GitHub Finds for AI Builders

OpenBMB — the Chinese AI lab behind the MiniCPM model family — released MiniCPM5 2B, the second model in the MiniCPM5 series. It is a 2.5B dense transformer built for on-device deployment: laptops, edge hardware, and resource-constrained scenarios where a full server GPU is not available. The model features a 131,072-token context window and native tool calling, enabling agent workflows without any cloud dependency. In benchmarks, it achieves an average score of 53.9, exceeding all compared models in its class. It runs locally via Ollama, LM Studio, llama.cpp, and MLX.

Vincentwei1021/anything2explainer

anything2explainer is a skill for Claude Code and Codex that generates complete narrated explainer videos from any topic. You install it once by linking the repository into your skills folder, state your topic, and the agent pipeline takes over: it researches the topic with sources, writes the narration, generates word-boundary-aligned TTS voiceover, creates a shot-by-shot storyboard, and dispatches parallel build agents that write each frame as a Remotion (React) component. The output is a 1280×720 H.264 MP4 with synchronized voiceover, subtitles, chapter cards, and a progress bar. A 2–3 minute video runs in approximately 1 hour of wall clock time; a 3–5 minute reference-tier video takes approximately 2 hours. English and Chinese are both supported natively. You are consulted at exactly four checkpoints.

unslothai/unsloth

Unsloth is described by its authors as "the first desktop app to run and train models." Its web UI, called Unsloth Studio (launched with the command `unsloth studio`), lets you browse a model catalog, download, and run LLMs, diffusion, TTS, and embedding models locally. A drag-and-drop dataset builder accepts PDFs, CSVs, and DOCX files for fine-tuning without any code. Runs on Windows, macOS, and Linux with NVIDIA, AMD, Intel GPUs, CPUs, and Vulkan. You can optionally connect to ChatGPT or other cloud providers. Benchmark tables cover models from Google (Gemma), OpenAI (gpt-oss), Alibaba (Qwen), and Meta (Llama). Authors benchmark fine-tuning at 2× faster with 70% less VRAM versus standard pipelines, and MoE model preparation at 12× faster with 35% less VRAM. The project has 75 000+ GitHub stars.

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 is a platform that supplies AI coding agents with current, version-specific library documentation on demand. Without it, agents rely on training data that can be months or years out of date, producing code that calls deprecated APIs or uses old patterns. Context7 intercepts library questions and pulls live documentation from the source, placing it directly into the agent's context window. It works in two modes: as an MCP server (native tool calls) or as a CLI skill (no MCP required). A free API key from context7.com raises rate limits. Supported agents include Claude Code, Cursor, and OpenCode.

langgenius/dify

Dify is an open-source LLM app development platform. On a visual canvas you load your PDFs, Word files and spreadsheets, select a model — Claude, GPT-4, Llama, DeepSeek, or any OpenAI-compatible endpoint — and get a working chatbot or agent without writing code. Over 50 built-in agent tools are included: Google Search, DALL-E, Stable Diffusion, WolframAlpha, and code execution. Every app gets a REST API out of the box, so you can connect it to a Telegram bot or embed a widget on your site without a separate backend. Runs on Dify Cloud (sandbox plan includes 200 free GPT-4 calls) or self-hosted at home.

HKUDS/nanobot

Nanobot is an ultra-lightweight, open-source, self-hosted personal AI agent framework written in Python, developed by researchers at the Hong Kong University of Data Science. It runs as a WebUI you open in your browser, as a terminal client, or through chat apps — all connected to the same agent. The agent maintains long-term memory across sessions, uses tools through MCP integrations, can route between different AI models, delegate subtasks to subagents, run scheduled automations, and expose an OpenAI-compatible API for other apps. On macOS and Linux, install with a single curl command; on Windows, a one-line PowerShell script. No complex setup — open the WebUI and start talking. Works with any model provider: OpenAI, Anthropic Claude, or local models via Ollama. Released seven months ago; already close to 48,000 stars.

bilawalsidhu/gods-eye-view

God's Eye View is an open-source 3D globe simulator that displays live public data: aircraft transponders, ship beacons, satellites, earthquakes, traffic, and public cameras — all from open sources. A voice agent powered by OpenAI's Realtime API lets you control the globe by talking: 'Take me to Tokyo', 'How many flights are over Texas right now?', 'Switch to night vision', 'Outline the state of Texas'. The agent has 28 tools spanning camera control, world annotation, live-layer queries, and console commands. An AI-generated HUD overlays five-word intelligence-style scene descriptions that update as you move. The project went viral on YouTube (5M+ views) and reached #1 on GitHub Trending in August 2026. Released two and a half months ago; already near 20,000 stars. Install in one click via Pinokio, or run locally — keyless access covers flights, ships, satellites, earthquakes, and cameras. Full photorealistic 3D rendering uses Google Photorealistic 3D Tiles, accessible for free through a Cesium ion Community account or directly via a Google Maps API key.

achimala/dream-loop

Dream Loop is an agent skill that builds games, apps, and visual scenes through a closed creative cycle. First, the AI 'dreams' — generates a high-quality target screenshot of what the final result should look like. Then it builds the project using that image as a reference. A separate AI critic compares the live screenshot to the target and sends feedback. The agent loops back and revises until the critic is satisfied. Optionally, the AI can dream again — generating a better target based on the current state of the project. Works with any agent that has image generation, vision input, and subagent support; tested with GPT-6 Astra in Codex, and strong models like Claude Fable 5 should work too. Install with one command: `npx skills add achimala/dream-loop`.

Z-Image-Turbo is an open text-to-image model from Tongyi-MAI, Alibaba's AI research division. It is a distilled version of their larger Z-Image model, trained to match the original's quality while running in just 8 inference steps instead of the usual 50. A browser demo is available on Hugging Face — no GPU or account required on your end. You type a text description, set the width and height between 512 and 2048 pixels, and optionally set a seed for reproducibility. The model renders text inside images in both English and Chinese. The underlying weights are open for download.

SenseNova U1.5-8B-MoT is a multimodal model from SenseNova (a Chinese AI company) that handles both image generation and editing in one interface. Supported operations include object replacement, local edits, text correction on images, and partial scene changes with the rest preserved. A standout feature is readable text rendering — both English and Chinese text come out clearly, which most image generation models still struggle with. The model supports up to 4K resolution output and follows complex multi-step instructions with visual references.

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

An agentic plugin marketplace that ships 94 domain-specific plugins — 202 agents, 183 skills, 105 commands, and 16 multi-agent orchestrators — from a single source that adapts natively to Anthropic Claude Code, OpenAI Codex CLI, Cursor, OpenCode, Antigravity CLI, and GitHub Copilot. Each plugin covers a specific domain: Python development, frontend, SEO, security, ML, infrastructure, documentation, and more. Installing a plugin loads only its components into context, not the full marketplace. Individual skills can also be installed via `gh skill install wshobson/agents` or `npx skills add wshobson/agents` without cloning the repo. The marketplace has been live for about a year and has accumulated nearly 40,000 stars — one of the most-starred Claude Code plugin collections.

shanraisshan/claude-code-best-practice

A community guide to Claude Code best practices with 83 categorized tips covering prompting, agents, CLAUDE.md, hooks, MCP servers, workflows, memory, and more. Includes curated skill collections, real-world workflow examples from major projects, and a comprehensive index of everything new in Claude Code. Covers hot features like Ultrareview, Devcontainers, Auto Mode, Computer Use, and the Agent SDK. Updated continuously by contributors with input from the Anthropic team.

jtydhr88/screenwriting-skills

A Claude Code plugin that adds 12 specialized screenwriting and dramaturgy skills, compiled from 19 professional books on craft — McKee's Story, Save the Cat, Egri's Art of Dramatic Writing, and nine others — plus analyses of Chekhov's complete plays and six Ozu Yasujiro screenplays. Each skill covers a distinct area: story structure, premise and theme, character and conflict, dialogue, scene craft, format and adaptation, and industry business. The skill frontmatter descriptions are in English so both English and Chinese prompts trigger them; the methodological content inside the skills is in Chinese, which Claude reads and applies regardless of what language you write in. Installation is via Claude Code's plugin marketplace — one command adds all 12 skills.

inclusionAI/Choruz

Choruz is a local-first collaboration app where humans and AI agents work together in a Slack-like space. Each agent — Claude Code, Codex, Pi, Grok, OpenCode, or any webhook-driven agent — runs its own real CLI in a dedicated workspace with access to files, a terminal, and its own task board. You interact with agents through direct messages, group channels, threads, and mentions, just like a team chat. Agents can hand work to each other automatically. The app includes an integrated browser-based terminal, file browser and editor, SSH support for remote machines, and optional bridges to Slack and Telegram. Currently in Developer Preview — interfaces and data formats may change between versions.

github/spec-kit

Spec Kit is an open source toolkit from GitHub for spec-driven development with any AI coding agent. Instead of describing features informally, you write a structured specification defining what to build and why — not how. The CLI (specify-cli) adds slash commands to your agent (/speckit-specify, /speckit-plan, /speckit-tasks, /speckit-implement, /speckit-converge) that guide you through a full workflow: establish project principles once, describe the feature in a spec, get a technical plan, break it into tasks, implement, then converge the result against the original spec until it matches. Optional extensions cover bug fixing (assess → fix → test) and idea assessment (intake → research → define → shape → decide). Works with GitHub Copilot, Claude Code, and Codex. The project reached 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.

I've been running Qwen3.8-27B through Ollama all week — Qwen call it their most capable open model to date, and the numbers back it up: SWE-bench Pro at 61.7% and OSWorld-Verified at 84.3%, which means it can actually navigate your computer autonomously. Context goes to 262K tokens natively, extendable to 1M. The quantized version fits in 32 GB of Mac RAM. What I personally liked: you can dial thinking depth down when you don't need full reasoning — much faster for everyday tasks.

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.

Lightricks released LTX-2.5 this week — an open video model you run locally from text, images, or existing clips, no cloud subscription needed. What sets it apart: multishot support keeps your character looking consistent across multiple scenes, and audio generates in sync with the video, not pasted on afterward. I hadn't seen an open local model pull off both in one go before. You need a CUDA-capable GPU to run it — this isn't a laptop project.

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 — known for flagship multimodal models — built H3 to generate short video clips: give it a photo or a text description and it produces a clip with an audio track. The ModelTC team accelerated this with their Turbo variant, which requires several times fewer generation steps while delivering the same output quality. Ready-made ComfyUI workflows are bundled in, so nothing extra needs to be configured from scratch. The model handles both image-to-video and text-to-video inputs. It crossed 1.26 million downloads in its first month. A no-install demo is available at the team's official studio at x2v.light-ai.top.

yang0/handraw-style

A gallery of 216 numbered hand-drawn art styles for AI image generation, split into seven categories: international editorial illustration (001–035), narrative picture-book styles (036–054), modern graphic design (055–082), Japanese contemporary illustration (083–123), Chinese contemporary illustration (124–154), web and media aesthetics (155–200), and Chinese illustration supplements (201–216). Workflow: open the built-in HTML gallery, note a style number, type 'number + theme' — the skill returns bilingual prompts and passes a reference image to the generator only when the style cannot be activated by name alone.

lnkiai/m3e-canvas

I opened this in a browser tab and in three minutes had a rough sketch of the screen I'd been trying to describe to my agent for a week. It's built around Google's Material 3 Expressive design system — the same visual language Android uses — so the components look like real mobile UI right from the start. You drag them onto a phone-sized canvas, link screens together, and when you're done the tool writes the description for you — not code, just words your agent can understand. Everything lives in localStorage, no backend, no account.

pierrenade/short-video-generator-AI

Tried this on a forty-minute interview that had been sitting on my hard drive: gave it the YouTube link, walked away, came back to eight vertical clips with subtitles burned in and no watermark. It found the good moments on its own — not just the loudest, but the ones where something actually happened. The transcription runs on your own machine, only the 'is this interesting?' question goes to an outside model. MIT licence, no subscription, works up to 1080p.

vinzdg/codenotch

Installed this in thirty seconds — it sits on the edge of the screen and shows how much of my Anthropic Claude Code limit is left, right next to Cursor and Codex. It also watches Antigravity, which pulls usage from Google's quota endpoint, so that one is covered without a separate login. Before this I had to open each dashboard to find out why my agent suddenly went slow. Now I see the rings change colour, know instantly when I'm close to the wall, and can switch tools before the session blocks. macOS only, native, auto-updates.

ahujasid/camera-to-blender

I pointed my phone at my desk lamp, took a photo, and thirty seconds later there was a 3D model of it sitting in Blender. The setup took about ten minutes total — you run a small server, the phone connects over local network, and the AI turns the photo into geometry automatically. The model was recognizable and I could start adjusting it right away. If you make 3D scenes and want real objects in them, this shortcut changes the whole workflow.

