped a single HTML file that runs an interactive night walk through a Kyoto mountain temple — no build step, no npm install. I cloned it and opened index.html immediately. The WebGL camera moves as you scroll through five chapters, layering procedural fog, rain, lanterns, and falling leaves over pre-rendered scene plates. The no-build constraint is a deliberate design choice: the whole experience is one file you can study, copy, and remix.

jd-opensource/JoyAI-Video-Edit

JD's research team open-sourced their real-time video editor: a 16B multimodal diffusion transformer that edits frames as they stream in, hitting 30.19 FPS at 720×1280 resolution end-to-end. The README shows five demo categories — style transfer, subject replacement, background changes — all running on live input without buffering the full clip first. I was skeptical about 'real-time' claims, but 30 FPS at 720p from a 16B model is genuinely impressive.

sv-number/mcp-server

Open source MCP server which connects your AI agent to the commercial SMS code provider. Your agent will be able to request a disposable private number of any country, to wait for the SMS code and complete or cancel the number activation process. All this is done via stdio without using SDKs. The provider is commercial, the free tier is not offered.

SMNETSTUDIO/WeChat-AI

SMNETSTUDIO’s open source self hosted WeChat-AI platform enables running LLM based role play chatbots on WeChat using Tencent’s iLink API and any OpenAI compatible language models. Around one third of the docs/UI are in Chinese. End users can create their own AI persona with unique personality and conversation history, while Admins can manage the bots, tokens and moderations via the admin dashboard. You can also visually define the conversation flow using the Chatflow editor without coding. Multi node support and Cloudflare Worker LB are also available.

oil-oil/codex-deepseek-subagent

An installable Codex skill using npx. with chinese readme file. Register DeepSeek v4 Flash as a native custom subagent for OpenAI Codex Desktop App. Once installed, the main agent will be able to send tasks to DeepSeek automatically. The skill handles task routing, securely storing credentials, and automating acceptance tests.

NikoDemon80/ComfyUI-H3-Motion-Context

This is a ComfyUI custom node pack that allows you to chain video clips for MiniMax H3 AI Video Generation Model. H3 jointly generates images and audio, but the model does not support chaining clips. This plugin allows you to feed the latent vector from the end of the prior video clip into the current one. It preserves the motion speed, direction, and audio continuity between the clips. The patches will validate the maths during startup and ensure they are still valid by comparing it with the latest ComfyUI code. If the code has changed and invalidates the patches they won’t run at all.

mewamew/my_ai_town

A life sim game with Chinese readme file. Godot 4.7. The game has LLM driven AI agents as citizens. Each citizen has a location, can remember things, have relationships, and make autonomous decisions. You can log in as a player avatar, make town wide announcements, change the weather, and see the reaction of the citizens.

Sateezg/codex-bridge

codex-bridge is an open source Claude Code plugin that connects Claude to the Codex CLI. It provides 2 features: image generation/editing (favicons, OG images, mockups, icon sets) with gpt-image-2 and task automation/delegation with GPT-5 subagents. These use your ChatGPT subscription credits rather than Claude’s context and token balance, allowing you to keep Claude focused on high level decision making.

Larryvrh/ComfyUI-MiniMax-H3-Turbo

ComfyUI-MiniMax-H3-Turbo is a plugin that introduces two custom nodes in ComfyUI, which can accelerate the generation process of MiniMax H3 (a Chinese company model) from around 20 steps to 4–8 steps. It supports different types of MiniMax H3 checkpoint (bf16 full, int8, prune) and can be installed with ComfyUI-Manager.

huangserva/ComfyUI_MiniMaxH3_Director

This is a mirror repo of AIMixer/ComfyUI_MiniMaxH3_Director. It contains 5 validated ComfyUI workflow JSON files for MiniMax H3 Director, an AI video generator developed by Chinese company MiniMax which can generate videos with synchronized audio. Workflows include text2video (T2V), image2video (I2V), first last frame (FL2V), reference-based video (R2V) and timeline video editing (V2V). Tested on NVIDIA RTX 4090 with ComfyUI 0.30.0.

eternityspring/shuohao-skills

shuohao-skills is a standalone skill pack for Claude Code and Codex. The novel-characters skill takes any novel as input and outputs a character bible (character profiles, AI art prompts, voice prompts, and 3-view turnarounds). There’s also an install script that checks if Claude Code and/or Codex is installed and then symlinks the skills, so that when you run git pull the skills will automatically update, without needing to be reinstalled.

xmarre/ComfyUI-Spectrum-MiniMax-H3

Spectrum Apply MiniMax H3 is a ComfyUI node that reduces the number of full transformer evaluations during MiniMax H3 video generation. It fits a Chebyshev ridge regression model to actual transformer hidden features and uses it to predict those features on selected future steps — cutting compute without changing the final generation heads, audio, or video reconstruction on every step. The README documents the speed/quality tradeoff honestly, including known failure cases like anatomy distortion and audio degradation with aggressive settings.

T8mars/comfyui-minimax-h3-audio-T8

A ComfyUI plugin that introduces 14 nodes for complete audio/video AI generation using MiniMax H3. This plugin will ensure correct functionality for video+audio generation, text2video+audio, image2video+audio, ff/lf video and ref based generation. It also ensures that audio/video is handled correctly where the original ComfyUI MiniMax H3 integration is lacking. All without needing any extra pip packages — it works using only the original ComfyUI dependencies.

openTrinity/mycontext

MyContext is an open-source desktop application (Electron + TypeScript) that builds a personal knowledge graph from your instant messages, documents, and meeting records. Instead of every AI session starting from a blank slate, it provides a persistent local context layer that language models and local agents can query. A "digital self" feature reads incoming messages and drafts replies in your communication style. The project gained attention for keeping all data strictly local — nothing leaves your machine.

joeseesun/qiaomu-seo

The qiaomu-seo agent skill works on Claude Code, Codex, Cursor, or any other compatible agent platform (e.g., an agent client) with a simple one-line command. This then gives the agent complete access to professional SEO tools such as technical SEO, keyword mapping, crawlability check, site migration, fixing SEO tags (JSON-LD/Canonical), etc. Most importantly, it also includes independent SEO testing for various AI search engine interfaces (e.g., Google AI Overviews, ChatGPT Search, Microsoft Copilot, and Perplexity).

google-gemma/gemma-translator

Google Creative Lab (Google’s experimental division) has developed an open-source voice translator that can run Google’s Gemma 4 model completely offline, using the LiteRT-LM runtime. No internet required once configured! It captures speech, transcribes, translates using Gemma 4, and speaks the translated text back in the user’s target language. Source code, deployment scripts and even 3D printing files are available!

Accio-org/RealReplicaBench

RealReplicaBench is a benchmark developed by the Accio team from Alibaba International. It evaluates AI agents on 107 real-world task workflows including product listing, shipping logistics, spreadsheet editing, and API/MCP calls. These tasks include CLI, web browser, document editing, and API/MCP interaction, with verifiers for each task type. The current leaderboard shows Claude Opus 5 achieves 56–62% task completion rate based on the evaluation harness used.

ZzzLc0405/photo-abstract-editorial

This is a structured SKILL.md prompt for Codex, which tells Codex how to convert any image you upload into an editorial vertical layout. It includes original photo, abstract art generated by the original photo(not filtered or painted) and a poem title. This repo is mainly in Chinese, but I also provide some English prompts in the references folder. It got more than 1200 stars in the first week, one of the most popular Codex skills.

SandAI-org/MAGI-2-preview

MAGI-2 Preview is an audio-visual generation model with a single unified architecture for both T2V and I2V (with text prompt or text prompt + image reference respectively) from Sand.ai consisting of 114 billion parameters, with only 6 billion activated at each token via a Mixture-of-Experts approach (MagiMoE). This model is capable of generating 10 second videos with synchronized audio from either text prompts or text prompts + image references. This model performs a preview denoising stage followed by upsampling to 1080p resolution. Model weights (307 GB) are available on Hugging Face with Apache 2.0 license.

criptogus/HermesOffice

This product is based on the GenOffice project (Apache-2.0), with Hermes Agent AI included on top. Each program has the same AI panel integrated. The AI system works on a block-level basis, analyzing the structure of documents and making exact changes within blocks. The rest of the file remains untouched, which means that all bytes are the same.

AMAP-ML/LongHorizon-Harness

LongHorizon-Harness is an open-source state management and execution tool by the AMAP-ML team for running on top of existing tools like Claude Code CLI or Codex CLI. It introduces a role separation between a Manager who knows the target goal and verified history, an Executor who executes each step with a clean context, and an Auditor who verifies results before committing changes to the verified history. On WeaveBench, task success rate increases from ~50% to ~80%. On OSWorld 2.0, the desktop full task completion rate is tripled. This tool climbed to #1 on Hugging Face’s Daily Papers weekly ranking on August 6th, 2026.

thebuggeddev/anatomy

Open source Interactive 3D human anatomy viewer application created using Three.js + GPT 5.6 Sol. Hosted on OpenAI platform (vinext/Cloudflare). This application allows you to visualize a 3D body model and communicate with GPT using the interface. Collected ~1800 stars in 5 days which is quite high for an anatomy viewer app.

fuxicodex/Fuxi

FuXi is a terminal AI coding agent that can read, edit and execute commands on your projects just like me and my Claude Code agent. It can also easily swap out AI providers such as Anthropic, OpenAI, Gemini, DeepSeek, Qwen and more without having to reinstall anything and will automatically failover and use cost optimized routing. Simply say “install FuXi” to your agent to run the installation script from fuxicode.com and then start it up and run “/model” to select an AI provider. FuXi is all in one file and does not require any dependencies.

mikiarlo3/awesome-growth-hacking-skills

This is an open-source directory of curated AI agent skills for growth hacking and marketing. The directory includes skills for SEO, AI search, content, email, paid media, sales, CRO, social media, and GTM flows. Optimized to work with AI agents such as Claude Code or Cursor. Over 400 stars within first 24hrs.

magicrew/doc7

doc7 uses any OpenAI compatible vision model (such as LM Studio or Ollama locally or a remote server) to convert PDFs, Microsoft Word & Powerpoint documents, scans, screenshots, graphs, diagrams, and equations to Markdown. Instead of using traditional OCR methods which attempt to extract text from images, doc7 uses an image-to-text approach where it sends the entire page to the model and asks it to reconstruct all of the semantic information such as tables, equations, figures, and references. On the “Attention Is All You Need” paper, doc7 was able to get all 15 facts correct whereas MarkItDown+OCR got 9 and Docling got 3. This tool is gaining popularity since there is often a need to convert documents when creating RAG systems or agents that work with complicated documents.

leonickson1/Swiftlet

Swiftlet is an open source Swift & Metal inference runtime for the Qwen3-Next/Qwen3.5/3.6 Mixture-of-Experts family models developed by Alibaba’s Qwen team under the Apache 2.0 license. By loading only the tiny dense core of the model into memory and streaming the large sparse experts from disk on demand, it enables running a 35B model using just 2.6GB of memory on an M5 Mac at 7–11 tokens/sec. On an iPhone 17, it can run the same model in approximately 2.5GB of memory at around 1 token/sec. This project gained popularity due to its unique ability to stream large MoE weights from disk instead of keeping everything in memory.

Binaryify/open-kimi-ppt-skill

This is a non-official Kimi Slides Skill developed by developer Binaryify after reverse engineering Kimi’s presentation platform. Not associated with Kimi or Moonshot AI. This skill allows AI code agents to make fully functional presentations. Always outputs 2 files. An editable PPTD project file (which can be viewed & edited using a local browser editor) & a final PPTX file with embedded fonts & fade transitions. Also allows you to take a PPTX file & turn it into a PPTD project which can then be edited. Compatible with any agent that adheres to the SKILL.md spec (e.g., Codex, Claude Code, Cursor, WorkBuddy).

Accio-Lab/RealReplicaBench

RealReplicaBench, developed by the Accio team at Alibaba International, is a public benchmark for long horizon business workflow tasks such as browsing the web, using command line interfaces (CLIs), editing files & spreadsheets, calling APIs/MCPs, executed within new containers with either a deterministic or LLM assisted verification process. This benchmark is getting popular recently as it provides a transparent and repeatable method for comparing AI agents on real world tasks instead of trivial examples.

sophiamyang/finger-frame-effect-ai

Hand tracking app that overlays an AI generated video on a video of the two-handed finger-frame gesture. When a video is uploaded, it is fed into Google’s Gemini Omni Flash video model, which restyles the video using a selected style (anime, claymation, watercolor, 3d movie animation, custom prompt) and then recreates the video with the same movements. This is achieved by tracking both hands in each frame with MediaPipe Hand Landmarker and blending the new video into the tracked frame using a quad. It’s a purely browser based experience, just need a Gemini API key.

MarcosSete/awesome-free-ai-course-notes

A handpicked collection of official free Machine Learning / Artificial Intelligence (ML/AI) course notes in written format only, from universities such as MIT, Harvard, Princeton, Stanford, UC Berkeley, Caltech, Cornell, Oxford, LMU Munich, KAIST, USP Brazil, and more. The selection is strict: all are in written format only, not slides or videos, and must be easy to read like a textbook chapter. All notes are official, permanently free, and can be accessed without logging in. See also the Awesome Free AI Books repository (featured here in late July), created by the same author.

firecrawl/anydoc

anydoc is an open source Rust library for converting Word, PowerPoint, Excel, PDF, and 8 other office documents into clean, structured Markdown, under 5ms.

elayadesign/ai-design-skills

Library of skill files (structured markdown rules) to extend the capabilities of your Claude Code, Cursor, Codex, or Windsurf AI agent. This first skill teaches the AI how to build out a landing page from start to finish (intake questions, page structure, conversion copy, complete design system including typography, spacing, radius, motion). If you don’t have access to a terminal, you can attach this skill file to any Claude.ai Project or chat. Received 358 stars in first few days!

disler/super-simple-software-factory

A python utility which ensures reproducibility in multi-step AI agent workflows. This means that each run will produce identical results. The steps and completion conditions are defined by regular code, but within each step the ai-agents perform their cognitive work in isolation, passing control to the next step once they’re finished. To install, copy the `.claude/skills/sssf` folder into your repo and run `/sssf install` inside Claude Code, and it sets everything up for you.

aigclink/geolook

Self-hosted open source platform for Generative Engine Optimization(GEO). GEO is a process of optimizing AI Search Engines (ChatGPT, Perplexity, Gemini, DeepSeek, etc.) to proactively mention and reference your brand while answering user queries. GeoLook can help you sample AI responses from all 15 engines, diagnose issues as to why your brand isn’t mentioned at all or lower in the search results, create tickets for implementations and acceptance criteria along with automated checks to validate if your brand was mentioned more often after the fix. Data is hosted locally in the work/ folder. No data is uploaded to a vendor cloud. Also support Chinese ai engines (DeepSeek, Kimi, Doubao, MiniMax), can monitor both CN and global markets.

openai/ten-proofs

Ten new results in mathematics and theoretical computer science formally verified using machine checkable Lean 4 proofs. These include binary codes, spherical codes, non-sofic groups, Connes’s Rigidity Conjecture, quantum game repetition, closet vector problem, Erhart volume conjecture, multicolored Ramsey numbers, extremal combinatorics.

genspark-ai/genoffice

GenOffice is an AI native office suite for macOS and Windows of Mainfunc, Inc.(Genspark). It includes DOCX word processor, XLSX spreadsheet editor, PPTX presentation maker and PDF editor. Different from other office suite which adds a chatting box on its side, GenOffice implements AI edit at block level. The AI only edits your desired paragraphs/slides/cells but keeps all other data unchanged bit by bit. All model calls are proxied through Genspark services. No need to install the local API key.

FareedKhan-dev/kimi-k3-in-c

C99 engine for running Moonshot AI’s Kimi K3 (2.78T parameter mixture-of-experts language model) fully on CPU without BLAS or CUDA and without an ML framework. Streams a 1.56TB checkpoint from NVMe and only maintains 8GB of active experts in memory. Gets identical results for all memory sizes between 8GB and 224GB.

Audio8-AI/Audio8_TTS

Audio8_TTS is a multilingual text-to-speech model developed by Audio8. It’s based on a DualAR architecture and has 0.6 billion parameters. With this model, you can produce high-quality speech that sounds like any voice you choose, all from a short audio clip. This model supports 11 languages, including English, Chinese, Japanese, and various European languages. Hosted on Hugging Face with an Apache 2.0 license.

s1dashu/animated-voiceover

A skill for agents to generate full-length animated educational videos based on one sentence, including research, narration, visual design, voice consistency, and clip editing. This skill can generate 1–5 min animated knowledge videos. It helps agents to do research, write narration, choose visual styles, direct multi-frame scenes, set up voice references, and edit clips. There are six default visual styles and support for Codex. The main way of executing videos is by using LibTV CLI and Seedance video generation. Note that the README file is available in both English and Chinese, while the SKILL.md file (skill guide) and other production documents are all in Chinese. This skill is mainly for Chinese-speaking people to produce videos.

xdash/FDE-the-Guidance-Book-of-Forward-Deployed-Engineer

This free ebook (in Chinese) by Fan Bing (XDash), author of Growth Hacker. It covers the Forward Deployed Engineer (FDE) role, popularized at Palantir but now being recruited at OpenAI and Anthropic. It also explains why 95% of corporate AI efforts fail to deliver a positive ROI even when they have working models. It covers everything from identifying opportunities, closing deals, driving deployments, keeping customers, and growing adoption within organizations.

microsoft/skill-recorder

Microsoft has released an open-source tool called Skill Recorder, which allows you to record your actions on the screen (mouse clicks, switching between applications, etc.), optionally with voice commentary (transcribed using the OpenAI Whisper model), and then use the GitHub Copilot CLI to analyze the recording and determine its purpose and sequential nature. As a result, you get a SKILL.md file containing a well-structured process that an AI agent can execute upon request or on a regular basis.

Infatoshi/netherite

This is a self-developed C/CUDA version of Minecraft 1.11.2 which is bit-to-bit identical to the original Java version (same output frames). It also has a CUDA-batched RL environment which can run 7,200 independent Minecraft instances in parallel on a single GPU card. This allows large-scale RL agent training with only a gaming GPU card.

PipeNetwork/kimi-k3-mlx

PipeNetwork has ported Moonshot AI’s Kimi-K3 to Apple’s MLX framework (Apple Silicon). Kimi-K3 is a 2.78 trillion parameter multimodal MoE model (104 billion parameters activated per token) developed by Chinese AI firm Moonshot AI. It has a 1M token context length. This port includes a complete implementation of the text tower (Kimi Delta Attention, Attention Residuals), the 3D aware video vision tower MoonViT, and a multimodal wrapper, which were all numerically tested for parity with the original PyTorch implementation.

perplexityai/numbat

Numbat is an open source tool for monitoring and securing AI agents from Perplexity AI. Numbat listens for events from local AI coding agents like Claude Code, Codex, and OpenClaw and records a standardized event trace of all their actions (files opened, shell commands run, network services connected to). It can evaluate detection rules using CEL expressions and decide to block or report on suspicious multi-step action sequences. It also does this when re-playing past sessions from a stored artifact.

xikhar/persona

Persona is a free and open source desktop application using Electron that shows a realtime animated 3D character (VRM format) synced with the voice playback of your AI code assistant. It listens to the audio playback events of Codex and plays corresponding animations of the character. There’s also a local MCP server included so Codex can control what animation to play and whether to hide/show the window without having to send any audio data across networks.

wassgha/rescript

A web app that transcribes your media using Whisper AI (via transformers.js) and provides per-word timestamps and automatic speaker tagging. Once transcribed, the video can be edited by removing words from the transcript, and the video will be trimmed accordingly. Finally, the video can be exported to an MP4/M4A file without requiring the data to be uploaded anywhere.

gavamedia/deltafin

Deltafin is a research project which enables running Moonshot AI’s 2.8 trillion parameter Mixture-of-Experts Language Model (Kimi K3) on a single Apple Silicon Mac/Linux system by streaming model experts from a local disk cache. Kimi’s K2.x family (on which K3 is based) follows a DeepSeek V3 architecture. The research notes explore the inference methodology relative to DeepSeek V4 models. Deltafin also comes with an OpenAI compatible API server which can be used to replace the ChatGPT API for any application. With an M1 Max & 64GB RAM, deltafin produces approximately one token per 15 seconds. Local mode requires around 1.7TB of disk space whereas streaming mode only requires 215GB of disk space (at the expense of slower inference).

digimata/quill

Quill is a simple menu bar application for macOS which records your calls locally (including your own voice and system sounds), and transcribes them with the on-device Parakeet TDT 0.6B v2 speech model. You get a speaker-annotated transcript of the call (them vs. you), with zero bytes uploaded. Once, you download the models (~600MB) and you’re good to go offline.

achimala/TheLongSilence

Long Silence is a procedurally generated 3D space exploration game that can be played directly in a browser tab, without downloading any assets. This game was written completely by Claude Opus 5 using WebGL2 and includes a custom flight model, auto-exposure, and atmospheric scattering. In this game, you are sent on a survey ship to investigate a mystery; 40,000 years ago, over four days, 900 populated planets fell silent.

yanhua1010/self-media-content-workflow

This is a collection of nine Claude Code skills developed by community user yanhua1010 for self-media content creation, including brief writing, account planning, trend tracking, competition monitoring, native content writing on different platforms like WeChat, Xiaohongshu, X, and short videos, content analysis, and delivery & storage. The system is tool-independent and can be used with Claude Code, Codex, and Cursor using the standard skills cli. There are five required manual checks before publishing, and there is no automatic publishing function.