PhiloLabs/fable51-worlds

I typed 'a night market in Tokyo' and ten minutes later walked around it in a browser tab. Every building, sign, and food stall was generated from scratch — no downloaded city data, just agents building it. The ready-made worlds they published open instantly with no account needed. To generate your own you need a Claude API key. The gap between 'I want to see this place' and 'I'm standing in it' is now one sentence.

tinyhumansai/openhuman

OpenHuman is an open-source personal AI desktop app for Mac, Windows, and Linux, built in Rust and available as a plain installer — no terminal or config files required. On first run, it pulls your email, calendar, documents, and chats, compresses them into local Markdown files in a SQLite-backed Memory Tree, and re-syncs every 20 minutes. It connects to over 100 OAuth integrations, 5,000 MCP servers, and 90,000 Skills, routing each workload to the right model automatically. TokenJuice compresses tool output before it reaches the model, cutting token usage by up to 80%. A Privacy Mode enforced in the Rust core means one switch stops all inference from leaving the machine. The app also orchestrates multiple agents simultaneously using visual durable workflows, reaches you through 17 messaging channels including Telegram, WhatsApp, Slack, and native email, and offers built-in image and video generation.

tigerless-labs/agent-memory

Agent Memory is a local-first, agent-agnostic long-term memory runtime for AI agents. Memories are stored as plain Markdown files in a single folder on your machine — greppable, git-trackable, and portable. An SQLite index sits beside them as a rebuildable cache; deleting it loses nothing. Supports Claude Code, Codex CLI, and any agent that can run a shell command. Writes trigger at conversation boundaries, not at the agent's discretion. A separate sleep-time Manage pass consolidates, ages, and archives by value — and can only propose deletions, never execute them without your confirmation. Benchmarked at 52.9% accuracy on LongMemEval-S vs 35.8% for MemCore and 5.8% with no memory. Runs on Linux, Windows, and macOS including Apple Silicon.

sickn33/agentic-awesome-skills

Agentic Awesome Skills (AAS Core) is a catalog of 2113 ready SKILL.md playbooks — reusable instruction files that AI agents can read and execute. Claude Code, Codex, Cursor, Google Gemini CLI, GitHub Copilot, and other agents can inspect your project and choose exact skills from the local AAS catalog. The agent owns the selection: AAS Core's read-only compose_stack tool validates the choice in memory, then the aas CLI persists it as aas-stack.json and shows an immutable plan before any change. Skills cover Git workflows, testing, documentation, SEO, security, content generation, and hundreds more scenarios. Available since 2024, the project has accumulated 46,000 stars across the AI coding community. Free, MIT license, community-maintained.

Rion-Wu-tech/wechat-intelligence-hub

WeChat Intelligence Hub is a Codex skill that converts your local WeChat (Tencent's messaging app) message history into structured personal intelligence. It extracts: contacts you haven't replied to, commitments you've made, business opportunities mentioned in group chats, reconnect leads, and generates time-range intelligence reports on demand. The system has two components — a read-only WeChat CLI reader (wechat-cli) and the intelligence engine (wechat-intelligence-hub) — installed together automatically. The reader accesses only your locally stored WeChat database; no data leaves your machine. Full database access requires authorized local credentials (the database encryption key); a demo mode with fictional data works without them. Installation is done by giving Codex one prompt from the README. The project README is primarily in Chinese.

punkpeye/awesome-mcp-servers

A directory of Model Context Protocol (MCP) servers — an open standard created by Anthropic that lets AI agents securely interact with local and remote resources. Covers production-ready and experimental servers for files, databases, and external services including OpenAI API, Apple Shortcuts, Apple Reminders, Google Drive, Slack, Notion, and GitHub. Works with Claude Code, Cursor, Claude Desktop, and other MCP-compatible agents. Implementations in Python, TypeScript, Go, Rust, and more.

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 Alibaba's Qwen team's text-to-speech model, available as a free browser demo on Hugging Face Spaces. It supports 10 languages — Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian — plus regional dialect variants. The model accepts text instructions to control tone, emotion, and speech pace, so you can make the voice sound upbeat, calm, or dramatic without extra tools. Voice cloning works from a 3-second audio clip, letting you reproduce a specific person's voice. End-to-end latency is around 97ms — responses come back almost instantly. The demo runs on ZeroGPU; no registration or installation is needed.

okf-memory/okf-agent-memory

OKF Agent Memory is described by its authors as 'a Domain-Neutral, Git-Native Persistent Project Memory for AI Agents based on the Open Knowledge Format (OKF) v0.2' — OKF is a memory standard developed by Google. When context windows reset, agents lose all architectural decisions and session facts. This tool stores memory as plain Markdown files with YAML frontmatter in a `knowledge/` directory inside your repository — files that live in git alongside your code and are read automatically on the next session start. An embedded MCP server connects Claude Code or any MCP-compatible agent without requiring external databases or additional services. A Progressive Disclosure system ensures the agent receives only the context it needs at each step — authors report 80% reduction in token consumption on long projects. An in-memory BM25 search index responds in under 300 microseconds. Supports both local and cloud LLMs including LM Studio and Ollama.

NousResearch/hermes-agent

Hermes Agent is described by its authors as 'The self-improving AI agent built by Nous Research.' It runs 24/7 on cheap cloud infrastructure — from a $5/month VPS or near-free on serverless backends like Modal and Daytona (hibernates when idle) — and accepts tasks via Telegram, Discord, Slack, or terminal. What distinguishes it: the agent learns from every task it completes, creating reusable skills so similar work is handled faster next time, and it builds a persistent memory of who you are across sessions. Supports any LLM backend — Claude, OpenAI, Nous Portal, or a self-hosted model. Seven execution environments are available: local, Docker, SSH, Singularity, Modal, Daytona, and Vercel Sandbox. Open-source under MIT license, installs in one command on Linux, macOS, and WSL2.

headroomlabs-ai/headroom

Headroom is a context compression layer built by Headroom Labs that intercepts everything your AI coding agent sends to the LLM — tool outputs, logs, RAG chunks, files, and conversation history — and compresses it before it reaches the model. JSON data compresses 60–95%, code around 20%, with answers remaining identical to uncompressed baseline. It attaches to your existing agent in one command (headroom wrap claude), works as a transparent proxy on port 8787 with zero code changes, ships as a Python or TypeScript library, and runs as an MCP server. Compatible with Claude Code (by Anthropic), Codex (by OpenAI), Cursor, Grok, Copilot, Cline, Goose, OpenHands, and many more. A shared cross-agent memory store keeps context consistent across Claude, Codex, Gemini, and Grok. The headroom learn command mines past failed sessions and writes corrections to CLAUDE.md to avoid repeating mistakes.

Pollen Robotics, the French company behind the Reachy humanoid robot, released a browser simulator for Microduck under the pollen-robotics namespace on Hugging Face. The space runs actual reinforcement-learning policies trained on physical Microduck hardware — not canned animations. Physics simulation runs in WebAssembly; the neural network policy executes via onnxruntime-web, all inside the browser tab without sending data externally. Two locomotion variants (legged and wheeled) are available, along with six behaviors: running, sitting, somersault, ball kicking, roller skating, and squatting. Concurrent visitors appear as semi-transparent ducks moving through the same scene in real time.

VoxCPM2 is a speech synthesis model from OpenBMB, the organization behind the MiniCPM model family. It covers 30 languages, outputs audio at 48kHz quality, and was trained on over 2 million hours of multilingual speech data — all at 2B parameters. The browser demo on Hugging Face supports three modes: voice design from a text description alone (no recording needed), controllable voice cloning from an uploaded audio sample, and an advanced mode with full style control. Upload a short recording and write any text — the model synthesizes it in that exact voice. No recording? Just describe the voice: age, accent, tone, and the system generates it. No installation required, no GPU needed.

rtk-ai/rtk

RTK (Rust Token Killer) is a high-performance CLI proxy, built as a single Rust binary, that sits between your terminal and your AI coding agent and compresses command outputs before they reach the model context. It applies four strategies — smart filtering, grouping, truncation, and deduplication — across over 100 supported commands including file operations (ls, cat, grep), git, test runners (pytest, cargo test, jest), build tools (TypeScript, ESLint), package managers, AWS CLI, Docker, and Kubernetes. Integration with Claude Code, GitHub Copilot, Cursor, Gemini CLI, Windsurf, and 12 more tools works via hooks or plugins. The reported reduction in bash output is up to 90%, with less than 10ms overhead. Available via Homebrew, curl install, Cargo, or pre-built binaries for macOS, Linux, and Windows. Apache 2.0 license. The project has 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 system from DeepLethe, built in Rust, that ingests documents and builds a bitemporal knowledge graph — meaning every fact carries timestamps for when it was valid and when it was recorded, preserving the full history of evolving understanding. Supported input formats include PDF, DOCX, PPTX, XLSX, CSV, Markdown, HTML, and connectors for Notion, GitHub, Jira, WebDAV, RSS, and S3. Search combines full-text (Tantivy), vector (pgvector), and graph traversal with RRF fusion, and streams answers with inline citations. A built-in AI agent searches documents, walks the knowledge graph, and queries mounted databases. Multi-user permissions, an append-only audit trail, ontology workbench, and forward-chaining reasoning are included. Released one month ago, already over 4000 GitHub stars. Currently v0.1; the database schema evolves with forward-only migrations — back up before upgrading.

affaan-m/ECC

ECC is an enhancement system for AI coding agents — Claude Code, Codex, Cursor, and others. It ships 68 specialized agents covering planning, code review, security, and build repair; 286 skills for TDD, research, documentation, and more; 94 commands; and a memory layer that learns from each session. After install, the agent follows a structured workflow — plan, test, implement, review, verify, remember — instead of jumping straight to code. It also includes AgentShield, which scans hooks, MCP configs, and agent files for injected prompts or leaked secrets.

aaif-goose/goose

goose is a general-purpose open-source AI agent with a native desktop app for macOS, Windows and Linux, a full CLI, and an embeddable API. It works with 15+ AI providers — Anthropic Claude, OpenAI, Google Gemini, Meta (Llama models via Meta's Model API), Ollama, OpenRouter, Azure, and more — through your own API keys or existing subscriptions. goose connects to 70+ extensions via the Model Context Protocol (MCP) open standard: file access, web browsing, databases, email, and beyond. Built in Rust for speed and portability. The project is part of the Agentic AI Foundation (AAIF) at the Linux Foundation, evolved from Block's original goose project. You can download the desktop app directly without any configuration, or install the CLI with a single command.

Hy4 preview is a new-generation MoE flagship model developed by the Tencent Hy Team. Total parameters: 770 billion, with 49 billion activated per token during inference — this architecture gives large-model quality without computing all weights simultaneously. Context length is 1 million tokens (roughly 750,000 words), enough for an entire codebase or a full book in one pass. The model covers four domains according to Tencent: software engineering (understanding, planning, debugging long-horizon development tasks), office work (converting context into documents, spreadsheets, and financial models), game development (turning text prompts into playable prototypes), and scientific reasoning across AI, physics, and mathematics. License: Apache 2.0. Deployment requires 8 tensor-parallel GPUs — this is a server-class model.

Ornith-1.5 is an agentic coding model from ornith-ai, trained via iterative self-improvement: the model solved real codebase issues, received feedback on results, and improved across training iterations. At 4-bit quantization it weighs 22 GB — the fast coder of its class, comparable to Claude Opus 4.6 medium on real codebase tasks, with the best multi-turn conversation quality measured among local models. On SWE-bench Verified: 79%; Terminal-Bench: 67–68%; MCP-Atlas (real agentic tasks with tools): 70.2%. Context window: 262K tokens, extendable to ~1M with YaRN. Supports native tool calling, MCP, and OpenAI API compatibility. peculiar-ragdoll packaged the original ornith-ai weights as GGUF for local deployment via Ollama, llama.cpp, and similar runtimes.

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 developer control center for running and managing multiple coding agents at once. Instead of opening each agent in a separate terminal or window, you connect them all — OpenHands, Anthropic's Claude Code, Codex, Gemini, or any agent that implements the Agent-Client Protocol — into one web interface running on your own machine. From there you start conversations, create automations triggered by GitHub events or Slack messages, set up schedules, and switch between local, remote, and cloud backends without losing context. Each backend runs agents in an isolated environment, and you can share an agent server with your team for shared workflows like automated code review. The project has been open-source for 18 months and accumulated 86,000 stars; a commercial cloud offering exists alongside the free self-hosted version.

nexu-io/html-anything

html-anything is a local-first web app from the nexu-io team that turns any input — Markdown, CSV, Excel, JSON, plain text — into a publish-ready single-file HTML document by driving whatever coding-agent CLI you already have installed. On startup it auto-detects Claude Code, Cursor Agent, OpenAI Codex, Gemini CLI, GitHub Copilot CLI, OpenCode, Qwen Coder, and Aider. The tool ships 75 composable skill templates across nine output surfaces: magazine articles, keynote decks, résumés, posters, Xiaohongshu and tweet cards, web prototypes, data reports, and Hyperframes video scripts. The agent streams the result live in the browser preview, then one click exports it: juice-inlined CSS pastes into WeChat without reformatting, a 2× PNG lands straight in the X or Instagram composer, or you download the standalone .html file. No extra API key is required — it reuses the session you already have active via claude login, cursor login, or codex auth. Launched four months ago, already at 8 k★. Apache 2.0 license.

n8n-io/n8n

n8n is a fair-code workflow automation platform that has been running for seven years and has over 200 k stars. It combines a visual canvas with the option to write custom JavaScript or Python where needed. With 1,500+ integrations and 9,000+ community workflow templates, you connect services visually — trigger a workflow from an incoming email, a new CRM entry, or a webhook, then route data through conditional logic, transformers, and AI agents that decide what to do next. Native AI capabilities let you embed any OpenAI-compatible model (including Anthropic, Google, or local Ollama models) directly into a workflow step. You can self-host it with a single Docker command or use the n8n cloud. The source code is always visible; additional enterprise features are available under a separate 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 the first natively multimodal model in ZhipuAI's GLM-5 series — unlike previous GLM versions, it processes both text and images in a single request. It uses a Mixture-of-Experts architecture: 320 billion total parameters, with 18 billion active per request. Supports up to 300,000 tokens of context — enough for several large documents at once. Benchmarks show it surpasses the previous GLM-5.2 across the board. Designed for agentic tasks, coding, and document analysis involving images. Unsloth packaged it as a GGUF for Ollama, but running it requires server-grade hardware with hundreds of gigabytes of RAM.

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.

diegosouzapw/OmniRoute

OmniRoute is a self-hosted AI gateway that sits between your coding tool (Claude Code, Cursor, Cline, Copilot, Codex) and the AI model providers. It exposes a single OpenAI-compatible endpoint at localhost:20128 and routes requests across 352 providers and 1,200+ models, switching automatically when a quota is exhausted. Out of the box, with zero API keys, it works through keyless free providers; with keys added it stacks ~1.51 billion free tokens per month across documented free tiers. It also runs a token compression layer (RTK + Caveman) that reduces the output bash commands send to the model by 15–95%, which lowers effective costs further. Available as an npm package, a Docker image, a desktop app, and a PWA. Has MCP and A2A support. The project has 600+ contributors and ships a Russian-language README.

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 a reference architecture from Anthropic for two Claude-powered agents: a shopping agent embedded in your customer-facing storefront, and a merchant agent for back-office staff. The shopping agent searches products, compares options, fills the cart, answers policy questions, and remembers customer preferences across sessions. The merchant agent explains sales performance, updates listings, adjusts prices and promotions — and every write is staged for human approval before it takes effect. Four verticals are included as runnable demos: retail, travel, telecom, and entertainment. A Claude Code plugin scaffolds a new commerce agent from your description of the stack and business rules. Requires Python 3.11+ and Node 22, plus a Claude API key.

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 27-billion-parameter model from Alibaba that ranks among the top open models for reasoning, math, and coding tasks. ISTA-DASLab (Austrian Institute of Technology) released a GGUF build using non-uniform quantization: each tensor is compressed according to its individual sensitivity, which preserves more model quality at the same file size. The build explicitly includes the vision projector, enabling the model to process screenshots, diagrams, and photos alongside text. File sizes range from 8.4 to 11.8 GB depending on the quantization level — fits on any laptop with 16 GB of RAM or more.

subsy/skill-cabinet

Skill Cabinet is a local browser UI that scans every agent skill folder on your machine — .claude, .codex, .cursor, .agents, Hermes profiles, and any ~/.*/skills path — and shows all discovered skills in one place. For each skill you see the file body, YAML frontmatter, GitHub origin, and whether it is a real file, a symlink, or a broken reference. The UI also finds duplicate skills installed across multiple agents and flags elevated-risk patterns in skill bodies. Quarantine removes a skill from all live agent drawers without deleting it from disk — restoring puts it back. You can also delete one or many skills at once. Launch with `npx skill-cabinet` (Node 20+ required); the browser opens automatically on port 3781.

stablyai/orca

Orca is a desktop app for Mac, Windows, and Linux that orchestrates multiple AI coding agents — Claude Code, Codex, Pi, OpenCode — running in parallel, each in its own isolated git worktree. You fan one or several prompts across agents, watch them work simultaneously, compare the outputs, and merge the winning branch. A mobile companion app (iOS and Android) lets you monitor agent progress and send follow-up instructions from your phone. Design Mode lets you click any element in a live Chromium browser window and send its HTML, CSS, and a screenshot directly into an agent's prompt — useful for UI corrections without copy-pasting. Orca launched six months ago and has collected ~60k stars.

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

Open-source SEO and GEO skills for Claude (MIT license), built on the Ryze MCP connector. Eight skills pull real data from your own accounts: Google Search Console for actual rankings and clicks, GA4 for traffic including AI referrals from ChatGPT, Perplexity, and Claude, Google Ads for keyword volumes and CPC, and DataForSEO (already wired in) for competitor keywords, backlinks, and SERPs. Skills cover: full site audit with CTR anomalies and decaying pages, keyword research with clustering, rank tracking between any two periods, competitor keyword gap analysis, backlink profile check, AI visibility reporting, content brief generation, and paid-vs-organic keyword overlap. Install is two steps: add the Ryze MCP connector in Claude settings, then install as a Claude Code plugin. Data flows through your own accounts with no markup.

Player-YN/PawWork_ZhuaZhua

PawWork is a Chrome MV3 extension that works directly on pages you already have open. You activate Paw Mode, click on any element on the page — a data table, a text block, an image, or a link — and describe what you want in a sidepanel. The AI generates: live spreadsheets (Univer), design canvases (tldraw), editable documents, or full HTML sites from the selected content. For complex tasks it shows a plan for your approval before executing. All generated files are stored locally in IndexedDB and OPFS — nothing leaves your machine. Generated code runs inside an isolated WASM sandbox. You bring your own model key; OpenAI, Anthropic, and other providers are supported.

open-webui/open-webui

Open WebUI is a self-hosted AI platform designed to run entirely offline. It provides a feature-rich browser interface for interacting with AI models: connect local models via Ollama or any OpenAI-compatible API — including OpenRouter, LMStudio, GroqCloud, Mistral, vLLM, and dozens more. Key features: full chat history with search, file and image uploads, voice input, RAG (web search, documents), model management UI, detailed user roles and permissions, and a plugin/tool system for extending capabilities. Installs via Docker, pip, or Kubernetes. Has grown to 150,000+ stars over three years and is the 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`.

This Hugging Face space runs Wan-AI's Wan2.2 image-to-video model (14B parameters) directly in the browser via ZeroGPU. Upload a photo, write a text prompt describing the motion you want, and the space generates a short animated clip. No GPU, no environment setup, no installation required. The space uses AOT (ahead-of-time) compilation for noticeably faster inference than a standard deployment of the same model. You can adjust video length and inference steps.

PhoneLLM Alpha 1 is a 32B open-weights model from Pipecat, designed specifically for voice agent workloads. Most language models struggle with voice: high time-to-first-token, losing track in multi-turn conversations, and unreliable tool calls mid-conversation. PhoneLLM addresses these directly — low latency responses, reliable multi-turn dialogue, and accurate tool invocations. According to Pipecat, it performs on par with GPT Terra 5.6 but at 94% lower cost and 1.3 seconds faster P95 time-to-first-token. BSD-2 license.

thedotmack/claude-mem

Claude-Mem is a persistent memory compression system built for Claude Code. It automatically captures tool usage observations during sessions, generates semantic summaries, and injects them into new sessions — so Claude maintains continuity of knowledge about your project after every restart. Memory can be backed by the cloud service at cmem.ai (free trial) or your own OpenRouter, Gemini, or Anthropic key. Also supports OpenCode and Antigravity CLI. Includes a real-time web viewer, natural-language history search via a skill command, and a `<private>` tag to exclude sensitive content from storage. The project has grown to over 92,000 stars since launch.

Leonxlnx/taste-skill

Taste Skill is a portable set of design instructions for AI coding agents that prevents them from producing generic-looking interfaces. The authors call it an 'Anti-Slop Frontend Framework' — rules the agent loads as a design brief that steer it away from predictable patterns. Three tunable dials control layout variance (centered and clean vs. asymmetric and modern), animation intensity (hover only vs. scroll and magnetic effects), and visual density (spacious vs. dense dashboards). The repo ships several skill variants: the general v2 default, a minimalist variant in the Notion/Linear style, an industrial brutalist variant, an expressive high-motion variant, and a redesign skill for improving an existing codebase. There are also image-generation skills to produce visual reference boards first, then hand the frames to the coding agent for implementation. Works with React, Vue, and Svelte — and with any major coding agent.

JuliusBrussee/caveman

Caveman is a skill for Claude Code and 30+ AI coding agents that trains your AI to respond without filler: no preambles, no summaries of what you just said, no explanations of things you already know. According to the authors' benchmarks across 10 tasks, output tokens drop 65% on average. The biggest wins are in browser queries: 15,704 tokens compressed to 121, a 130× reduction. The skill runs as a transparent proxy wrapper or installs directly as a Claude Code plugin. Authors are honest about the limits: only output tokens shrink, and the skill itself adds ~1–1.5k input tokens per turn, so on short conversations the net effect can be negative. The main benefits are faster responses and easier reading.

GLM-5.3 is the flagship open-weights language model from Zhipu AI (Z.ai), the team behind the GLM series. The model tops coding benchmarks including Terminal Bench 3.0 and DeepSWE, outperforming all prior open-weights competitors. All improvements come from post-training, not architectural changes — Zhipu focused on making the model handle agent workflows and large codebases better. Context window is one million tokens. Available in English and Chinese, accessible via vLLM, SGLang, and OpenAI-compatible APIs.

Connected it through the OpenAI-compatible API and the difference from previous Qwen versions is immediate. Qwen released a 180B multimodal model that processes text, images, and video in a single request with a native 262K-token context, extendable to 1M. GPQA Diamond hit 91.7%, SWE-bench Pro 62.5% — plugged it in as my agent's brain and tasks that tripped up the last version now go through cleanly. One caveat: it's an experimental architecture preview, and the license is qwen-community rather than Apache, so commercial use needs checking.

Tested it through the Kimi API — 1 million tokens of context means you can load an entire project and ask it to figure things out. Moonshot AI released what may be the world's largest open-weight multimodal model: 2.78 trillion parameters, understands text, images, and video in one request. Connects as an OpenAI-compatible API so your agent needs zero code changes. HuggingChat integration is already live for free browser access.

Ran it on CPU — fast, under 0.5 GB, sounds clean. Audio8 released a tiny TTS model with zero-shot voice cloning: give it text and 3–5 seconds of reference audio and it generates speech in that voice. Multiple languages supported. Free for projects under $2M annual revenue — covers most indie use cases without a subscription.

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

A Claude Code plugin built specifically for AI-generated code. It automatically measures the complexity of each function, finds the most tangled hotspots, and refactors them — flattening nested conditions into cleaner, direct logic. After every pass it generates a before/after complexity table, so you can see exactly which functions improved and by how much. Works on any language supported by standard complexity analyzers, and respects your project's own linter thresholds when configured.

openclaw/openclaw

OpenClaw is an open-source AI assistant that runs locally on your devices and connects to the messaging channels you already use: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, and more. All requests and model responses pass through a local Gateway — the central component that manages sessions, tools, and channel connections. You choose the model (any compatible one), connect channels and tools — the same configuration works for a single user or a trusted team. Optional companion apps add voice, Canvas, camera, screen, and device-local actions. Developed by the non-profit OpenClaw Foundation with community contributions. Supports macOS, Linux, and Windows; installs via curl or PowerShell in one command.

mouredev/hello-sdd

This repository accompanies a course by Spanish developer educator mouredev on Spec-Driven Development (SDD) with AI agents. The core idea: before handing a task to an AI coding agent, you write a structured specification so the agent follows a concrete plan instead of improvising. The repo provides three reusable template files: AGENTS.md (context the agent needs about the project — works as Claude.md), spec.md (numbered functional requirements in EARS notation, edge cases, out-of-scope, acceptance criteria), and prompts.md (ready-made prompts for each SDD phase: constitution, spec, clarification, planning, tasks, implementation, validation, and change). A complete example project built step-by-step with SDD is included. There is also a reusable spec-generator skill for Claude Code in .claude/skills/spec-generator/ — it interviews you about your project and generates a spec.md following the template. The skill is also linked for opencode in .opencode/skill/.

jub0t/Concat

Concat is a free, open-source video editor built as a direct replacement for CapCut — without watermarks, paywalls, or subscriptions. A native Rust engine handles media processing while a React interface provides the editing experience, and everything runs locally on your machine. AI-powered auto-captions are generated by a bundled Whisper model: your audio is never sent anywhere. Local text-to-speech is also included. Feature set covers multi-track editing, split/trim/merge, transitions, speed control, voice filters, styled text overlays, and reusable templates. Available for macOS (unsigned binary — one quarantine-removal command in README), Windows, and Linux. Currently in Alpha at version 0.2.0, so the core workflow is functional but rough edges may appear.