XYZ-AI-Lab/AxisAgentic

AxisAgentic is an open-source Python runtime from XYZ AI Lab for running AI agents on long, multi-step tasks. Every agent action is written to an append-only trace log, which can replay exactly what the model saw at any stage, evaluate performance across benchmarks like BrowseComp and Humanity's Last Exam, or export curated trajectories for supervised fine-tuning. The flagship example built on top of it is XYZ-Aquila, a deep web search agent that topped common baselines on six agentic search benchmarks. The framework works with any OpenAI-compatible model endpoint.

sunny-glow/Auto-BenchMax

This is an open source pipeline developed by community member ‘sunny-glow’ to generate training data that doubles your LLM’s scores on tool-use/agency benchmarks. You can find the full pipeline, training code (Qwen3-Coder-30B trained using axolotl + DeepSpeed ZeRO-3), and first iteration synthetic dataset (1,312 examples) in this repo. Using this pipeline, we were able to improve our score on the MCP-Atlas public benchmark from 19.1 to 40.4 after one iteration. This pipeline supports both Execute-then-Extract for rule-based benchmarks and Construct-then-Verify for LLM evaluated benchmarks. You can even use this pipeline with just one instruction from Claude Code using the provided skill!

MarcosSete/awesome-free-ai-books

This is an open source curated list maintained by community developer MarcosSete. This repo contains a list of more than 50 genuinely free AI/ML books. All the links point to their official author/publisher resources, none of them are pirated. The list includes books on Deep Learning (Goodfellow, Bishop, D2L), Reinforcement Learning (Sutton & Barto), Natural Language Processing/LLMs (Jurafsky & Martin till 2026), Math for ML, ML Systems, Computer Vision, Generative Models (Diffusion Models, GANs, VAEs), AI Safety and Alignment, Graph Neural Networks, etc.

deerwork-ai/deer-workflow

Deer Workflow is a TypeScript workflow runtime developed by ByteDance as a demo of DeerFlow 3.0. You can specify your tasks using natural language, and then Deer Workflow will use Codex CLI (or Claude Code) to automatically generate and run the corresponding workflow in a step-by-step manner. The workflow has an observable execution process and allows multiple parallel steps without any manual orchestration code.

slvDev/esp32-ai

I ran into this while browsing and immediately ordered an ESP32-S3 — the idea of a 28.9M-parameter language model running on an $8 chip at 9.5 tokens per second felt almost absurd. Pulled the repo, checked the code — it's clean, real implementation, not a demo wrapper. The trick is 4-bit quantization with Per-Layer Embeddings borrowed from Google's Gemma, storing 25M of those parameters straight in flash with only 512KB SRAM on the device. It generates short coherent stories locally, no server, no Wi-Fi required, and what I liked most is the author explains the entire architecture with actual numbers in the README — no hype, just engineering.

Jakubantalik/thinking-orbs

I installed this locally in literally two minutes — npm install thinking-orbs — and six animation states were ready to use: Working, Searching, Solving, Listening, Composing, Shaping. Each one is actually thought through, not just a spinner variant. What struck me most is it's built on pure 2D canvas with no ctx.filter, no SVG filters, no WebGL — the author clearly uses this themselves and cared about performance. Respects prefers-reduced-motion, auto-pauses when the tab is hidden, ARIA labels out of the box — exactly the small things that separate a quality library from a weekend project.

mshumer/Claude-of-Duty

I'll be honest — I opened this expecting another vibe-coding toy. But the code is 55,000 lines across 11 real subsystems: HDR pipeline with cascaded shadow maps in a sampler2DArray, a custom physics engine using binned-SAH BVH, AI soldiers with navmesh pathing, Web Audio synthesis without a single prerecorded sound file. What I found most interesting in the README is mshumer's observation that parallel agent fanout actually performed worse than sequential single-owner passes — that's a non-obvious, honest insight about agentic development. FPS improved from 12-17 to 28-30 across three refinement rounds.

gnipbao/story-to-handdrawn-video

I stumbled across this and immediately started counting how much manual work it replaces. You feed it Chinese story text or a sequence of hand-drawn images and it outputs a 3:4 vertical video with handwritten captions, black-and-white plate reveals that transition to color illustrations, and optional page-flip effects — all driven by Remotion with a TypeScript and FFmpeg pipeline underneath. What surprised me is that it's packaged as an agent skill for Claude Code, Codex, and Kimi Code, so you can describe a scene in natural language and get an animation back. The output is a silent H.264 MP4 at 1080×1440, which is a thoughtful choice — leaves room for voiceover in post.

vercel-labs/scriptc

From the first look it's clear this isn't another toy 'compile TS to exe' project — Vercel Labs is taking this seriously. scriptc compiles TypeScript to a native binary via LLVM or C, with no Node.js or V8 bundled inside. Startup time is 2.4ms versus Node's 47ms, binary size 170-200KB, memory usage 1-4MB versus Node's 67-116MB. It already supports async/await via stackful fibers, monomorphized generics, JSON with runtime-validated casting, fetch, and http/https — and uses differential testing to verify byte-for-byte parity with Node.js. This is beyond the proof-of-concept stage.

VictorTaelin/OptMem

OptMem is a persistent memory system for AI agents. It’s a 426 token system prompt and a single standalone python script that saves memories to a local binary tree. Installing OptMem takes only 1 curl command, and integrating it into your code requires adding just 1 prompt block to your CLAUDE.md or AGENTS.md file. This tool became viral due to its ability to address a fundamental pain point (loss of context after each session) while providing a simple and universal solution.

RongleCat/grok-app

Grok App is an open source unofficial desktop client of xAI’s Grok Build CLI developed by Chinese developer 铁柱AGI (@cgnot996 on X). It provides a GUI for the CLI agent with multi-session, per project access control, media preview and scheduled automations. It became popular since xAI released Grok Build and people were looking for a desktop GUI to interact with Grok Build daily.

makecindy/cindy

Cindy is an open-source desktop + mobile (Electron / iOS / Android) application that allows you to run Claude Code, Codex, and other AI models in a single environment. Cindy enables seamless switching between models, allowing you to keep your agent’s memory, skills, and tools intact. Cindy can be used to control your browser, computer, and phone, and can accept tasks via messages and calendar invites. Cindy is available as open source software (Apache-2.0), and hosts some features on a cloud service which comes at an additional cost.

Kritt-ai/open-kritt

Open-kritt is a self-hosted security research platform created by the Kritt team to orchestrate AI agents (such as Claude Code, Codex, or any other model through the OpenAI/Anthropic API) to analyze code repositories and discover real security vulnerabilities. Instead of querying a single model against all the code, we split the problem into small prompt chains and execute them concurrently using several agents. We then remove duplicates and sort the results by severity. We developed open-kritt based on our internal tooling which has already resulted in over $1.5M in verified bug-bounties (researcher: Blockian).

danilo-znamerovszkij/draw-your-font

An npm CLI & Claude Code skill that turns your handwritten letter photos into a legit font file (TTF/WOFF/WOFF2). Claude’s vision model detects and tags all the characters, then our deterministic code extracts the blobs for proper font metrics (cap height, x-height, descender depth) for a functional font file, not a mosaic. All local, no uploading, accounts, subscriptions.

dacnay816y62-hub/cinema-dna-21x9x3

Generate movie frames based on plot or theme, with three interconnected 21:9 aspect ratio images arranged vertically. Each image follows a structured checklist (character relation, space hierarchy, realistic lighting, color scheme, avoiding CG look). Most of the docs are in Chinese.

mikehasa/agentacct

agentacct is a local first tool which parses session logs generated by Claude Code & Codex on your computer to create a local web dashboard on localhost:127.0.0.1. You can see how many tokens were used by each agent per day, estimated costs (with clear distinction as being estimated based on a local pricing table), and what tasks were performed. There are no accounts, telemetry, or cloud sync required.

SeanJ1ang/design-judge-skills

(Readme, documentation are Chinese main version, also have English README)

KinetiNode/claude-fable-5-system-prompt-clean

Adapted version of Anthropic’s released Claude Fable 5 / Mythos 5 system prompt. The original includes a bunch of Claude specific wiring such as XML norms, MCP/tool interface, product behavior and artifacts which is of no use outside of Claude. This version removes all of that to focus on the generalizable behavioral advice for being honest about uncertainty, effective communication, writing good code, making reasonable trade-offs and minimizing meta-commentary. Available in small (fewest tokens), medium (balanced) and full versions.

andrewyng/openworker

OpenWorker is an open source AI agent developed by Andrew Ng, founder of DeepLearning.AI and Co-Founder of Coursera. Instead of being a chatbot, it performs actions such as writing documents, replying on slack when mentioned (@OpenWorker), updating calendars, filtering emails, and more. It integrates with over 25 different tools like GitHub, Jira, Notion, Gmail, Google Calendar, and more. It uses MCP to communicate with any accessible tool. When performing any important action like sending messages or changing events, it will prompt for confirmation.

yc-duan/fastctx

FastCtx is a Rust MCP server that exposes popular repository actions (file io, code search, subprocess management) through a set of structured tools with clear input/output. Rather than dealing directly with shell commands, a code agent can make use of 9 different MCP tools. This open source project (dual MIT/Apache-2.0) has become popular due to the rising usage of MCP by AI agents to augment their functionality, as well as the reduced latency of a compiled language like Rust.

powerycy/goutoujunshi

Goutoujunshi is a open source Codex skill based on python which provides an AI relationship consultant. It has a core database of 19 knowledge document and 16 practical resource in psychology, communication tactic, attachment style, family law, and safety. It also processes users’ feelings and gives strategic advice, saves different relationship information in different profile, and outputs some specific conversation scripts. It openly recognize the diversity of sexual orientation and relationship type. The project is mainly written in Chinese.

pireel/pireel

Pireel Studio is a backendless, open-source AI video editor for talking-head content, built by pireel.com. The video editor (storyboarding, kinetic text, graphic themes, timeline, WebCodecs export) works in the browser without a server and without a user account. An MCP agent plugin (npx skills add pireel/pireel-agent) allows AI coding agents like Codex or Claude Code to control the video editor via natural language. The video bytes are stored in OPFS; nothing needs to leave the browser. Licensed AGPL-3.0.

Vincentwei1021/video-shotcraft

video-shotcraft is a skill for Claude Code & Codex that provides your agent with a suite of tools for creating videos. With 106 shot recipe cards, 161 motion previews, and a 36 second template (Ink Press), your agent can orchestrate all aspects of video creation — from capturing screenshots of your product to creating 2.5D camera animations, syncing cuts to music, and adding sound effects — all from one prompt.

thesysdev/appless

AppLess is an experimental app built by the team behind OpenUI, an open standard for generative user interfaces powered by language models. There are no apps, no home screens. Everything is dynamically generated by a language model (Cerebras) using openui-lang, a streaming DSL that renders to native Cupertino components on iOS and Material 3 on Android. It’s completely offline — you’ll need a free Cerebras API key.