hkqr/my-free-code

My Free Code is a local gateway that sits between your AI coding agents (Claude Code, Codex, Cline, and others) and AI providers. Agents call the gateway using standard Anthropic or OpenAI-compatible protocols, while the gateway routes to the actual provider: Anthropic, DeepSeek, Groq, xAI, NVIDIA NIM, OpenRouter, and hundreds more. Tier routing (Fable, Opus, Sonnet, Haiku) is configured via environment variables — for example, MODEL_SONNET=deepseek/deepseek-chat. On provider failure it automatically falls back to the next in chain without duplicating already-streamed output. Includes a web admin UI at /admin, token counting endpoint, model discovery, authentication, and provider health monitoring. Runs on Python 3.10+ at http://127.0.0.1:8082.

google-gemini/gemini-cli

An open-source AI agent by Google that works in your terminal. The free tier gives 60 requests per minute and 1,000 requests per day using a personal Google account — no payment required. It has a 1M-token context window, which means it can hold an entire large codebase in view at once without needing to break it into pieces. Built-in tools include Google Search grounding, file operations, shell commands, and web fetching. Supports MCP (Model Context Protocol) for custom integrations and extensions. Fully open source under Apache 2.0.

damejan80/tokentab

Tokentab is a local token usage tracker for AI coding tools: Claude Code, Codex, Cursor, and Gemini CLI. It reads the session logs these tools automatically save to disk and calculates token counts and costs — real numbers from logs, not estimates. Data is broken down by day, project, and work type (coding, debugging, refactoring). The web dashboard runs on localhost:4747 using Python's standard library with no external dependencies beyond Rich for terminal formatting. Pricing is handled by a local lookup table in prices.py with no network requests. If a tool isn't installed, it's skipped. The CLI supports daily, monthly, and custom date range filters; the web server can be skipped entirely if you only need a terminal summary.

browser-use/browser-use

Browser Use is a Python library that gives an AI agent full control over a web browser — it opens pages, clicks, types, fills forms, and extracts structured data, exactly as a human would. You describe the task in natural language and the agent handles execution. The library integrates directly with AI coding tools: a single prompt installs it as a skill in Claude Code, Codex, Cursor, or similar agents, after which you simply tell your agent what to do in the browser. Typical use cases shown in the README: fill in a job application from a resume, extract a followers list as CSV, research and compare products, monitor price changes. It works with any LLM via API — Anthropic, Google, OpenAI, or Browser Use Cloud. The project has 111k stars and active development from 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

A skill that turns OpenAI's ChatGPT web app into the planning and review brain for your coding sessions, while OpenAI's Codex retains full execution ownership. No ChatGPT API key required — it connects through the official web UI via a secure, OAuth-protected, read-only MCP bridge. Your repository is never uploaded: ChatGPT reads only the lines it needs from your current workspace on demand. The skill self-updates daily from GitHub. Setup is one paste to Codex: it installs dependencies, builds the project, and runs the initial setup automatically. The only step you do manually is log into ChatGPT or enter a 2FA code. Documentation is in both Chinese and English.

tt-a1i/simplify-codebase

An agent skill for Codex and Claude Code that finds and safely removes accidental complexity from existing codebases. The core principle, as the authors state: 'deleted lines are an outcome — the durable gain is deleting a fact, state, contract, or concept that no longer needs maintenance.' Before removing anything, the skill builds a proof record: what maintenance burden it creates, who the consumers are (production, test, dynamic, external), what observable behavior would be surrendered, and the smallest check that would expose a wrong cut. It works in two modes: Survey (read-only audit of candidates with ranked evidence) and Change (apply proven simplifications with independent validation per batch). It treats public APIs, dynamic loading, stored formats, migrations, and architectural constraints as first-class evidence — if a real consumer exists or a boundary is unresolved, the skill recommends retaining rather than forcing a deletion. Primary README is in Chinese; an English version (README.en.md) is available.

OnlyTerp/opengrok

OpenGrok is a Python tool that adds a model picker to Grok Bot, letting you choose any supported AI provider per agent — xAI Grok, Zhipu GLM, Anthropic Claude, Google Gemini, DeepSeek, or local llama.cpp models. Every provider mapping is backed by a wire capture in the repository: the tool's authors describe their rule as 'evidence or it doesn't ship.' The setup command detects your Grok Bot installation, adopts existing bindings or asks three questions, and writes a config. The picker lets you test a binding live with one click before saving. A doctor utility baselines your machine on setup and monitors files and services — after any Grok Bot update it reports exactly what moved and what needs to be re-applied. Provider maps hot-reload without a restart. Works on Windows, macOS, and Linux with zero external dependencies beyond Python.

oboroge0/hayamimi

Hayamimi (Japanese for 'quick ear') is a Python tool for real-time multilingual speech-to-text, running entirely on CPU with no GPU or cloud API. It routes each utterance to the best specialist model for its language using quantized (INT8) ONNX models — Japanese, Chinese, Korean, Cantonese, English, and 24+ EU languages each go to a dedicated model, with everything else falling back to Meta's Omnilingual ASR. On real broadcast Japanese audio it achieves 5.8% character error rate, less than half of Whisper large-v3-turbo at 13.8%, at 10-50x realtime speed on a 6-core CPU. Partial subtitles update every ~0.5 seconds while you're speaking; a finalized line lands ~100ms after you stop. The browser dashboard shows live subtitles, speaker labels (S1/S2), and translations. There's also an OBS overlay for stream captions. Memory stays under 2 GB by default. Requires Python 3.10+, ffmpeg, and ~3.1 GB of downloaded models (a minimal Japanese/English install is ~1.1 GB).

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

A Claude Code skill that does what SEO agencies charge for — except it's an AI agent, not a team of contractors. It audits your site across five channels: classic SEO (sitemaps, meta, structure), AEO (Google AI Overviews and Bing Copilot), GEO (citations in ChatGPT and Perplexity), LLMO (what LLMs know about your brand), and NEO (Naver search and AI Briefing, which the author calls 'the channel most global SEO guides ignore'). For each channel it returns a scorecard with concrete evidence, not generic advice. After the audit the agent fixes technical issues, adds llms.txt and JSON-LD, and creates pages on the 'one question = one page' principle. It then schedules a remeasurement and proves the change with numbers. The author applied this exact playbook to a real Korean stock research site and recorded 1.54 million search impressions in 30 days with zero ad spend.

KKKKhazix/sun-style-writing

A .skill file for Claude Code, Codex, Kiro, and compatible agents that encodes 10 narrative techniques extracted from a specific Chinese literary text. The techniques include plain description (show actions, not emotions), contrast (juxtapose numbers and facts), repetition with variation, foreshadowing and callback, object narration (let things speak instead of characters), imagery, time montage, silence (stop where the character stops knowing), anticlimax, and circular structure. After installation the agent writes more concretely and stops substituting generic emotional language for actual scenes. The techniques are universal and work in any language, not just Chinese.

kacperkapusciak/goldie

Goldie generates App Store screenshots and preview videos directly from the iOS simulator. It captures screens using the argent testing framework, wraps each in a device bezel (iPhone 17 Pro and other models), adds a background and headline, joins clips into a preview video, and checks the result against Apple's upload rules. Works with SwiftUI, UIKit, Flutter, React Native, and Kotlin Multiplatform apps. The studio interface lets you pick font families (Merriweather, DM Sans), layout templates (editorial, showcase, magazine), and background colors. The key integration: Claude Code and Cursor can trigger Goldie via a simple text prompt — the agent runs it and returns finished files without you touching a terminal. Output screenshots are 1320×2868 px for 6.9" devices; preview videos at 886×1920 H.264, 15–30 seconds.

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 and text-to-speech space by k2-fsa that supports over 600 languages. You upload a reference audio clip of a person's voice, type any text, and the system synthesizes speech in that voice and language. It runs on HuggingFace Zero (free GPU tier) via a Gradio interface, so nothing needs to be installed. The model can adjust speaking speed and duration of the output. It uses a combination of Qwen3-0.6B for text processing, OpenAI Whisper for reference transcription, and k2-fsa's own OmniVoice model for synthesis — total about 1.6B parameters for the full pipeline.

Qwen3.8-27B from Alibaba is a powerful multimodal model, but in its original form it runs a long reasoning chain before every reply — useful when you need deep analysis, but slow for everyday tasks. DavidAU fine-tuned it using the Cold Fusion + GAIN method: the model now produces reasoning chains 50–90% shorter while maintaining the same answer quality. Multi-Token Prediction (MTP) variants push output speed above 90 tokens per second. The model is packaged as GGUF and runs locally in LM Studio, Ollama, or llama.cpp. Context window is 256,000 tokens, extendable to 1 million. Three reasoning intensity modes let you balance speed and depth per request.

totec448-spec/chat-on-steroids

Chat On Steroids is a desktop app for Windows, macOS, and Linux that turns ChatGPT into an agent that can actually work with your local files, run terminal commands, and — on Windows — control the screen, mouse, and clipboard. It runs a local MCP server and bridges ChatGPT to your machine through OpenAI's Secure MCP Tunnel. You keep using ChatGPT in the browser; the app sits in the system tray as the permission layer. You choose which folders it can access, whether it can run commands, and which capabilities are on. A bundled Chrome extension adds session recording, Compact & Resume, and an experimental multi-agent goal loop via OpenRouter.

StarTrail-org/PixelRAG

PixelRAG is an open-source visual retrieval system from Berkeley SkyLab and BAIR: instead of parsing pages to text, it renders documents — web pages, PDFs, images — as screenshot tiles and retrieves over those images directly. Visual structure that HTML parsing discards — tables, charts, infographics, layout — is preserved in the screenshots, so a vision model can answer questions about it. The project ships a pre-built hosted index of 8.28 million Wikipedia pages, queryable for free at pixelrag.ai without an API key. For Claude Code users, the repository includes a pixelbrowse skill that adds a /screenshot command: Claude renders a page with Playwright and reads it as an image, seeing charts and tables the way a person does. No MCP server, no backend, no GPU — the skill just calls the pixelshot command installed locally on your machine.

guillaumemeyer/watermarks-remover

watermarks-remover is an agent skill and local Python service that strips AI provenance marks from text and files you own. It works at two layers: deterministic scripts remove invisible Unicode characters, exotic spaces, and bidirectional control marks; a statistical rewrite hook handles token-sampling watermarks embedded in word choices. Coverage spans Claude (Anthropic), Gemini/SynthID-Text (Google), OpenAI, and open-LLM watermarking schemes. Beyond text, it removes C2PA, EXIF, XMP, and document property metadata from PNG, JPEG, WebP, PDF, DOCX, XLSX, PPTX, EPUB, HTML, Markdown, MP4, WAV, FLAC, and more. It installs as a Claude Code plugin via the marketplace — the skill calls a local service over HTTP, so the agent host needs no Python. A PostToolUse hook runs automatically on every Write, Edit, and MultiEdit call.

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.

experientiallabs/experiential

Experiential is an open-source AI model gateway and router from Experiential Labs. It exposes a single OpenAI-compatible API endpoint that proxies to any combination of hosted providers (OpenAI, Anthropic, Gemini, Azure, Bedrock, Fireworks, OpenRouter) or local models — using your own API keys (BYOK). You can set per-agent and per-use-case spend limits, track model traffic and costs in a dashboard, and then train a custom router from your actual production traces: easy requests route to a cheaper model automatically, complex ones go to the best available. The hosted platform at platform.experientiallabs.ai lets agents like Claude Code, Cursor, and Codex set themselves up via a plain-language prompt.

Ornith 1.5 35B A3B is a mixture-of-experts reasoning model for software engineering, built by ornith-ai. Despite the 35B parameter count, only 3B are active during inference — making it faster than a dense 35B model. On SWE-bench Verified it reaches 79%, a benchmark score that was out of reach for local models until very recently. The context window spans 262,144 tokens, long enough for an entire codebase in one session. It supports tool calling out of the box and fits naturally into agentic coding pipelines. This GGUF repository provides quantized versions ready to run via Ollama or llama.cpp without any additional setup.

MiniMax H3 Turbo is a video generation model from MiniMaxAI, a leading Chinese AI startup. It generates short video clips from a text prompt or a reference image. MiniMaxAI published an official browser demo on Hugging Face Spaces — the model runs on MiniMax's own servers, so no local GPU or installation is required. The space includes style controls that let you adjust the visual character of the generated video. This is the quickest way to evaluate H3 Turbo's output quality without setting up ComfyUI or renting server compute.