Blaizzy/nativ

Nativ is a native SwiftUI macOS app developed by the author of mlx-vlm for running AI models on Apple silicon. Includes an embedded mlx-vlm server, scrounge up compatible models in your Hugging Face cache, and expose both OpenAI- and Anthropic-compatible local endpoints. With a chat UI, model library, performance dashboard, and tight integration with coding agents like Claude Code and Codex.

yuwen-cool/yuwen-publish-precheck

yuwen-publish-precheck is an open-source agent skill for Claude Code, Cursor and Codex that helps you precheck your drafts based on Chinese social media regulations (Douyin, Xiaohongshu, WeChat Video). It uses a 2-step process: first, a keyword matching method to identify potential violations, then, an AI semantic analysis method to judge whether a specific keyword violates the regulations in different contexts. For example, it may be okay to say something in your life story but should not be used in business presentations. Besides, we also provide suggestions for replacing those problematic keywords. Once you replace them with the suggested alternatives, our tool will automatically check whether the modified draft still contains any violations.

tandpfun/wardrobe

I fed a few outfits to the AI to see how well it could automatically detect clothes, and it worked pretty well! It could tell the difference between jackets, shirts, and pants even when there were multiple objects in frame. It leverages gpt-4.5-mini for vision and gpt-image-2 to make clean cutouts, and stores everything locally on your device. What really blew me away was the ability to model an editorial layout, where detected objects are laid out on top of a reference image. The drag-and-drop UI is powered by Node 22+ + Vite, so it only took me ~3 min to get up and running.

nethical6/conversation-steganography

At first I was like “lol ok” but after reading the README it looks like this is a tool for encoding a message using AES-SIV, then using a local LLM (Llama 3.2 or GPT-2) to map the encrypted bytes to tokens that make up a natural language conversation, thus hiding your message within a benign chat. This happens entirely client side, with key derivation using PBKDF2-HMAC-SHA-256 with 600,000 rounds. The wizard will automatically download the model (~500 MB to ~5 GB).

pablostanley/yoinks

I typed npx yoinks and 30 seconds later I was downloading a YouTube video. The TUI displays all available resolutions alongside their approximate file size so you can pick your poison before downloading. It’s backed by yt-dlp downloaded as a standalone binary, so no Python needed. The UI is written in Ink (React for the terminal) and the MP3 only mode with bundled ffmpeg-static is what makes this different than just calling yt-dlp yourself.

MatinSenPai/Aether-GUI

I checked out Aether-GUI and sure enough the one click connect button does exactly what it says. It fires off the Aether binary in a pseudo terminal, automatically responds to interactive prompts, and verifies the tunnel by connecting to the local SOCKS5 proxy on 127.0.0.1:1819. In the advanced tab you can select between MASQUE, WireGuard, and WARP-in-WARP tunneling protocols, along with setting the scan mode and IP version. Written using Tauri v2 + React 19 + Rust backend, much snappier than Electron apps.

hoainho/img2threejs

I read the README and immediately got excited about the pipeline architecture. Img2Threejs is a single reference image that generates a procedural Three.js GROUP object factory with pivots, sockets, and colliders for animating the object. This is rigorous because it has eight stages of building (blockout, structural, form refinement, material, surface, lighting, interaction, optimization) where the agent has to get approved by a vision model at each stage before proceeding. No external Python deps, 12 scripts, MIT licensed.

v-modal/vmodal_sdk_flutter

This is the official Flutter SDK from VModal, an AI-powered video search engine. This SDK allows mobile applications to integrate with the VModal gateway, providing users with the ability to search videos based on natural language queries, audio within videos, and even onscreen text (through multimodal semantic search). This repo was just created four days ago, but already has 621 stars. Apparently, there’s a lot of interest in searchable videos in mobile apps!

xiejunjie524/handdraw-story-video

Handdraw Story Video is a Codex skill and Python/Node.js pipeline developed by Chinese engineer xiejunjie524. It generates 35–45 second 720×960 videos at 30 fps from 7–9 story illustrations. Each frame starts off as a black & white line drawing before gradually filling in with low saturation color. Handdraw Story Video supports integration with AI image generators for source illustrations, and built-in checks for timing, duplicate assets, and caption length restrictions. Readme is in Chinese.

OlegSotnikov/sallyport

Sallyport is a free Mac app (Apple Silicon, macOS 14+) developed by Oleg Sotnikov that securely stores API/SSH keys locally and makes authorized requests on behalf of AI agents using MCP without exposing the key to the agent. Recent supply chain attacks and prompt injections have demonstrated that exposing credentials to an agent’s environment is becoming risky.

erickong/penecho

PenEcho is a free online drawing canvas where you can draw, write equations, or sketch diagrams and get AI responses right next to your drawing, which you can move around, accept, or reject. Works with Claude CLI, Codex CLI, OpenAI/Anthropic API. The canvas is a sparse 20,000x20,000 grid created by a tiny open source community. The project got lots of stars because it allows people to natively incorporate spacial, visual thought into their AI processes, which isn’t possible in a chat interface.

Wan-Video/Wan-Dancer

Wan-Dancer is a 14B diffusion model released by the Wan-AI team, who has also developed the Wan2.1 open video generation series related to the HumanAIGC research of Alibaba. By taking a music track and an image of a person as input, it outputs videos longer than 1 minute with 5 types of dance styles (Chinese traditional, K-POP, street, tap, latin). The core idea is to decouple the global keyframe generation and local temporal detail generation through a hierarchical framework, which addresses the problem of identity loss and movement repetition in the prior arts, thus enabling more than 20 seconds of video generation.

pyang5166/gbro-collage-broll

gbro-collage-broll is an open source agent skill(mostly Chinese docs) for AI agent environment. It’s a 3-step workflow. First, it gives you visual metaphors(free). Then, it generates static halftone paper collage frames for your approval(cheap). Last but not least, it renders a 9:16 5 second 720*1280 mp4 file with gemini-omni-flash-preview(paid). Halftone editorial style(bright solid background color, b/w photos, stop motion collage) is the branding of this skill.

PromptPartner/agentsmith

Agentsmith is a portable configuration tool for any AI agent that understands CLAUDE.md, AGENTS.md, GEMINI.md (e.g., Claude Code, Codex, Gemini). It helps you quickly get set up with a minimal config via a wizard using a universal base + a work type profile (e.g., software dev, devops, marketing, docs, data analysis, research, design, admin). It also ensures a plan -> implement -> verify -> ship process & a session hand off protocol so that sessions can continue seamlessly. We’ve been developing this for 6 months while working fully autonomously & have now open sourced it as a reusable tool kit.

vshulcz/deja-vu

deja-vu is a zero dependency go binary that parses session logs generated by various ai code agents (claude code, codex, opencode, aider, gemini cli, cursor, antigravity, grok build) and indexes them locally (no model downloads, no external apis, no leaving your computer).

TobiasLee/Rebuttal-Skill

The Rebuttal-Skill by TobiasLee is a Skill Definition (not an Application) which can be used with Claude Code, OpenCode, and Gemini CLI. It provides a two part process for handling academic rebuttals: Part 1 extracts review comments to atomic issues, determines reviewer intent, assigns severity levels (FATAL–MINOR), and generates a prioritized list of P0–P3 experiments/analysis plans; Part 2 generates structured rebuttals based on evidence after receiving author feedback in a Direct Answer + Evidence + Revision format. Part 0 provides a preliminary evaluation if it’s even worth writing a rebuttal and suggests a path for resubmitting if it’s not likely to be worth it.

stackblitz/bolt-slides

An open source React presentation framework developed by the StackBlitz team (creators of Bolt.new). Each slide is a web app component. Includes an agent skill (.bolt/skills/slides/SKILL.md) that teaches AI agents how to style, combine, and generate full decks. Slide library includes over 20 slide types such as graphs, code editors, timelines, & 3D cards.

Shadow-Weave/HMS

ShadowWeave’s Holographic Memory System (HMS) is a reproducible long term memory research framework designed for the LongMemEval benchmark, which requires the model to answer questions based on evidence from different previous sessions. It consists of an intermediate evidence ledger (a structured timeline of when, what and where) generated before generating answers, rather than just giving the model a list of retrieved evidence snippets. We also provide a wrapper (wrap_openai()) that can be used with any OpenAI API client, allowing you to automatically keep and look up your memories without modifying your code.

Blueturboguy07/cue

cue is a locally running electron app for Mac OS created by an independent developer. It’s an open-source alternative to the paid Cluely service. cue runs a transparent floating panel on top of your desktop and takes three streams of input: screen video, microphone, and meeting audio (the other party). These are then piped into your preferred AI model (OpenAI, Anthropic Claude, Google Gemini) which returns predictions in real-time. There is no backend server, data is only sent to your chosen AI provider.

Fei-Away/Codex-Dream-Skin

Codex Dream Skin is a non-official reskinning tool for the OpenAI’s Codex Desktop App, an AI code generator. This tool allows you to change the UI theme of the Codex Desktop App to your favorite image and colors using CDP (a browser debugging protocol), which changes the style of the UI dynamically without modifying the Codex program files and code signature. It also allows you to revert the changes back to default. First install the official Codex app and quit it once. Then simply download the .dmg file (for Mac) or Setup.exe file (for Windows) from here and install it (no need to clone the repo or install Node.js).

xai-org/grok-build

Grok Build is the terminal AI coding agent developed by SpaceXAI (xAI). It is an open-source tool licensed under Apache 2.0, implemented in Rust. Grok Build is a full screen interactive TUI that interacts with your code base, modifies files, executes shell commands and performs web searches. It operates in three modes: interactive mode, a headless mode suitable for scripting and continuous integration (CI) workflows, and an embedded mode using the Agent Client Protocol (ACP) within various code editors. The source code is periodically synchronized from the SpaceXAI monorepo and no external contributions will be considered.

William-Lu-stack/Flawless

Flawless is an AI native SRE control plane for Kubernetes and Cloud Infrastructure. It offers a Chat based Incident Console with Contextual Diagnostics, Automated Evidence Gathering, Dry Run Preview of Changes before making them and a 2D/3D Topology Map of your whole infrastructure. It can support multiple AI Models and maintains an Audit Trail of all Remediation Actions performed.

QuantumByteOSS/quantumbyte

QuantumByte generates an app from natural language input, creates a dedicated repo with live preview, and runs a separate requirements harness on it. It self verifies each business requirement (pass/fail/inconclusive), and will automatically start a repair process if any fail. It uses Claude Agent SDK under the hood, with Next.js for the UI, PostgreSQL + Redis for orchestration.

Intuition-Lab/personal-model

Persome logs your MacOS activities and builds a multi-layered personal model of you ranging from specific data points to an overall mental model of your thought process and priorities. This model is made available to AI agents using the Model Context Protocol (MCP) so they can take action based on your real behavior instead of making assumptions.

pengchujin/jzsub

This is a Codex Skill named JZSub. It automates the whole process of generating bilingual subs burned in mp4 video from any video source like Youtube, Bilibili, Vimeo, Twitch… all you need to do is provide the url of the video. It will automatically download the best quality file, transcribe the audio, translate using GPT, and finally output the mp4 video with bilingual subs burned in. All this in a single step!

modiqo/waggle

Waggle is a multi-agent AI artifact reference system. Rather than having each agent copy all the context from the last, which increases token count by ~15x, Waggle generates a short (~30 byte) token that any agent can lookup using MCP. Implemented in Rust, using SQLite for persistence, Ed25519 signatures, and Cloudflare Workers for scaling. Also includes support for tracking who has viewed what artifacts (receipts), versions, and revocations.

Kappaemme-git/codex-first-customer-finder-skill

This Codex skill takes in a startup idea/product URL, creates an ICP (Ideal Customer Profile), and uses public data sources like Reddit posts/threads, forums, etc., to find potential first customers for your startup. For each potential customer, it provides a link to where the data was found along with an evidence-based score. It also outputs an HTML file with all the data at the end.

jakubkrehel/skills

A suite of agent skills implemented as slash commands for Claude Code. These skills focus on different aspects of UI design like animations, typography, layout grids, and color schemes. They can be run from within an agent session using npx. This project includes a CLAUDE.md integration file, which makes it easy to use with Anthropic’s code agent.

withmarbleapp/os-taxonomy

I opened this expecting another AI toy and found a proper knowledge graph: 1,590 micro-topics mapped across 8 primary school subjects with 3,221 prerequisite dependencies, published by Marble (Generative Spark, Inc.). The licensing is unusually thorough — ODbL for the database layer, CC BY-SA for the authored content, with SHA-256 checksums and a CITATION.cff included. The interactive visualization at withmarble.com/curriculum is the clearest prerequisite map for K-8 learning I've seen, and the JSON schema is clean enough to slot directly into an RAG pipeline.

vinhhien112/Three.js-Object-Sculptor-Codex-Plugin

I tried this Codex plugin on a reference photo of a desk lamp and was genuinely surprised — it runs through seven build passes (blockout → structural → refinement → materials → surfaces → lighting → interaction) and outputs pure TypeScript/Three.js geometry with no GLBs or downloaded assets. The quality gate uses AI vision to score the rendered output against the reference image before advancing to the next pass. The README makes an explicit design decision I respect: it does not do photogrammetry — it reconstructs the soul of the object, not its mesh, which means the result is actually usable in games and interactive scenes.

mereyabdenbekuly-ctrl/clodex-ide

Clodex is a local-first Electron IDE that treats model output as untrusted input from the start — that design choice runs through the whole codebase. I set it up and the thing that immediately stood out is the Guardian system: sensitive operations like file writes and network calls require independent authorization rather than relying on the model's judgment. The agent runtime has proper turn management, cancellation, and recovery; memory is append-only with provenance tracking. They shipped the public preview with 3,322 tests passing.

op7418/guizang-material-illustration

This is a Claude Code skill that converts articles, data, and screenshots into labeled explanatory diagrams — text baked directly into the image, not a chart you have to style yourself afterward. I tested the chart beautification mode: feed it a screenshot of a messy spreadsheet and it reconstructs a clean material-design visualization with correct data. Supports 10+ diagram types including process flows, hierarchy maps, and scientific mechanisms. What impressed me most is the QA checklist in the README — it verifies label accuracy, data integrity, and readability at thumbnail scale before shipping.

xiaotianfotos/homerail

HomeRail converts one-off agent chats into reusable, auditable DAG workflows designed to run on a home server or a small datacenter node. I liked that the primary interface is voice-first — Chinese by default, with ASR/TTS/VAD all included — and the CLI is the fallback rather than the main path. The `hr dag supervise` command gives a real-time stream of DAG execution; `hr replay` lets you step backward through any run. Workers run in isolated Docker containers per DAG run, which means parallel workflows genuinely can't pollute each other's workspaces.

osama-fawad/Pekingman

Pekingman provides a complete cognitive pipeline for virtual NPCs to perceive the world via multiple modalities, remember past experiences, think like humans, keep emotions consistent, and react on the fly. It aims to enable lifelike virtual characters in open-world games, metaverse, and interactive digital humans.

AlephAITech/WorkBuddyGuide

WorkBuddy is an Office Assistant for daily assistance (e.g. HR, admin, sales, engineer) that can complete tasks such as reading files locally, generating reports, spreadsheets and presentations as well as knowledge management. WorkBuddy can access multiple models(Hunyuan/DeepSeek/GLM/Kimi/MiniMax), MCP Server & Skills. Official document repo, but actually it’s a bluebook written by community. Introduce how to install it and do the first task, then give some examples, multi-agents, RB Roadmap etc. Mostly simplified chinese (only README supports bilingual). License: MIT.

yetone/kill-ai-slop

The Kill AI Slop project is a list of 32 visual and copy anti-patterns that tell the world your product was generated by AI. These include indigo gradients, glowing cards, ‘Powered by AI’ labels, and excessive emojis. In addition to this field guide, there’s also an agent skill that will scan the code of any web project and automatically remove these anti-patterns. Available in 4 languages, the project itself serves as a model of good design by being intentionally sparse.

Raymondhou0917/speak-human-tw

An agent skill for Traditional Chinese that detects 35+ AI writing patterns — overly formal structures, filler phrases, mechanical bullet lists — and rewrites them to sound naturally human. It runs inside Claude Code, Codex, or Cursor with a two-phase workflow: first it annotates every pattern it found and suggests changes, then asks for approval before touching a single word.

cosmtrek/mindwalk

What is Mindwalk? Mindwalk is a visualization tool that turns the agent’s action log into a 3D interactive map of your project. It shows files with different colours based on how the agent has interacted with them: if they have been read/searched/edited. Also, you can time travel by scrubbing through the timeline to see exactly what the agent did step-by-step, making it much easier to debug your project. Claude Code Codex

CoscoAI/luxy-aisre

Luxy is an AI-based Site Reliability Engineering (SRE) platform for Kubernetes. It unifies alert, log, topology and runbook data into one automated workflow — discovery, diagnosis, preview, validation, execution, verification. It works with any OpenAI compatible endpoint, can be installed using helm, and has a chat interface for making operational decisions. Luxy received more than 300 stars since launch.

yusukebe/ax

ax is a CLI tool that acts like ‘curl for the AI era’. Fetch webpages and get back structured, token efficient content that’s tailored for LLM agents instead of just plain old HTML. Find outlines, pinpoint text locations, pull out fields, parse tables, all in one go. Written in Bun and just one runtime dependency. Runs entirely locally, no need for any cloud APIs or API keys.

WeZZard/jlens-qwen36

jlens-qwen36 is a visual debugger for running Qwen3.6-27B (4-bit quantized) locally on Apple Silicon using MLX and visualize the model’s “thought process” across layers. By leveraging the Jacobian lens (inspired by Anthropic interpretability work), we can visualize the tokens that the model is biased towards at different depths and positions. This is represented as a clickable matrix. This project gained over 300 stars within a few days when folks realized they could now introspect into cutting edge models on their own device!

ronak-create/FableCut

A non-linear web video editor powered by a single Node.js server with no npm dependencies. The unique part about this project is that the whole project timeline is stored in a JSON file, allowing any process (including AI agents) to access and edit it. With an included MCP server, Claude Code and Claude Desktop can manipulate the project timeline directly. A REST API allows other systems to do the same. Changes made by AI agents are pushed to the browser client using Server-Sent Events and revision numbers ensure safe multi-writing.

Doriandarko/texts-to-transformer

This is a Python script that trains a decoder-only Transformer model using Apple’s MLX framework, on your own iMessage data. Snapshotting the Messages DB, anonymizing contact info, creating a custom byte-level BPE tokenizer, leaking-free dataset splitting, training loop, and serving your new chatbot in the command line. All done locally, no cloud needed!

Nanako0129/pilotfish

Pilotfish is a one-prompt-install orchestration layer for Claude Code that routes each task to the right Claude model automatically: Fable 5 handles planning and architecture decisions, while Opus, Sonnet, or Haiku carry out the actual implementation. The idea is that you almost never need the most expensive model for the entire session — only at the thinking stage.

yynxxxxx/Codex-X

Codex-X is a multi-platform desktop application that provides a GUI frontend to OpenAI Codex. You can easily change AI providers, inject system prompts, and configure your authentication visually rather than through environment variables or configuration files. Codex-X also supports TOML configuration files and allows freeform instruction templates for complex prompt engineering. Within its first week, Codex-X received 642 stars because it solves a very real problem: while Codex’s API is great, its ecosystem is lacking.

SmileLikeYe/agent-chief

Agent Chief is a local-first attention filtering layer for managing multiple AI agents. Using a three-tiered “worthiness” system of hard coded rules, similarity classifier, and LLM judge, Agent Chief will either interrupt you, forward the request to another agent, or archive it silently. In the first week, this project received 319 stars from users working with multiple AI agents at once.

Robbyant/lingbot-world-v2

AI world modeling engine using Wan 2.2 to generate 720p 60fps video from text descriptions. The world includes two cooperating agents: a pilot agent that handles character actions and a director agent that drives dynamic environmental effects. This project demonstrates an ‘unbounded interaction horizon’, meaning long videos with consistent scene information over long periods of time.

Robbyant/lingbot-vla-v2

LingBot-VLA 2.0 is a Vision-Language-Action foundation model for robot control, pre-trained on 60,000 hours of data across 20 robot configurations plus egocentric human videos. It represents actions in a unified 55 dimensional space and leverages sparse MoE layers to enable a single model to generalize well to robots with vastly different embodiments, something that most robotic foundation models fail to do. This model has only received 311 stars in less than 48 hours, which highlights how rare large scale, open source cross-embodiment robot foundation models are.