Tencent/WeMM-Embedding

WeMM-Embedding is a family of universal multimodal embedding models developed by the WeChat Vision team at Tencent. The models encode text, images, videos, visual documents, and interleaved multimodal inputs into a shared vector space, enabling cross-modal retrieval with a single query — for example, finding video clips or documents that match a text description, or locating images similar to an uploaded photo. Available in three sizes — 2B, 4B, and 9B — with all weights published on Hugging Face under the tencent/ namespace. The models run via standard transformers or sentence-transformers and are compatible with vLLM and SGLang for production serving. A Matryoshka embedding design lets you truncate vectors to smaller dimensions (from 64 to 4096) without retraining, trading accuracy for storage and speed.

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 an open-source local AI video production tool for Windows, created by the Chinese team "有温度" (You Wen Du). You upload 10–30 seconds of reference audio for voice cloning, paste a Chinese text script, and choose from 12 visual templates (from minimal whiteboard to retro collage and cyberpunk). The system automatically splits the script into scenes, generates illustrations, animates them with hand-drawn strokes, syncs subtitles, and exports a finished MP4. All source files, API keys, and output videos are stored locally; multiple computers on the same network can share a single task queue. Voice cloning uses a local IndexTTS 2.5 service; image and text generation go through the OpenLux API.

alchaincyf/huashu-excel

huashu-excel is an agent skill installable in Claude Code, Cursor, Codex, and any agent that supports the skills standard. It adds a full Excel analysis pipeline: before computing anything, the agent runs a structured health-check that catches misplaced headers, merged cells, subtotal rows mixed into the data body, duplicate records, text-formatted numbers, and inconsistent date formats. After cleaning, it cross-verifies computed sums against the table's own subtotal rows. If anything doesn't reconcile, the skill reports the discrepancy explicitly and withholds the final number until it's explained. The author demonstrated this on a real Chinese sales table: a standard pd.read_excel() approach returned a total 161% above the true value with zero errors or warnings. huashu-excel caught all three root causes — a subtotal row counted as a store, a regional subtotal counted again, and a duplicate row — at the diagnostic stage. The only external dependency is openpyxl.

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 video generation model from the Wan-AI team: given an image and a text prompt describing the desired motion, it produces a 720P video at 24 fps. The r3gm space on Hugging Face runs Wan 2.2-14B through an optimized inference stack (fp8 quantization + ahead-of-time compilation) on free ZeroGPU infrastructure — no sign-up required. Wan 2.2 is a major upgrade over Wan 2.1: trained on 83% more video data, it features a new mixture-of-experts design that splits the generation process across specialized blocks, making motion more fluid and stylistically consistent.

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

Open-Higgsfield is an open-source alternative to Higgsfield AI: a web app for AI image and video generation through a unified interface. Instead of multiple separate services, it brings together 40 models — 12 for images and 28 for video, including Flux, Kling, and Veo. A single prompt bar handles everything; each model retains its own settings. You can request up to four results per prompt, and the last 60 requests are stored locally in the browser. The gallery has four sections: images, video, assets, and favorites. A hosted version is available at openhiggsfield.ai — no installation needed, but a generation service API key is required.

cloudflare/computer

Cloudflare Computer is an SDK from Cloudflare that gives AI agents a virtual computer: a persistent filesystem backed by SQLite in a Durable Object, and a code execution environment, all running on Cloudflare's infrastructure. Three execution backends are available: a full Linux container mounted via FUSE, a shell runner using just-bash in a Dynamic Worker, and a JavaScript module runner. All three share the same durable filesystem through a single API endpoint. The repository includes a ready-made MCP example (examples/mcp) that connects an agent to the workspace directly via MCP tools — so your Claude Code or other agent can read and write files and run commands without you building the plumbing. Cloudflare explicitly marks it as preview only and not suitable for production, but the examples folder has runnable demos for every backend, including a tutorial that writes a markdown file and converts it to PDF via pandoc in a container. The project reached 8700 stars in under three months.

bam-bam-2/solo-skills

Solo Skills is a kit of 26 AI agent skills published by a Korean solo entrepreneur who built 49 automated workflows for his one-person business. The skills are structured as SKILL.md files — a format that Claude Code, Codex, and OpenCode read and execute directly: the agent treats each skill as a ready-to-follow procedure. The kit covers creating product demo videos using iframe capture and ffmpeg instead of a screen recorder, compiling PDFs from HTML via Paged.js, publishing Threads threads, sending Discord reminders through launchd, and offloading heavy jobs to a remote machine. A standout skill named claude-codex-fallback runs Claude first and silently switches to Codex only when a rate limit is hit. Some skills are tied to Korean services (Naver Mail, Threads API), but the rest work anywhere. Installing takes two commands: clone the repo and copy the 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 browser-based image generation space built on Black Forest Labs' FLUX.2 Klein — the compact variant of their second-generation image model. Where the full FLUX.2-dev reaches 32 billion parameters, Klein runs at 4–9 billion, making it faster and less memory-hungry. The key feature of this space is simultaneous multi-LoRA support: you can load two or more style adapters at once and control the weight of each independently, which lets you blend visual aesthetics in a single generation pass without switching tools. It runs on Hugging Face's Zero GPU infrastructure, so nothing needs to be installed on your end. The space also carries an mcp-server tag, meaning agents can connect to it programmatically.

Free AI Detector is a browser-based tool by Lynote that analyzes text and estimates how likely it is to be flagged as AI-generated. Paste any text and the tool returns a score showing how 'human' or 'AI-like' the writing reads. No account, no installation — it runs as a web app on Hugging Face Spaces without requiring any local setup. It appeared on August 5 and collected 330 likes and eight community discussions in three weeks. AI text detection services typically sit behind paid subscriptions; this one removes that barrier entirely.

rome-os/rome

Rome is an open-source agentic OS (MIT) where agents don't just answer questions — they build persistent, reusable capabilities. Describe a task in plain language and Rome turns it into a Rome App: a purpose-built interface with agents, persistent data, schedules, and hooks. Email triage, price trackers, code review loops, and morning briefings all become separate apps that keep running while you're offline. A community App Store lets you install apps built by others or publish your own. The self-evolution loop: each new task you describe widens what your Rome instance can do. Runs via Docker with a one-liner quickstart, or try Rome Cloud (preview) with no setup needed.

localai-org/kimodo.cpp

LocalAI community ported NVIDIA's Kimodo text-to-motion model to C++/GGML. The model takes a text prompt or a precomputed LLM2Vec embedding and generates 3D body motion in SMPL-X format: joint rotations and root translations. Inside are two separate components: a Llama 3-based text encoder and the Kimodo-SMPLX-RP-v1 diffusion model — both available as ready-to-use GGUF weights from the LocalAI-io organization on HuggingFace, downloaded via a single script. Runs on CPU or Vulkan-compatible GPU; built from source with CMake. A browser-based demo server opens at localhost:8094 — enter prompts and preview animations. The C++ port is Apache-2.0; model weights are non-commercial research-only.

kunchenguid/backpass

backpass is a CLI tool that closes the feedback loop between agent sessions and your AGENTS.md or CLAUDE.md file. It reads local transcript stores from seven agent harnesses — Claude, Codex, Pi, OpenCode, Grok, Cursor CLI, and Hermes — directly from disk, no API keys or uploads required. Every proposed edit comes with verbatim quotes from real sessions as evidence, and a new rule only passes if it appears in at least two independent sessions — one session is not enough. One run proposes at most five edits, described as one gradient step: small, noisy, and safe. Secrets are redacted before they touch any model call. Writing only happens with `backpass apply`, which shows each proposed change with its evidence for you to accept or reject individually. The tool requires Node.js 22.5+ and the acpx utility on your PATH.

ApodexAI/FrontierAgent

FrontierAgent is an open-source agent runtime and terminal TUI by Apodex AI, built for long-horizon tasks: research, file analysis, and document preparation. Two workflow modes: ReAct, where one stateful agent works end-to-end, and Agent Team, where a coordinator breaks the task into parallel assignments, dispatches sub-agents, and synthesizes results. The TUI shows each task's live status, active tool calls, and action history. File-mutating steps show a diff and require approval; sessions are checkpointed and resumable. The filesystem is sandboxed: inputs are read-only, working state lives in /workspace, and persistent deliverables go to /outputs. Connects to any OpenAI-compatible API endpoint — Apodex, OpenAI, or your own local model server. Open source, Apache 2.0.

Tried MiniMax Music3 — at 2.4B parameters it generates full songs up to 5 minutes with vocals, instruments, and proper structure (verse, chorus, bridge, outro). You feed it lyrics with [Verse]/[Chorus] section tags and a music description specifying genre, BPM, instrumentation, and mood — the model builds the whole musical narrative from there. Under the hood it's a hybrid: a global 8B LLM handles overall structure, a local 0.6B LLM handles detail, and a Flow Matching module generates 32 kHz stereo WAV. One thing that gives me pause — there's no listed license, so commercial use is unclear.

Set up MiniMax H3 — a 33B transformer from MiniMaxAI that turns text and images into video with synchronized stereo audio at up to 2K resolution. It splits into three components: context preprocessing, 768p base generation, and 2K upscaling (though H3-Regenerate-2K remains closed-source for now). Running it on image + scene description tasks — gets you 4-15 seconds at 24 FPS. First open-weights model I've seen with native stereo audio baked into the output, no separate audio pipeline required.

Deployed DeepSeek-V4-Flash-0731 via vLLM on a server setup — a 304B MoE model from DeepSeek under MIT license. The 'Flash' comes from DSpark speculative decoding for faster inference while activating far fewer parameters than the total count. Terminal Bench 2.1 score is 82.7 versus 72.1 for DeepSeek-V4-Pro Preview, DeepSWE jumped from 12.8 to 54.4, and AutomationBench doubled from 12.8 to 25.1. Three reasoning_effort levels — low, high, max — let you tune speed versus depth per task.

Unsloth released Dynamic 3.0 GGUF quantizations of Qwen3.8-27B — Alibaba's 27-billion-parameter multimodal model — for local deployment via Ollama and llama.cpp. The model understands both text and images, supports a native 262K-token context window (extendable to 1M via RoPE scaling), and is optimized for agentic workflows and tool calling. Dynamic 3.0 offers 12 quantization levels ranging from IQ1 (~6 GB) all the way to BF16 (~55 GB), letting users choose the best balance of quality and hardware fit. Unsloth claims their approach achieves better accuracy than competing quantizations of the same size by compressing critical weights less aggressively.

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 from ornith-ai, built on top of Qwen3.5 and Gemma4 foundations with continued pretraining and reinforcement learning. Its key distinction is self-improvement training: the system continuously generated new tasks for itself, solved them, and improved without relying on fixed human-curated datasets. The model accepts both text and images as input, supports tool calling, and works inside agentic pipelines such as MCP-based setups. In agentic mode with the OpenHands harness it scores 70.6 on SWE-bench Verified — strong for a 9B-class model. A quantized GGUF version is available on Hugging Face, running in roughly 6 GB at 4-bit precision.

TRELLIS.2 is Microsoft's second-generation 3D generation model, running as a Hugging Face Space with no installation required. It takes an image as input and outputs a 3D model using the Trellis2ImageTo3DPipeline. The space offers six render modes to preview the result: Normal, Clay render, Base color, and three HDRI lighting environments (forest, sunset, courtyard). The first TRELLIS became one of the most-forked 3D generation projects on Hugging Face, with over 100 community spaces built on top of it. TRELLIS.2 continues the series with higher fidelity generation.

Anima-2.9B is a text-to-image diffusion model built on NVIDIA's Cosmos-Predict2-2B architecture and fine-tuned from circlestone-labs/Anima. Gazingstars123 expanded the original 28-layer architecture to 40 layers using interleaved insertion with zeroed-out output projections, reaching 2.9 billion parameters. The model was trained on 1.7 million additional anime and illustration samples with a July 2026 knowledge cutoff. Training used mixed captions from Gemini and Claude — both tag-style and natural language descriptions — without score-based tags. It loads as a single-file diffusion checkpoint (.safetensors) in ComfyUI. License is non-commercial only.

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 self-contained Skill file (structured Markdown with embedded Python) for Claude Code and Codex that aggregates Chinese A-share market data from 19 sources into 54 callable endpoints with no authentication required. The 11-layer architecture covers: real-time quotes and order books via mootdx and Tencent, research reports and PDF downloads from Eastmoney, sector signals and smart-money flows from Tonghuashun, limit-up pools, board attribution, ETF options with Greeks and implied volatility, interactive investor Q&A from cninfo, and PMI macro data from the People's Bank of China and National Bureau of Statistics. Fallback sources are built in for when a primary API is rate-limited. The project has been active for three and a half months and reached over 9 000 stars. The primary README is in Chinese; an English version (README_en.md) is available.

open-gsd/gsd-core

GSD Core is a context-engineering and spec-driven development framework for AI coding agents: Claude Code, Codex, OpenCode, Cursor, Windsurf, Copilot, Antigravity CLI, and more. It addresses context rot — the quality degradation that accumulates silently as an agent fills its context window. The solution is a disciplined five-step phase loop (Discuss, Plan, Execute, Verify, Ship) where all heavy research and execution work runs in isolated subagents, each starting with a clean 200k-token context. Structured files — STATE.md and CONTEXT.md — bridge session boundaries so nothing is lost between conversations. The framework has been available for three months, reached over 8 600 stars, and installs via npm, adding /gsd-new-project and /gsd-onboard commands. Multiple plans can run in parallel execution waves. The project includes a GEMINI.md context file for Antigravity CLI, which is the official successor to Google's Gemini CLI (Google sunset Gemini CLI in June 2026).

kgoedecke/doop

Doop is an open-source alternative to Paper.design — a multiplayer design canvas where humans and AI agents work side by side in real time. Agents connect through the built-in MCP server and stream HTML designs into frames piece by piece, so you watch the design build up live in your browser. Claude Code joins with one command: `claude mcp add --transport http doop http://localhost:4300/mcp`. Each canvas frame is real HTML in a sandboxed iframe; the agent fills it incrementally, reviews a screenshot, and corrects anything that looks off. Multiple people and multiple agents work on the same canvas simultaneously, with live cursors, per-frame editing indicators, and a shared activity feed. A design memory feature lets you pin exemplar frames that get distilled into style rules every agent follows. Self-hosts without configuration: `docker compose up` or `npm run dev` with an embedded Postgres — no external services needed.

furkankly/zoetrope

Zoetrope reads the JSONL transcript that Claude Code writes to ~/.claude/projects/ and renders the session as an animated flow graph. The main agent sits at the top; subagents and workflow groups appear beneath it, with tool calls surfacing as live chips showing the tool name, call count, and elapsed time. You can follow a running session live, replay a finished one at adjustable speed (default 8×), or scrub through the timeline freely. A minimap appears when the graph grows beyond the screen. The same engine is compiled to WebAssembly and available at zoetrope.furkankly.dev — drop a transcript file there and get the same graph in the browser without installing anything. Everything stays on your machine.

anywhere-labs/dsh-desktop

DSH Desktop is an independent community open-source project — not affiliated with DeepSeek — that packages DeepSeek Harness into a native desktop application for Windows x64 and macOS Universal. It bundles the local Web UI, Host service, and plugin system from a fixed upstream version, then layers on top a native window, system tray icon, auto-start on login, and auto-update. No Node.js, no terminal: download the installer from GitHub Releases or dshdesktop.cn, run it, and DeepSeek Harness appears as a tray app with the full DSH plugin ecosystem available. The project composed additional plugins via the official DSH plugin mechanism, so upstream upgrades don't break the extras. It reached nearly 20 000 stars in 11 days.

Qwen3.8-2.4T-A95B is the most capable model in Qwen's open-model family, released publicly at Qwen-Max quality level for the first time. It uses a Mixture-of-Experts architecture with 2.4 trillion total parameters, of which 95 billion are activated per token — organized as 92 layers with 512 experts (10 routed + 1 shared per token). The architecture combines Gated DeltaNet (linear attention) with traditional Gated Attention for efficient long-range reasoning. Native context window is 262,144 tokens, extensible to 1,010,000 tokens. The model excels at coding, research, and long-horizon agentic tasks: Terminal Bench 2.1 — 86.6, SWE-bench Pro — 67.7, GPQA Diamond — 92.6. Text-only; thinking mode is always on.

Ornith-1.5-35B-A3B is a mixture-of-experts model from Ornith Team (Deep Reinforce), built specifically for coding and agentic tasks. It has 35 billion parameters total but activates only around 3 billion per token, making it faster than a dense model of the same scale. On SWE-bench Verified it scores 79%, compared to 73.4% for Qwen3.6-35B and 52% for Gemma 4-31B; on the agentic MCP-Atlas benchmark — 70.2% vs 62.8% for Qwen3.6-35B. Context window is 262,144 tokens, extendable to roughly 1M with RoPE scaling. It supports function calling via XML compatible with the OpenAI API, and chain-of-thought reasoning in <think> blocks. Licensed MIT. A ready-to-run GGUF version is available at 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 from inclusionAI with 7.9B total parameters, activating only 1.3B per token. Its KDA-MLA architecture with a 128-expert sparse MoE (8 routed + 1 shared expert active per token) enables fast inference: 86–90 tokens per second on a MacBook M4 Pro in FP8, using about 8.34 GB of RAM at an 8K context window. The context window extends to 256K tokens with YaRN scaling. A thinking mode (enable_thinking=true, temperature 1.0, top_p 0.95) triggers step-by-step reasoning before producing a response. Available via SGLang (pre-built Docker image), a special vLLM branch, and an open Ollama PR for Apple Silicon.

The Agent Memory Leaderboard from agentmemories.ai is the first public benchmark dedicated to long-term memory systems for AI agents. It evaluates systems across four dimensions: storing and retrieving facts across sessions, handling long multi-session conversations, building user-specific personalized memory, and temporal reasoning (correctly ordering events). There are two tracks: academic open-source methods and industry production APIs. The first public release launched in August 2026 and already includes both research systems and commercial products. Runs as an interactive Hugging Face Space — no installation needed.

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 written in Rust that weighs around 20 MB — no Java, no Python environment, no bundled Chromium. It supports 90+ databases including MySQL, PostgreSQL, SQLite, Redis, MongoDB, DuckDB, ClickHouse, Elasticsearch, SQL Server, Oracle, and many others. The built-in AI SQL assistant accepts plain-language descriptions, writes SQL, explains queries, optimizes them, and runs AI-generated statements through built-in safety checks before execution; it works with Claude, OpenAI, or local models via Ollama. A separate Rust-powered MCP server (`npx @dbx-app/mcp-server`) exposes your DBX connections to Claude Code, Cursor, Windsurf, and other MCP-compatible agents, with per-connection access policies: read-only, safe write, or full access. Available as a desktop app for macOS, Windows, and Linux; a Docker image for team deployments; and a web version. DBX launched in late April 2026 and reached 16,000+ stars in four 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 an installer wizard that assembles a personal AI assistant on top of Claude Code from four separate open-source repos: memory (ai-memory-vault — persists everything Claude learns about you and your tasks in plain text files, no size limit), voice (backtalk — hold a key, speak, get a spoken reply in about one second), face (ai-visualizer — full-screen animations that sync with listen / think / speak states), and optional hands (barehands — gesture control through a webcam, no headset). Every component is optional; the wizard asks what you want before installing. If anything breaks, you describe the problem to Claude and it fixes it using the bundled troubleshooting guides. All components update via git pull without touching your personal data or agent identity. Runs on macOS (Apple), Linux and Windows; on macOS, git is supplied by Apple Command Line Tools, which macOS will offer to install automatically on first use.

ifixai-ai/iFixAi

iFixAi is an independent auditing tool for AI agents that launched in late April 2026 and reached over 11,000 stars in four months. It runs 49 inspections grouped into 5 core pillars: Fabrication (does the agent make up facts or use tools it wasn't granted?), Manipulation (does it escalate privileges, break its own policy, or succumb to prompt injection?), Deception (does it sandbag, pursue side goals, or fail silently?), Unpredictability (does it behave consistently?), and Opacity (does it escalate to humans and handle off-topic correctly?). The result is a letter grade A–F with a per-pillar breakdown. It works with any provider — OpenAI, Anthropic, Gemini, Azure, Bedrock, OpenRouter — or your own agent via HTTP endpoint. Available as a Claude Code plugin, Codex plugin, CLI wizard, or a one-command skill for Cursor, VS Code, Windsurf, Cline, and others. To produce a citable grade, a second independent provider acts as judge — not the same model being tested. Full suite takes under two minutes; estimated cost is ~$12–18 when using claude-sonnet-4.6 as judge (at OpenRouter rates, mid-2026). Apache 2.0 license.

freestylefly/awesome-gpt-image-2

Awesome GPT-Image-2 is a structured prompt library for GPT Image 2, built by independent Chinese developer 苍何 (freestylefly). It contains 532 reverse-engineered cases and 20+ industrial-grade templates across 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. The core idea is 'Prompt as Code': each prompt is split into atomic, composable parts (subject, lighting, materials, layout, visual details) designed for agents, scripts, and automation workflows. The main template documents are written in Chinese; the gallery at gpt-image2.canghe.ai is in English and lets you browse, copy full prompts, and filter by style. An agent skill for Claude Code and Codex is available via npm — it selects styles, templates, and scene tags from the same dataset used by the gallery site. The project launched in late April 2026 and reached 12,500 stars in four months.

FlashML-org/FreeToken

FreeToken is an edge-native inference engine for running large open-weight MoE (Mixture of Experts) models on consumer hardware. Released about a month ago and already at 3300 stars, it comes from the Berkeley/MIT FlashML team — Matei Zaharia (co-creator of Apache Spark and Ray), Song Han and Ion Stoica — backed by a paper on arXiv (2608.16157). The key technique is bandwidth-adaptive CPU–GPU co-execution with a global LRU expert cache and semantic KV cache checkpoints: instead of loading the entire model into VRAM at once, it distributes layers between GPU and host RAM and caches the hot experts. This allows running 290B+ models — DeepSeek-V4-Flash, Qwen3.6-35B-A3B, GLM-5.2 — on NVIDIA RTX 30/40/50 gaming GPUs. The API is compatible with Anthropic and OpenAI, so it connects directly to Claude Code, Codex and OpenCode. Available as a desktop app for Windows and Linux with GUI, plus a CLI via 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 for generating parametric 3D CAD designs. You describe what you need in natural language, attach reference images and key dimensions, and the AI agent builds the geometry in an autonomous loop: it generates candidate designs, runs geometry checks in a real environment, and commits verified designs as new baselines. The result is not an image — it is a fully parametric source file built on the build123d library and the Open CASCADE Technology kernel (the same engine used by FreeCAD and CATIA). Parameters stay editable: change a dimension in the workbench and the geometry regenerates automatically. Export formats: STEP for interoperability with other CAD tools, STL for 3D printing, and color-aware 3MF for multi-material printing. Requires Node.js 20.19+ and Python 3.10–3.13.

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 text-to-speech model from Index Team at Bilibili that performs zero-shot voice cloning: one reference audio clip is enough to capture a voice's tone and timbre. It supports 8 emotion modes (happiness, sadness, anger, and more) controllable via prompt, plus speech rate and phonetic pronunciation control. Supported languages are Chinese, English, Japanese, Spanish, and Arabic. Russian is not supported. The model requires Python 3.10–3.11, a CUDA GPU with approximately 6 GB VRAM, and is installed by cloning the GitHub repository and running uv sync.

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 harness from TrueFoundry — a runtime layer between any LLM and the outside world, released a month ago and already at 3,400 stars. It handles model calls, tool execution, isolated code sandboxing (via Daytona), session management, and human approval checkpoints: the agent pauses before critical actions and waits for your go-ahead, with a full audit log of everything it did. Supports OpenAI, Anthropic, Google Gemini, and any OpenAI-compatible endpoint. Skills — git-backed instruction packs — plug in without touching the core runtime and update live. Three access modes: a chat interface, an HTTP API for embedding into your own products, and embeddable UI components for building custom interfaces.

PrimeIntellect-ai/prime-agent

Prime Agent is an open-source coding and research agent from Prime Intellect, released four months ago with nearly 18,000 stars. It runs in a persistent IPython environment and can execute code, manage files, spawn parallel subagents, and communicate with them directly without routing through the user. Its defining feature is self-improvement: the /refine command reviews past work and rewrites the agent's own supplemental prompts and skills to perform better next time. Recurring workflows can be packaged as reusable skills. A long-running autonomous mode with heartbeat signals lets the agent work in the background for hours. Runs on macOS and Linux; on first launch, choose a subscription or bring your own API key from OpenAI, Anthropic, or another provider.

openai/codex-security

Codex Security is the official OpenAI CLI and TypeScript SDK for AI-powered security scanning. It uses Codex to identify vulnerabilities in your code, validate findings, and propose patches. Adding `--patch --patch-severity high` applies fixes for high and critical findings; `--create-pr` commits the verified files and opens a draft GitHub PR. For monorepos, `scan-components` runs separate scans per component and merges results by root cause. Deep scan mode runs multiple subagents for up to 96 hours. Supports OpenAI, Amazon Bedrock, OpenRouter, and any OpenAI-compatible provider — including Anthropic models via OpenRouter. Released two months ago and reached 10,000+ stars. Requires Node.js 22+ and Python 3.10+, plus a ChatGPT account with Codex Security access 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 translates Claude Code, Codex CLI, Claude Desktop, and Grok Build requests to any LLM provider — 40+ built in, plus any OpenAI-compatible endpoint. Configured through a browser dashboard at localhost:10100. Supports Anthropic, Google Gemini, xAI, Kimi, Azure OpenAI, Ollama, DeepSeek, Tencent Cloud, and more. Model routing uses the provider/model syntax directly in the tool's model picker. Additional features: a ChatGPT/Codex account pool with load balancing and session affinity (active conversations stay on the same account); combos — a virtual model ID that routes across multiple providers with failover or round-robin. OAuth is available for xAI, Anthropic, and Kimi. ocx stop fully restores the original configuration.

cobusgreyling/loop-engineering

Loop Engineering is an open-source collection of patterns, starters, and CLI tools for designing systems that orchestrate AI coding agents — Claude Code, Codex, Cursor, and others. The core idea: instead of prompting the agent each session, you design a 'loop' — a system that prompts the agent on a cadence with persistent state, token budgets, and skill files. The `loop init` CLI scaffolds STATE.md, LOOP.md, budget files, and constraint files in five minutes, then prints a Loop Ready Score. `loop doctor` shows the top-3 next actions. Additional tools include loop-audit (readiness score), loop-cost (token spend estimator), loop-sync (drift detection), loop-context (memory manager), loop-worktree (git worktree isolation), loop-gate (path denylist enforcement), and an MCP server for patterns lookup. Boris Cherny (Head of Claude Code at Anthropic) is quoted: 'I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do.' Released two months ago; reached 10,500+ stars.