Robbyant/lingbot-video

LingBot-Video is an open-source Mixture-of-Experts video generation model trained on over 70,000 hours of embodied intelligence data for text-to-video and image-to-video tasks. The training applies multi-reward optimization across aesthetics, physical consistency, and task completion — an unusual combination for video models of this scale. It gained 360 stars in under two days because truly open MoE video models remain rare.

oso95/scroll-world

A Claude Code skill that builds immersive scroll-driven landing pages. It interviews you about your brand, then uses Higgsfield (GPT Image 2 for isometric diorama scenes, Seedance for camera-flight videos) to generate connected visual sequences. A custom scrub engine synchronizes video playback to scroll position so the camera flies through each scene without cuts or jumps.

nexu-io/motion-anything

Motion Anything is an agentic motion engine that lets you animate anything based on natural language inputs (no keyframes, no timelines). You don’t need any design tools! It comes preloaded with 403 carefully curated motion recipes and can export to CSS, React, Lottie, MP4, and GIF all at once via one prompt. This zero npm dependency tool was a huge success with vibe coders looking for easy ways to achieve high quality UI motions.

mira-wm/mira

MIRA is a 5 billion parameter latent diffusion world model which generates Rocket League frame by frame from actions. This is a collaborative research project between General Intuition, Kyutai and Epic Games. It is one of the first playable world models available for public use. It achieved 312 stars of genuine excitement of the coming world model era.

Tencent-Hunyuan/Hy3

Hy3 is a 295-billion parameter mixture-of-experts language model from Tencent, with only 21 billion parameters active per inference call. It is designed for reasoning, agent-based pipelines, and long-context tasks, and ships with deployment guides, fine-tuning docs, and quantization tooling. The team reports strong performance in coding, frontend development, and CI/CD automation, with lower hallucination rates and more stable tool-call behavior compared to prior Hunyuan releases.

SuperJJ007/CSSwitch

CSSwitch is a small macOS application written using Tauri 2.0. This application serves as a local proxy between Claude Science and any Anthropic/OpenAI compatible API server. The application runs Claude Science within its own sandboxed environment and dynamically modifies the API requests to allow you to replace Anthropic’s backend with DeepSeek, Qwen, Kimi, MiniMax, GLM, OpenRouter, or your own self-hosted endpoint!

simonlin1212/Vibe-Research

Vibe-Research is an Open Source Investment Research Platform which streams Real Market Data (Valuations, Fundings, Sector News in 12 Categories) into any LLM of your Choice. Claude Code Subscription, DeepSeek/OpenAI API’s & MCP Integrations are supported so that your AI can do the Analysis based on Structured Local Data rather than sending it to some Third Party Servers.

Robbyant/lingbot-vision

LingBot-Vision provides pre-trained Vision Transformer encoders that can learn spatial information from unlabelled images by predicting object boundaries. Available models include ViT-S to a 1.1B Large model, and weights are available via Hugging Face and ModelScope.

Pluviobyte/rnskill

rnskill is a collection of about 60 skills — ready-made task routines — for AI agents like Claude Code and Codex. There's one that rewrites text to remove the "AI feel" while keeping the author's voice, one that plans cinematic video concepts, one that builds dark-style product video scenes, one that makes typing-animation intros with synced sound, plus others for subtitles, covers, and video editing. To install, tell your agent to run `npx -y skills add Pluviobyte/rnskill --skill <skill-name>` (pick the skill you want from its list), or add the whole set with `-g --all`.

isjiamu/gzh-design-skill

This is a Claude Code agent skill that transforms a Markdown document into HTML ready for publishing on WeChat’s Official Accounts editor. It comes with 6 beautifully crafted themes ranging from sparse white space to elaborate editorial designs. There’s also a theme generator that allows you to create your own themes based on a simple textual description or an image example. And to ensure your content looks great on WeChat, there are two validation points to catch common rendering issues even before you start typing.

anthropics/jacobian-lens

The Jacobian Lens, by Anthropic, is a novel interpretability method that takes any internal activation in a language model and maps it linearly to the output layer. It then decodes it to a list of top-ranked vocabulary words. This provides a clear view of what the model was “thinking” about at any point in the model. In their paper, they propose that language models have a global workspace of verbalized internal representations. This is a direct view into their global workspace. You can use it on any HuggingFace decoder transformer model out of the box.

ai4s-research/open-science

Open Science is a Tauri desktop app that organizes AI-driven research into a planning, approval, execution, and review process. It comes with built-in agent capabilities for literature reviews, experimentation, and paper writing. All artifacts (figures, tables, reports) are fully traceable back to the underlying code, data, and conversations. It supports connecting to any LLM via OpenRouter or directly via the Anthropic API, or using a local model. MCP is used for dynamic integration of tools.

xt4d/GameBlocks

GameBlocks is a JavaScript skill library that plugs into Claude Code or Codex and provides self-explanatory building blocks for browser-based 3D games: spatial transformations, actor motion, world state, and coordinate frames. It was built because natural language descriptions of 3D behavior are inherently ambiguous — agents misinterpret direction, produce unstable motion, and break game state. With clear, inspectable module semantics, the agent can reliably generalize from the blocks rather than deriving 3D logic from scratch every time.

elder-plinius/T3MP3ST

This runs locally, and within an hour was already performing reconnaissance (nmap, dns fingerprinting, http probes) via my existing Claude Code agent, without needing any additional API keys. Key-less is the magic here; T3MP3ST doesn’t come with its own LLM, but hooks into your existing coding agents. Check out the impressive and accurate results in the README: 90.1% pass@1 on 104 black-box tasks, 58% hintless CTF solution rate with one run, 8/10 exact file/line/CWE match on held out CVEs. High risk post exploitation features are gated by explicit approval. That’s good.

HUANGCHIHHUNGLeo/claude-real-video

I pointed this at a 20 minute talk, and it pulled 47 keyframes based on scene changes, and passed those along with the whisper transcription to Claude. It produced a summary that followed the flow of the speaker’s argument, not just the slides themselves. Deduplication is done by downscaling and calculating pixel differences over a sliding window. Maximum 150 frames by default, scene change threshold 0.30.

jamesob/local-llm

Finally, a hardware guide! James O’Beirne includes actual configs and price estimates for hardware: $5,587 base system with used EPYC Milan CPUs, up to 4x RTX 6000 Pro, running GLM-5.2-594B at 80 tokens/second with a 460k token context window. The P2P GPU bandwidth is what really blew my mind: 50.4 GB/s bi-directional bandwidth with 0.37–0.45 microsecond latency over the PCIe switch. Docker images of the models are provided and ready to go.

ammaarreshi/Generals-Mac-iOS-iPad

At first I thought this was some kind of emulator trick to get a 2003 RTS game running on iPad. But after reading the README, it turns out this is using EA’s official GPL v3 source code release, compiling the real C++ code base for arm64, with DXVK/MoltenVK doing the directx->metal translation in the driver. The mention of being built with Claude Code as the main coding agent, with architectural guidance from humans, really caught my eye. Only 2,200 lines of code were added to GeneralsX to make it fully compatible with iOS, including campaign, skirmish, and Generals Challenge modes.

Archive228/loopkit

One liner install, 33 pre-built skills for Claude Code spread across 9 categories (debugging, security, front end dev, testing, refactoring, etc). First time I tried it, my adversarial verifier subagent pointed out that I had written a test that passed but wasn’t actually checking the correct invariant. Standing context document is only 60 lines, so very little token cost per request. Compatible with Claude Code, Cursor, Codex, and Gemini CLI.

synthetic-sciences/openscience

OpenScience is a self-hosted, open source AI toolbox for fully automated scientific research. Simply specify your research objective and OpenScience will autonomously conduct a literature search (across more than 30 databases like arXiv, PubChem, UniProt) to come up with hypotheses, write & execute code, and generate reports. It’s written in TypeScript & Python using Bun and comes with more than 290 prebuilt skills in fields such as biology, physics, and machine learning. This project exploded onto the scene in just a few days, getting from 0 to 598 stars, by truly automating AI-enabled research.

nagisanzenin/engram

Engram is a plugin that introduces spaced repetition (brief review intervals) to Claude Code and other coding agents. When you tell it to teach you something, it will prompt you to attempt to remember the information, after which another “assessor” agent will test your retention, and the FSRS-4.5 (spaced repetition scheduling algorithm) will determine when to schedule your next brief review session, just before you would have forgotten. Install the Engram plugin for Claude Code by running `claude plugin marketplace add nagisanzenin/engram` followed by `claude plugin install engram@engram`. Then you can begin learning new topics using the command `/learn <topic>`.

MaximeRivest/riddle

Riddle is an AI-powered journaling application for the reMarkable Paper Pro tablet. When you pause writing, it captures the page as an image and sends it to a vision language model. The model's reply is converted into handwriting using the Dancing Script font and animated stroke by stroke directly on the e-ink display. It supports OpenAI, OpenRouter, Groq, and local LLM servers, with two rendering modes — one inside existing tablet apps, one that takes over the e-ink hardware entirely. 322 stars in two days from the Rust and LLM communities.

HKUDS/OpenOPC

OpenOPC is a Python framework to build AI-driven organizations where every agent has a state machine for work items (execute, delegate, review, integrate, rework). Agents can hire specialized sub-agents as needed and accumulate institutional knowledge over time. Supports 9 verticals, integrates with Claude Code, Codex, Cursor, OpenCode as execution engines.

Kulaxyz/token-diet

Token Diet is an always-on efficiency skill that injects token-saving instructions into every session of your AI coding agent. It tunes the agent's reply style, documentation output, test verbosity, and context handling across Claude Code, Codex, Cursor, Windsurf, and Cline. Real Sonnet 5 benchmarks show an average −31% bill reduction, with output-heavy sessions dropping as much as 81% in token output.

uzairansaruzi/hermex

Hermex is an open-source SwiftUI app that pairs with a self-hosted Hermes backend to give you a full-featured mobile control panel for your personal AI agent. It supports streaming chat with file and image attachments, session history, cron-based task scheduling, skill browsing, and server-side file navigation — making your phone the remote interface for an AI running on your own machine. The project picked up over 360 stars in a single day, riding a wave of interest in privacy-first AI setups where people want mobile access without routing everything through a third-party cloud.

jmerelnyc/Talos

A free and open source GPU worker client for the Talos distributed inference network. This works alongside a local Ollama installation and uses WebSockets to connect to the network, accepting language model inference requests from other users and reporting uptime for payouts. Also comes with a dashboard to view GPU usage and tune resource allocations.

JasonLiu0826/ackem

Ackem (Autonomous Companion Keeping Emotional Memory) is an Electron + TypeScript based local first desktop application to host your own AI companion locally. The AI companion will save all the past conversations, emotions, personality information in your local directory only. You can use any OpenAI compatible api like Ollama, DeepSeek, Lm Studio etc. Ackem also provides STT/TTS functionality using your voice, live2d avatar, ability to load memory from text files, and gaming mode.

wlzh/dji-4g-vohive-mac

This article is an in depth tutorial about running a Linux virtual machine on Mac via UTM, then changing the USB ID of the DJI EG25-G modem from its own ID to the universal Quectel EC25 ID using AT commands so that VoHive and other Linux modem drivers can work out of the box. Supports both Apple Silicon and Intel platforms.

mekos2772/ios-location-spoofer

This is a JS script that will hijack Apple’s geolocation requests (WiFi & Cell tower) and respond with arbitrary coordinates. This script works within various iOS proxy clients (Shadowrocket, Surge, Loon, Quantumult X, Stash) without requiring any modifications to your device or apps.

Jia-Ethan/codex-keysmith

This is a simple single script python CLI that handles the process of adding a markdown instructions file to codex cli’s config directory for you. It copies the file to the .codex directory, adds model_instructions_file to the config.toml file and creates a time stamped back up of the original file.

ik190/macos-chatgpt-overlay-bar

This is a simple native macOS application that integrates ChatGPT into the system’s menu bar. With just one click, the chat window will slide down. It became popular due to the fact that it solves the biggest pain point of many daily ChatGPT users on macOS, which is context switching between browser tabs.

cclank/lanshu-animated-architecture-diagram

This is a Python tool that reads a JSON file (a plain-text data format) describing a system's architecture and produces three outputs: an editable diagram file (Excalidraw format), a static PNG image, and an animated GIF — 1210×1138 pixels, 20 frames per second, 41 frames — where glowing dots move along the connection lines and blocks pulse. Everything renders locally through Pillow (a Python image library), no browser or design app involved. To install: clone the repo, run `pip install -r requirements.txt`, then run `python3 scripts/render_animated_diagram.py --spec assets/default-spec.json --outdir outputs --verify` — works on any regular computer, no special hardware needed.

ricccrd/dd

dd is a container runtime for macOS that removes the virtual machine. Rather than running a complete Linux VM like Docker Desktop, it utilizes a JIT compiler to translate container code and performs Linux syscalls in userspace. Supports both ARM64 and x86-64 containers, provides a Docker Engine compatible API allowing you to use regular docker commands, includes a GTK4 desktop application and CLI. Developed by one person (Richard Hutta) with 35 releases and an MIT License.

Pluviobyte/video-production-skills

The repository packages four reusable video production skills: ai-motion-director for topic-to-motion planning, reference-video-replica-qc for recreating reference clips, dark-saas-magic-video for dark-themed SaaS product videos, and black-white-text-opener for typed-text intro sequences. Skills install via npx and integrate with Claude Code, Cursor, and GitHub Copilot.

offchainthoughts/Amber

Embedding corpora can be large and time consuming to create, but for many applications they are embedded repeatedly. This python library allows you to serialize the embeddings into a file (.amber) that contains cryptographic proofs of the integrity of the embedding. This means that users can verify your embedding via spot checks (i.e., re-embedding some portion of the corpus and checking that the results match). This also means that you can distribute your embedding file without the need to share your model. The repository contains formal proofs of correctness.

lycorp-jp/sim-use

observe transforms UI contents into a compact representation that’s more efficient in terms of tokens (16x less than accessibility json) and act performs actions such as tapping, swiping, and typing on elements by their aliases rather than pixels. It takes ~300ms to complete a round trip. It comes with a Claude skill and works with iOS Simulator and Android emulator/device. It was published at v0.9.0 in June 2026.

Leron-X/leronx

LeronX Engine is an open-sourced video production pipeline for LeronX Pro. It includes all steps from writing scripts and storyboards, planning scenes, TTS synthesis (11 languages), FFmpeg rendering with effects, subtitles, stock video matching, plugins, etc. Comes with 47 tests (97% coverage) and Docker. The proprietary parts like cloud rendering, billing, desktop app are not available here.

Kulaxyz/self-learning-skills

Self-learning skills is a skill for AI coding agents (like Claude Code, Cursor, or standards-based agents) that learns from their successes in a session, such as how to access a database, where to find credentials, or what deploy command to use, and turns those into reusable skills for future sessions. It has built-in safety measures to avoid storing secrets, and a confirmation step before knowledge is committed.

Krishnagangwal/CS-Fundamentals

This repository consists of various PDFs and markdown notes related to important computer science subjects which are required for any technical interview. It contains notes on Data Structures and Algorithms, DBMS and SQL, OOP, Operating Systems, Computer Networks, and System Design. Also includes Leetcode Problem sets, HR Interview Preparation and Companies Hiring for Entry Level.

Einsia/Browser-BC

Browser-BC is a browser extension that captures free form browser activity, which is then distilled by a local process to generate a SKILL.md file documenting how to perform the task using an AI agent. This skill can be installed on either Claude Code or Claude Desktop and run using the AI agent. It’s written entirely in python and runs locally (the only time it accesses the internet is to make a call to the LLM distillation service). V0.1.9 was released in June 2026.

diinki/linux-antiquity

Linux Antiquity is a linux desktop theme for Hyprland tiling window manager. It is inspired by Art Nouveau style and historical illustrations about astronomy & mythology. Theme includes theme picker, weather applets and complete config for Quickshell compositor. This is version 0.1 and mostly art project. There are a lot of refactoring to do.

CopilotKit/OpenTag

OpenTag is a self hosted AI bot for Slack that can read messages, answer questions, run tools and display outputs (e.g. charts, breakdowns) inline in threads. It’s built using the CopilotKit bot SDK and also works on Discord, Telegram and WhatsApp. Features include: No per seat licenses, self hosted, require human confirmation before running actions.

winsznx/theeleven

Theeleven runs 11 AI characters — each one plays a football tactical role like playmaker, striker, or defender — that watch a live match in real time and automatically open "over/under" betting markets on the X Layer blockchain (made by OKX). The technical core is PropMarketHook, a custom add-on (called a "hook") for Uniswap v4, a decentralized-exchange protocol, with a commit-reveal mechanism — agents lock in a hidden choice first and reveal it later — built directly into the blockchain. Users stake USDT0 (a stablecoin) with one wallet signature, and a relayer (a middleman service) pays the transaction fee for them; the project also exports MCP tools (a standard that lets other AI agents plug in) so other agents can use the same markets. To run it yourself: clone the repo, run `pnpm install`, then `pnpm --filter @regista11/web dev` to start the web app locally — you'll need Node.js and pnpm installed, plus several API keys and a wallet private key to make it actually work.

deepseek-ai/DeepSpec

DeepSeek's full-stack open-source toolkit for training and evaluating speculative decoding draft models. Ships three algorithms (DSpark, DFlash, Eagle3) with pre-trained checkpoints for Qwen3 and Gemma, multi-GPU training scripts, and a benchmark suite covering nine evaluation datasets.

bozhouDev/codex-orange-book

An unofficial community guide to OpenAI Codex — the 'Orange Book' — covering everything from installation to real-world case studies. Produced by Chinese developers and available as a downloadable PDF, it fills gaps the official docs leave open.

benchflow-ai/awesome-evals

A seriously curated library of 443+ annotated resources for building and evaluating AI agents: papers, blogs, talks, tools, and benchmarks. Assembled via a depth-4 citation crawl of 11,600+ papers by BenchFlow, with 146 reading notes and active pruning of dead links.

AlexandrosGounis/pdfx

A minimal extension of the PDF format that bundles multiple documents into one file via an embedded JSON manifest. A .pdfx file is still a valid, standard PDF — any reader opens it — while the PDFx viewer can split it back into individual documents on demand.

amplifthq/opentag

OpenTag lets your team @-mention a coding agent directly in Slack, GitHub, or Lark — the message is routed to Codex or Claude Code running locally on your machine, and the response comes back to the same channel. No cloud service in the middle.

tdeverx/contained-app

Contained is a native SwiftUI macOS application to manage Apple’s container CLI. It provides a glass-card view of containers, realtime CPU/memory/network sparklines, a terminal, logs viewer, file explorer, and import Docker Compose files. It’s made for Apple silicon Macs.

trotsky1997/OpenFugu

OpenFugu is an open reimplementation of Sakana AI's proprietary Fugu system — an orchestrator that uses a tiny 0.6B routing model to analyze each query and direct it to the most capable worker in a pool of LLMs. The measured result is 107% of the best single worker, meaning multi-model routing actually outperforms sticking with your single strongest model. The project covers all four stages: reading the architecture, running demos, training a custom router, and serving via API.

Th0rgal/open_oura

open_oura is a Rust toolkit that communicates directly with Oura Ring (Gen 3/4/5) over Bluetooth Low Energy using a reverse-engineered protocol. It retrieves heart rate, HRV, sleep stages, temperature, and activity data and stores everything in a local SQLite database. A real-time 3D motion visualizer is included as well. The project grew out of frustration with Oura's cloud-only data access model.

samasante/liquid-glass

This is a zero dependency React component implementing Apple’s famous glass morphism effect. Instead of just displaying a static screenshot, it acts as a dynamic lens over real DOM elements. Chromatic aberrations and refractions are achieved through SVG displacement filters and signed distance field rasterization while maintaining full interactivity with all underlying elements. Supports Chrome, Firefox, Safari (including iOS) and has 3 different modes: refract child elements, refract content using props, render with WebGL (for videos and canvases).

QwenLM/Qwen-AgentWorld

Qwen-AgentWorld is a Mixture-of-Experts language model with 35 billion parameters (only 3 billion active at a given time) and is trained to mimic the working environments of AI agents including web browsers, command line interfaces, operating systems, mobile applications, search engines, software development tools, etc. The model was developed by Alibaba’s Qwen team based on more than 10 million real-world interaction trajectories via a three-stage training process including continued pre-training, supervised fine-tuning, and reinforcement learning. The model has a 256K token context size and supports various inference engines like SGLang, vLLM, and Transformers, which make it superior to most of its proprietary counterparts on AgentWorldBench.

m1ckc3s/claude-status-bar

This is a simple macOS menu bar application written in Swift that displays a live Claude Code session status. The indicator will animate when Claude is busy, show the time elapsed since starting, and indicate when Claude is idling and ready for input.

krea-ai/krea-2

The official inference code and open weights for Krea 2, a scratch-trained text-to-image model developed by Krea AI. We provide two checkpoints: Raw (full-size base model for fine-tuning and LoRA training) and Turbo (distilled model that generates images in only 8 steps). According to Artificial Analysis, Krea 2 is the best text-to-image model as of June 2026.

BohemiaInteractive/CWR

The game engine for Bohemia Interactive’s Arma: Cold War Assault has been ported to C++20 using CMake and Clang, and is available for Windows x64 and Linux x64. The game’s assets are licensed under the Arma Public License while the engine itself is licensed under the GPLv3.

ksimback/looper

Looper is a Claude Code skill that requires a design phase prior to running an agent loop. It asks questions regarding your objective and what constitutes success, then produces a portable YAML specification and a Python runner. This runner includes cross-model review judges, iteration limits, no progress detectors, budget constraints, and an ASCII flow chart to visualize the process before execution.

kanavtwtgg/birds.cafe

A WebGL simulator that lets you control a formation of seagulls across the ocean using your keyboard arrows. The simulation includes day/night cycle, rain, lightning, and storm. There’s also some ambient music for the atmosphere. The best part is that it works on mobile too! This was published on June 21st and within a few days it got 731 stars without any marketing whatsoever, only by word of mouth from people sharing it with others.