Ariescar/anyCreature

anyCreature is an open-source AI pipeline that turns a one-sentence creature description into a game-ready GLB file with rigging, skinning, vertex colours, ambient occlusion bake, and animations. The workflow has five stages (00_START through 04_SHIP), each defined in cards that the executing AI session reads. The engine itself has zero runtime dependencies — one JSON spec in, one skinned GLB out via `node engine/cli.js spec.json out.glb`. Quality is enforced by independent 'blind' reader agents that see only a 24-pixel silhouette and must correctly identify the creature without knowing the original order. Neither gate is self-graded: if the silhouette does not read cleanly from a fresh context, the session revises until it does. Requires Node.js 18+ and Python 3.9+.

antirez/ds4

DwarfStar is a native inference engine built by Salvatore Sanfilippo (creator of Redis) specifically for DeepSeek V4 Flash, DeepSeek V4 PRO, and GLM 5.2 — not a general GGUF loader. The engine supports Apple Metal on Mac (primary), NVIDIA CUDA with multi-GPU, and AMD ROCm. Its 2-bit quantization targets only the routed MoE experts, leaving shared experts, projections, and routing untouched, which preserves quality at minimal model size. A built-in HTTP server supports micro-batching for multi-user setups. Two MacBook M5 Max or M3 Ultra machines can be linked via RDMA for tensor parallelism. SSD streaming is available for machines without enough RAM. Benchmarks in the README show 39 t/s generation on MacBook Pro M5 Max 128 GB. The project is in beta.

Yuan1z0825/nature-skills

Nature Skills is a library of 19 reusable AI research skills for Claude Code, Codex, and compatible agent frameworks (Node.js 18+ required). Skills cover academic writing: polishing text to Nature-journal English, drafting abstracts and introductions, composing response letters to peer reviewers. Paper analysis: a source-anchored Markdown reader with formula and figure rendering, paper-to-Chinese-presentation conversion. Research workflow: multi-source literature discovery with citation verification against Nature and CNS journals, statistical review, data availability statement drafting, patent generation from papers, and reference verification. Skills install globally via npx and persist across all sessions without re-configuring. Each skill encodes a reproducible workflow — not just a prompt template.

virgiliojr94/book-to-skill

Book to Skill converts any technical book, document folder, or file collection into a structured agent skill for Claude Code, GitHub Copilot CLI, Amp, or Codex. Instead of loading the whole book into context, it distills the content into frameworks, decision rules, anti-patterns, and per-chapter files — each chapter is loaded only when you ask about that topic. Benchmark by the authors: 24–51× fewer tokens per question versus dumping the full book at once. Supported input formats: PDF, EPUB, DOCX, MD, HTML, RTF, MOBI. The generated skill includes a SKILL.md index (~4,000 tokens), one file per chapter (~1,000 tokens each), a glossary, a patterns reference, and a quick-reference cheatsheet. Works for more than books: internal architecture docs, brand guidelines, research paper clusters, RFCs, and compliance specs — anything you reference often enough to wish you'd memorized it.

MiniMax-AI/MiniMax-H3

MiniMax's official repository with nine skills for its H3 video model (the one behind Hailuo). The key one, h3-prompt-writing, is a plain Markdown skill with no external API calls, so it works in Claude Code, the Claude Agent SDK, Cursor, Windsurf, Codex, LangChain — any agent that reads SKILL.md. It ships two first-party prompt guides: base-en.txt for text/keyframe modes and ref-en.txt for full-reference (Ref2VA) mode. The other eight skills are style-specific video generators (product ads, 3D animation, stop-motion, brand promos, subtitled clips and more) built for the MiniMax Hub canvas and not portable beyond it. Vendor: MiniMax; the model has an online API.

lencx/Minke

Minke is a native desktop application (macOS, Windows, Linux) for DeepSeek Harness, built by independent developer lencx. It combines agent chat, a file manager with syntax highlighting and diffs, browser, terminal, and plugin manager in a single window — panels stay visible alongside the active conversation without tab switching. Connects to DeepSeek via cloud API or to local models through LM Studio and Ollama, as well as any OpenAI-compatible endpoint. Mobile access works via Tailscale Serve: Minke opens as a PWA on iPhone or Android, letting you continue agent tasks and use the terminal remotely. Application data stays on your machine. Plugins are installed directly from the UI with a built-in GitHub catalog.

langchain-ai/openwiki

OpenWiki is a CLI from LangChain AI that creates and maintains wiki documentation for a codebase or personal knowledge base. An agent reads your source files and builds linked Markdown pages in the openwiki/ folder; each factual statement is tied to specific lines of code through Grounded Claims — when code changes, the wiki knows which pages are stale and what needs to be confirmed or updated. Documentation is output in Google Open Knowledge Format (OKF v0.2), making it portable to any OKF-aware tool. Supports 13 model providers: OpenAI, Anthropic, Gemini (including Google Vertex AI), Amazon Bedrock, and others. Integrates directly into Claude Code and Codex with one command (openwiki integrations install claude) — the agent then handles documentation using your existing model credentials. Nine built-in connectors: Notion, Slack, Gmail, X, Hacker News, LangSmith, and more. Auto-updates via GitHub Actions: code changes trigger a pull request with updated documentation. Released in June 2026 and already at 15k stars.

drumih/turbo-fieldfare

A Swift + Metal inference runtime purpose-built for Gemma 4 26B-A4B, Google's open model. The full model is 14.3 GB, but TurboFieldfare keeps only the shared 1.35 GB core and the FP16 KV cache in memory, streaming experts from SSD as needed — a working budget of about 2 GB, so it runs even on 8 GB Macs. Not a wrapper around MLX or llama.cpp but a model-specific engine; the repo documents 103 measured experiments across kernels, caching, I/O and decode. Ships a native Mac app, a CLI, a streaming model installer and a local OpenAI-compatible server. Requires macOS 26+.

BigPizzaV3/CodexPlusPlus

Codex++ is an external launcher and manager for the OpenAI Codex desktop app, built with Rust 1.85+ and Tauri 2.x. It attaches to the official app via Chromium DevTools Protocol without modifying app.asar or writing patch files to the installation directory. The tool offers four provider modes: official account only, mixed (account + API), pure API (fully custom Base URL and key), and aggregated provider with failover and load balancing across multiple services — each mode stores credentials separately. Session management covers local session scanning, bulk deletion, Markdown export, and token usage history. UI enhancements include plugin marketplace access, Chinese interface localization, conversation width and scroll-position restore, Stepwise next-step suggestions with configurable model, Upstream worktree creation, Zed Remote integration, and user script support. Distributed as signed installers for Windows x64, macOS Intel (x64), and macOS Apple Silicon (arm64). README is primarily in Chinese; an English README is available.

ayghri/i-have-adhd

i-have-adhd is a skill for Claude Code and Codex that restructures how your agent formats responses: action comes first, explanation follows. Ten rules govern every reply: start with the concrete next step or command, number multi-step tasks, end with one clear next action, suppress tangents, restate context each turn. Banned: preambles ('Great question!'), recaps, closers ('Hope this helps!'). Lists capped at five items. Time stated in minutes, not 'soon' or 'briefly'. Mistakes are reported matter-of-factly, without apology. The skill is fully customizable: fork the repo, edit SKILL.md, install your version in place of the upstream one. Released in May 2026 and already at 23k stars — because the problem is real.

alibaba/open-code-review

Open Code Review (OCR) is Alibaba Group's internal AI-powered code review CLI, open-sourced after two years of internal use across tens of thousands of developers. It reads Git diffs, sends changed files to a configurable LLM via an agent with tool-use capabilities, and returns structured review comments with line-level precision. The architecture is hybrid: deterministic engineering handles file selection, grouping, and rule matching, while the LLM agent focuses on dynamic context and comment generation. Benchmarked on 50 popular open-source repositories, 200 real pull requests, and 10 programming languages: OCR consumes about 1/9 the tokens of general-purpose agents like Claude Code at higher precision, while recall is deliberately lower to reduce noise. Supports plugins for Claude Code, Codex, and Cursor, plus CI/CD integration with GitHub Actions and GitLab CI. Compatible with OpenAI, Anthropic, and any custom provider.

Spielewoy/autoprompt-skill

Autoprompt is a skill for coding agents that according to the independent Terminal-Bench 2.1 benchmark improves scores by +14.61 points, corresponding to a 45% reduction in task failures. The benchmark leaderboard includes DeepSeek V4 Flash 0731 as one of the reference models. An interactive installer (npm install -g autoprompt-skill && autoprompt) detects your coding agent and installs the skill. Supports Claude Code, Codex, OpenCode, Kilo Code, VS Code Copilot, and Prime Agent.

pathwaycom/arc-task-gen

ARC Task Gen from Pathway generates original ARC-AGI-1-style reasoning tasks distribution-matched to the public evaluation set. The repository accompanies the paper introducing BDH-CQ, a new architecture that combines in-context learning via evolving recurrent memory with iterative reasoning in a continuous latent space. Key result: a 150M-parameter BDH-CQ model achieves 29.5% pass@2 on the ARC-AGI-1 evaluation set at $0.0007 per task — 11× cheaper than GPT-5.6 Luna Low. Generated tasks follow the standard ARC JSON format and work with existing evaluation harnesses.

odysseus-dev/odysseus

Odysseus is an open-source, self-hosted AI workspace that brings together chat with any language model, AI agents with tool support and MCP integration, deep web research, a document editor, email via IMAP/SMTP with AI triage, notes, tasks, and a calendar. The Cookbook module recommends models based on your hardware; the Compare view puts two model responses side by side. Data stays on your machine. License is AGPL-3.0.

nexu-io/open-design

OpenDesign — an open-source alternative to Claude Design, Anthropic's closed design product — is an open-source native desktop app for macOS and Windows, released in late April 2026 and reaching nearly 90 000 GitHub stars in four months. It acts as a design studio for AI coding agents: you describe the artifact, the agent builds it live inside the app with a built-in preview, and you iterate in plain language until it is ready. Supported output types include web and mobile prototypes, live dashboards, multi-slide decks, images, HyperFrame motion graphics and video. Export formats: HTML, PDF, PPTX, MP4. Supported agents: Claude Code, OpenAI Codex, Cursor, OpenCode, GitHub Copilot, DeepSeek Harness, Kiro, Kimi, Hermes and 20+ others — or any OpenAI-compatible endpoint via BYOK (bring your own key). A DESIGN.md design system lets you define brand colours, typography and components once, and every artifact the agent produces follows them automatically.

missuo/herdrm

herdrm is a native macOS app that works as a console for herdr.dev — a background runtime that keeps coding agent terminals running. The app aggregates all your local and remote herdr instances in one sidebar: Claude, Codex, Gemini, Grok, OpenCode and others, with live status (blocked agents bubble up, working ones spin, done get a check). Select an agent and a live terminal attaches directly to its PTY. Sends system notifications when any agent on any device finishes or needs input.

JustVugg/colibri

Colibrì is a pure C inference engine (zero engine dependencies, single-file core) for frontier MoE models on consumer hardware. It unifies storage, RAM, and VRAM into a single memory hierarchy: the dense shared core (~17B params, ~9.9 GB at int4) stays resident in RAM, while 19,456 routed experts (~19 MB each at int4, ~370 GB total on disk) are streamed on demand with a per-layer LRU cache. Supported model families: Kimi K3 (2.8T) by Moonshot AI, GLM-5.2 (744B), Inkling (975B), DeepSeek V4 Flash (284B), Qwen3.6 (35B-A3B), OLMoE (7B). The built-in OpenAI-compatible server lets any client that speaks the OpenAI API — including Claude Code from Anthropic — use a locally running model instead. Backends: CPU, CUDA, Metal, NUMA. Includes a web dashboard with real-time metrics and a live visualization of all 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

An Electron desktop app from Chinese company 赋范空间 (BeyonData) that wraps DeepSeek Harness in a graphical interface. Includes a plugin discovery center with AI-powered recommendations, one-click install and management, a Preset Square with official and community agent presets, an Application Center for standalone AI apps, and visual workspace management. The local dsh web service runs inside the app; data stays on your machine. macOS arm64 and Windows x64, open source, alpha stage.