HKUDS/AgentSpace

AgentSpace turns AI agents into proper team members — each with a defined role, a skill set, and access rights. It ships an AgentRouter layer that works with most major AI providers, plus built-in task scheduling, approval flows, and audit logs. The governance layer most agent frameworks skip is exactly what this project puts front and centre. That gap is probably why it hit 343 stars within two days of launch.

yo-WASSUP/Good-Badminton

Good-Badminton is an open source AI Hawk-Eye for badminton. Using YOLO model to identify shuttlecock and players, RTMPose to generate skeleton keypoints, finally get annotated videos, players’ heatmaps and movement stats (distance & speed each rally). Within few days after publishing, it already got almost 400 stars which is quite unusual for a niche tool like this one.

sums001/Windows-Copilot-API

This is a python library and fastapi server that makes Microsoft Copilot’s web interface work as an OpenAI compatible rest api. This allows you to use GPT-4/5 without an OpenAI key or paying a penny. The library includes support for cloudflare captchas, rate limiting, and docker! With 408 stars in just 4 days it’s clear that there’s a lot of demand for free model access.

SakanaAI/fugu

Fugu lets you dynamically distribute tasks amongst several large language models, each optimized for different subtasks, and aggregate their responses into a final answer. It uses two machine learned models (TRINITY, an LLM orchestrator, and Conductor, a reinforcement learning based agent-to-agent orchestrator) to manage the coordination between agents. Available via the Sakana API.

raiyanyahya/recall

Recall is a Claude Code plugin designed to overcome the cold start issue. When starting a new session, there is always a loss of context since you have to explain your project all over again. This plugin saves your session activity, and generates a summary of the context which is saved as two files (running history and regenerated context) in a .recall/ directory within your repository. The summarization happens completely offline using TF-IDF and TextRank algorithms in python. There are no API requests, no third party services and no tokens used!

overflowy/make-look-scanned

A Go CLI (or WebAssembly in your browser) that applies a set of degrading filters to each page of a PDF file: small rotation, grayscale, warm paper color, scanner noise, shadows on the edges, and jpeg compression. The output PDF looks just like the physical document you would have printed and then scanned. The CLI can also be compiled to WebAssembly, allowing you to perform the same transformations directly in your browser without having to install any software.

lyra81604/zhengxi-views

This Agent Skill summarizes 14 years of investment notes made publicly available by Zheng Xi, a popular fund manager at EasyFund, into a fully trackable AI system. Each answer contains references to supporting quotes from the original text without making up any information. It also includes actual performance figures of 8 of his funds and over 27,000 market funds, which are rated according to Zheng’s own criteria.

vercel/eve

Eve is a new open source framework by Vercel for creating robust AI agents. What sets Eve apart is its novel file system approach — agents are defined via files and directories, making them easily editable, understandable, and deployable. In just a few days, Eve has gained over 2,100 stars!

zhongerxin/Cowart

With Cowart, you can extend the Codex AI assistant with an infinite drawing canvas (powered by tldraw). This lets you draw out your thoughts, insert images, and iteratively modify them with help from the AI. All of this happens locally, without relying on external servers. In just a few days, Cowart has gained nearly 1,700 stars!

rebel0789/codexpro

With CodexPro, you can point ChatGPT’s Developer Mode to your code base and turn it into a hands-on coding agent. This works by establishing a secure tunnel through which ChatGPT can access your files, search code, and execute safe commands — all within the scope of your chosen project. In just a week, CodexPro has gained over 650 stars!

Forsy-AI/agent-apprenticeship

Agent Apprenticeship is a framework for teaching AI agents through experience. By capturing the step-by-step logs of tasks performed by agents and transforming them into reusable experiences, Agent Apprenticeship enables users to share their knowledge. It includes a CLI and comes packaged with a library of over 500+ real world tasks, as well as support for tools such as Cursor and Claude Code. In its first week, Agent Apprenticeship has gained over 600 stars!

Plaer1/junction

Junction is a VS Code sidebar that integrates your code editor with local AI coding agents. This lets you interact with these agents directly within your workspace, using a chat interface. Junction stands out because it provides a single panel for interacting with multiple types of agent backends, giving you flexibility and avoiding vendor lock-in. In just a week, Junction has gained over 500 stars!

ngrok/webernetes

This project is a TypeScript port of a subset of Kubernetes to run in your web browser, without any backend needed. It supports Pods, Deployments, ReplicaSets, ClusterIP & NodePort Services, rolling updates, and even HTTP communications between pods via DNS. This was created by the ngrok team to provide an interactive way to teach Kubernetes without having to maintain a cluster.

cloudflare/security-audit-skill

The Cloudflare coding agent skill automates a 6 stage security assessment process (reconnaissance, vuln scanning, validation, reporting, structure output, and independent verification) with each finding being validated by another agent to minimize false positives.

anthropics/launch-your-agent

The Launch Your Agent skill is an officially sanctioned Claude Code Skill made by Anthropic that helps founders and builders go from idea to a Claude Managed Agent in their own Anthropic account. After installing this Skill and running /launch-your-agent, it will ask you a series of questions to scope out what you need your agent to do and create a v0 of your agent. Once the v0 is created, it will deploy it to your Anthropic account using your API key. The Skill will then evaluate the performance of your agent based on your criteria, provide you with recommendations for improvement, and optionally schedule your agent if necessary. In the end, you’ll have a functioning managed agent in your Anthropic account along with a my-agent/ directory which contains a resumable launch script, evaluation scaffold, and a notes file for future versions (v1/v2). The skill is open source under the Apache-2.0 license and has gained nearly 400 stars in its first week.

aidenybai/cnfast

cnfast is a TypeScript library that provides a drop-in replacement for the commonly used cn utility function in Tailwind CSS applications. By leveraging call-site identity caching and V8 argument caching, cnfast is able to achieve an average speedup of 3.8x and a maximum speedup of 7x compared to tailwind-merge while maintaining identical byte output. Developed by Aiden Bai, creator of the popular million.js virtual dom optimizer.

boogu-project/Boogu-Image

Boogu-Image is an open source, Apache-2.0 family of image generation and image editing models implemented in PyTorch. There are three versions available: Base(with emphasis on diversity and legible text), Turbo(4 step process to generate photorealistic images) and Edit(for image to image transformation). Supports both English and Chinese texts. ‘Boogu Arena’ was released to showcase its ELO level which is comparable to top commercial products. Main selling point is that it has almost closed source quality with much smaller training data. It gained almost 400 stars in its first week.

dongshuyan/compass-skills

COMPASS is a local-first bundle of four skills for AI coding agents like Claude Code and Codex, each aimed at the moment things go sideways on a long project. A task-clarifier gates ambiguous or risky work and asks a few sharp questions before acting; a task-forest keeps a repo-local map of goals, subtasks and dependencies so the AI always knows what it's actually building; a session-handoff skill compresses the whole conversation state into a paste-ready prompt for a new session; and a profile-keeper remembers your communication style and risk boundaries. It all lives in your repo as plain files — task graphs you can even view as a diagram — and nothing is uploaded anywhere.

privatenumber/mac-ocr

mac-ocr is a command line utility to extract text from images and pdf files using Apple’s native Vision framework. The same technology that powers the OCR capabilities of macOS. It allows for local processing without requiring any internet connection. Additionally, it supports converting a scanned pdf into a searchable pdf with a selectable text layer. It’s available as a CLI, a NodeJS library, and even comes with an agent skill for AI assistants.

fivetaku/fablize

Fablize is a Claude Code plugin designed to promote a process over increasing the AI model’s intelligence. This helps ensure the model runs and checks its own HTML/code/graphs, finishes tasks, replicates bugs, and never claims a complicated task is finished without evidence. The plugin was inspired by a controlled experiment pitting two cutting-edge models against each other, revealing that the main distinction between them was their processes.

eli-labz/Third-Eye

Third Eye is an open source geospatial dashboard made with Next.js and MapLibre. It aggregates live public data feeds such as aviation, shipping, wildfires, earthquakes, weather, satellite imagery, news feeds and more into a high performance GPU accelerated map. It uses free public data sources and requires no API keys to use. Includes additional tools such as a recon suite and crypto wallet tracking. Pre-built docker containers for self hosting. Try out the live demo here.

duolahypercho/fusion-fable

The Fusion-Fable skill is a Claude Code skill that allows you to poll multiple cutting edge AI models about a single question. It will send the question to various AI model setups (such as running Opus 4.8 twice, running Opus and GPT-5.5 together, or even running a trio of models including Gemini 3.1 Pro), and then let each run free using web access and a shell. Once all models have had their say, Opus 4.8 will score and aggregate their responses, generating a final summary response.

Waishnav/devspace

DevSpace is a self-hosted MCP server that connects ChatGPT to your local machine. You install the CLI, point it at the folders you allow, expose it over a private tunnel and protect it with a password — then ChatGPT can inspect repos, edit files, run shell commands and even use git worktrees. It took off because it gives you AI coding help without shipping your files to a third party. The repo is real: TypeScript-heavy, 167 commits, CI in place, from a developer with a track record.

SunJaycy/GoldenEye-Recomp

Statically recompiled version of the Xbox 360 port of GoldenEye 007 that runs natively on Windows. This allows for mouse/keyboard controls, internet play, post processing effects, various graphical settings and higher FPS compared to the original. No game assets are included; you’ll need to provide your own legally obtained files.

royalbhati/sqltoerdiagram

An offline ER diagram generator for SQL CREATE TABLE statements. It features its own SQL parser, optimized canvas rendering for high table counts and an inline code editor with live preview. Relationships are highlighted on hover and focused by click on the pins. The position of the tables is saved across sessions and the diagram can be exported as PNG or SVG. This tool went viral because it solves a major hassle (understanding database schemas) effortlessly and completely offline.

NO6KIKO/gorest-2d-animation-spritesheet-generator

This is a local browser based editor to create 2D side scrolling game scenes. You can add sprites, backgrounds and objects as layers. Import/Create Sprite Sheets, Preview Animations and export assets to be reused locally. No need for any fancy desktop animation software. This was created with AI code assistance(Author likes to call it “Codex-assisted”). React frontend and a tiny server. Each scene has a page where all the animation clips are listed. So you can quickly go through all the animations.

mrtooher/fable-mode

A free skill for Claude that encourages the assistant to work more methodically by forcing it to create a plan, break down tasks into verifiable steps, delegate subtasks to cheaper models where possible, self-evaluate before returning results. This skill aims to replicate the structured approach of the new Fable models on top of your existing model. Note that this skill doesn’t make the model any smarter; instead, it helps you structure your workflow better for complex, multi-file/source tasks (and is useless for simple tasks). It became this week’s hottest new repo because everyone wants better, more robust AI workflows that produce fewer errors.