CopilotKit/openbot

OpenBot is an open-source agent platform from CopilotKit that gives each AI agent its own isolated computer: a real browser with its own logins, its own files, and only the tools you grant. Built on the AG-UI open protocol, so agents from LangGraph, Mastra, CrewAI, or any hand-written endpoint connect the same way. Every tool call goes through a gateway that decides, records, and only then acts — or refuses and names the rule. Still in alpha and actively developed.

browser-use/macos-harness

From the browser-use team: a minimal Python process that gives an LLM six raw primitives for controlling macOS — window screenshots, keyboard/mouse input directed to an app PID, Apple Accessibility tree, AppleScript, a real browser via CDP, and the local filesystem. The model writes any missing logic mid-task; no app-specific tools are needed in advance. macOS only, MIT licensed, anonymous opt-out telemetry.

bowenliang123/dsh-context

A DeepSeek Harness plugin that adds a Context tab and a /context slash command to any session. The tab shows a six-color stacked bar of what is currently in the model's context window — system prompt, tool schemas, your messages, injected context, assistant replies, tool results — scaled against the model's total window. History view shows one bar per model request so you can watch the window grow and see exactly when compaction or pruning happened.

bawadou/ai-data-extractor

A free open-source Python tool that auto-discovers and extracts complete conversation history from 10 AI coding assistants: Claude Code, Codex CLI, Cursor, Windsurf, Continue, Gemini CLI, OpenCode, Trae, Cline/Roo Code, and Aider. Extracts messages, code context, diffs, tool calls, timestamps and session metadata into a normalized JSONL format. Supports macOS, Linux and Windows. Standard library only — no dependencies to install.

vibeinging/deepseek-harness-desktop-app

DeepSeek Harness Desktop App is a local AI desktop workspace built on top of dsh, the official DeepSeek AI agent runtime. It integrates DSH sessions, project management, a file manager, a built-in multi-tab browser, Git Worktree isolation for parallel AI-assisted coding branches, Canvas for document editing with version history, and Office artifact creation (DOCX, XLSX, PPTX, PDF) in a single desktop window. The interface is primarily in Chinese with an English localization available. Currently verified on macOS Apple Silicon; Windows x64 build is in the pipeline.

SigmanticAI/apex-inference-chip

APEX is an open-source hardware tile (RTL) for a full transformer decoder layer, built by Sigmantic AI using their autonomous verification agents. It runs Qwen2.5-0.5B on real FPGA hardware at 0.56 tok/s — a 140× improvement from the initial bring-up. Every hardware block is bit-exact verified against a NumPy golden model before it ships.

myYangyunfan/dsh_desktop

dsh_desktop is a Windows and macOS desktop client that bundles the full DeepSeek Harness (dsh) runtime, including official plugins, inside a single executable. A built-in Node.js runtime means users just download and double-click—no terminal, no npm, no Node.js installation required. The bundle also includes dsh-openclaw-bridge, a plugin for connecting dsh sessions to Tencent's WeChat and QQ messaging platforms via OpenClaw and Tencent's official bot APIs. Extras include a balance widget showing real-time cost per session, Windows system-tray notifications when agent tasks complete, a desktop companion pet (a small whale), and a crash-recovery watchdog. The interface is in Chinese. Available as a portable exe and an installer version.

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 (version 0.1.2) you run locally alongside your AI app. It connects your AI agent to real online stores—Shopify, WooCommerce, eBay, Etsy, Amazon—and returns a ranked shortlist with machine-readable reasons for each recommendation. Seller payments are never an input to the ranking. Every response includes rejected offers and why they were ruled out, merchant-trust evidence, and a signed single-use purchase mandate that must be approved before any checkout can proceed.

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, locally-run MCP server for shopping agents. It connects to stores including Shopify, WooCommerce, eBay, Etsy, and Amazon, ranks candidates using only the criteria in your brief — price, delivery, features, merchant trust — and labels sponsored offers below organic results. The ranking logic is open and deterministic: every recommendation includes machine-readable reasons, and a local audit trail records each decision.

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 platform for team collaboration where AI agents participate as full teammates — they hold personas and memory, claim tasks from a shared Kanban board, coordinate with each other without collisions, and can send and receive real email. Two brain modes are available: Cumora Cloud (agents run in managed per-agent pods on the OpenAI Responses API) or BYOA — you connect your own Mac or VPS with your local Claude Code or Codex CLI, so the server never sees your provider keys.

Hisn00w/ASu-skills

ASu-skills is a plugin pack for Codex (OpenAI's coding agent) and DeepSeek Harness aimed at the Chinese campus job-recruitment cycle. It ships five skills: one that finds active open-source projects and submits small real pull requests to build a portfolio, one that rewrites your work experience and projects for a target role, one that generates an editable HTML résumé with PDF export, one that replicates the author's own high-density technical résumé layout, and one that turns emails and screenshots into a structured job-application tracker. The toolkit reached nearly 1,000 stars in five days during China's autumn recruitment season.

dsh-market/dsh-market

dsh-market is itself a DeepSeek Harness plugin that installs an App Store-style interface inside DSH Settings. Once active, it lets you browse and search the curated community catalog of 800+ plugins with category filters, star counts, and AppStore-style screenshots. Themes and skins get their own tab. It also handles updates, backup and restore of your plugin list as JSON, and hot-enable or disable of plugins without restarting. Installs are restricted to the curated awesome-dsh-plugin registry so random packages can't slip through.

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 integrates 10 visual tools into DeepSeek Harness for text-only models: image Q&A, pixel-level object grounding, object detection, cropping, tracing, pixel diff between screenshots, long-screenshot OCR, foreground extraction, dominant color analysis, and HTML screenshots. The toolkit uses Gemini 3.7 Flash as a built-in free vision engine (300 images per machine per day, no registration), with an option to bring your own OpenAI or Anthropic-compatible model.

xiaobright/dsh-anchored-standard

Anchored Standard is a community project — not affiliated with or endorsed by DeepSeek — for the DeepSeek Harness agent framework. It addresses a specific behavioral issue: DeepSeek V4 Pro conditions strongly on the tool catalog it sees on its first request. When started with the full Standard catalog, the model tends to add hedging phrases and indirect replies; starting with the Minimal preset's exact two-tool schema anchors the first response to direct, task-focused output. After the first tool call, the preset automatically exposes the full Standard catalog. In the Project2 evaluation, this approach scored 98–99 versus 91–92 for the Standard preset.

vercel-labs/eve-software-factory-template

Vercel Labs released Foreman, an open-source AI software factory template built on the eve framework. It connects to GitHub and Linear and runs every labelled task through a four-stage agent pipeline: a Classifier agent triages the request and asks clarifying questions if the task is ambiguous; an Analyst builds an implementation plan with acceptance criteria from a live checkout of the repository; an Implementer executes the plan in its own isolated sandbox, runs the project's own checks, and pushes a branch; a Reviewer independently reads only the pushed diff (not the Implementer's reasoning) and gives a verdict with evidence. Between runs, Foreman maintains a persistent factory brain — notes about the repository that every agent starts from. Deploy is a one-click Vercel button that automatically sets up GitHub and Linear connectors.

elie222/rakazo

Rakazo is an open-source alternative to Grok Bot, built by Inbox Zero Inc. (the team behind the Inbox Zero email productivity app). Each bot runs with one persistent thread, a sandboxed computer environment (Docker locally or E2B for remote deployments), long-term memory, scheduled routines, and the ability to spawn peer bots or short-lived subagents within a single turn. The full product runs without any Rakazo-operated cloud service — it is self-contained. Model selection uses a Pi catalog: the current release supports OpenRouter API keys, ChatGPT Plus/Pro via device code sign-in, GitHub Copilot, and SuperGrok; Claude Pro support is described as not yet available in the UI. Available on web, Electron desktop, and Expo mobile. The project was built using Cursor with Grok 4.6.

ccch1mneyyy/dsh-TUI

dsh-TUI is a community-built plugin for DeepSeek Harness — DeepSeek's open-source AI coding agent. It adds a full-screen terminal interface described by its author as Claude Code-style: pixel-whale header bar, live agent status line, streaming thought display, context progress bar with TPS meter, and a double-Esc session time-rewind feature. The DeepSeek Harness official WeChat channel featured it as a top community pick, which drove rapid star growth.

anywhere-labs/deepseek-harness-desktop

Anywhere Labs released a macOS and Windows desktop wrapper for the official DeepSeek Harness agent framework, which DeepSeek open-sourced on August 13, 2026 under MIT license. The app handles launching and managing the local Harness service automatically, adds system tray integration, and packages everything into a downloadable installer — no Node.js installation or command-line setup required. The main readme and interface are in Chinese. Upcoming features include a plugin marketplace, mobile remote control via iOS and Android, and integrations with messaging apps such as WeChat, Feishu, and Discord.

alchaincyf/deepseek-harness-orange-book

HuaShu (alias @AlchainHust), an AI-native independent developer who has never hand-written a line of code, published this Chinese-language guide to DeepSeek Harness within 24 hours of DeepSeek open-sourcing the framework under MIT license on August 13, 2026. The book records what the author actually saw running it on his own machine: the complete system prompt injected before each agent turn, the 129-line default boot configuration dump, three unedited session logs (standard mode, PTC, and creation mode), a step-by-step recording of the AI autonomously adding a new tool to its own toolchain in 19 steps (tool count going from 32 to 33), real cost accounting for a PTC session, and a list of 35 optional packages present in the repository but not installed by default. The readme contains a brief English summary; the main book content is in Chinese only.

zhu1090093659/dsh-web-ui

A BSD-3-Clause plugin and skin collection for DeepSeek Harness (an AI coding assistant by DeepSeek AI), developed by independent contributor zhu1090093659. Adds a kanban task board with cron scheduling, a git graph, a right-side panel with file tree and diff view, live token and TPS stats, mobile remote UI via QR pairing, SSH remote management, and 8 themes including Windows XP Luna and a Trading Terminal skin with live ticker.

Leutenegger/book-to-skill

An MIT-licensed Python CLI by developer Leutenegger that converts PDF, EPUB, DOCX, Markdown, HTML, RTF, and MOBI files into a structured agent skill — generating a SKILL.md index, per-chapter files loaded on demand, a glossary, patterns reference, and cheat-sheets. The author measured 24×–51× fewer tokens consumed compared to dumping the whole book into context. All processing is local; files are never uploaded.

i3T4AN/KADATH

KADATH is an open-source evolutionary multi-agent runtime by individual developer i3t4an, built on top of Hugging Face smolagents. It treats AI agent improvement as a biological selection problem: a population of agents attempts your goal across multiple epochs, each is independently graded against a locked benchmark, and then the kernel applies Darwinian selection — top 30% are preserved unchanged, the middle cohort mutates based on what worked for the elite, and the lowest-ranked are culled. In a published ten-epoch run, the best fitness score rose from 18 to 91, reflecting improvement across the competitive population rather than hand-tuning a single agent.

deepseek-ai/deepseek-harness

DeepSeek AI, the team behind the r1 and v3 models, published an open-source agent harness called DeepSeek Harness (dsh). The system is built around a single rule: everything — model connections, tools, the Web UI — is a separate plugin that can be installed or removed without touching the rest. Under the hood runs Cordis, a framework built for hot-swappable, composable components. The project is currently in developer preview and starts with a single npx command.

milind-soni/OpenMausBot

OpenMausBot is an open-source macOS desktop app by Milind Soni, inspired by Grok Bot (xAI) but fully independent and not affiliated with xAI. In a Telegram-style sidebar, each contact is a real AI agent — Claude, Codex, or Grok — running locally on your Mac using your existing credentials. Agents can get a cloud Linux desktop they drive autonomously, connect to 500+ apps (Gmail, Slack, GitHub, Notion) via Composio, and surface every shell command or file edit as an inline approval card before acting.

dmmulroy/anti-slop

An open-source Oxlint plugin by developer Dillon Mulroy with 15 rules that specifically target TypeScript and JavaScript patterns AI coding agents produce most often. Rules like no-module-mocking, no-unknown-returns, and no-chained-type-assertions catch the kind of shortcuts agents take when they write 'working code' instead of type-safe, evidence-based code. The project is designed to be vendored — copied directly into your repository — rather than installed as a fixed npm dependency, so the rules belong to the project and can be adapted.

0xnyn/airship

Airship is an open-source CLI (airship.design) that wraps your running dev server in a visual design canvas without touching your project files, build config, or dependencies. You click any element on the page, and Airship reads the exact file and line that rendered it from the dev build — then hands that context, your instruction, and the relevant frames to Claude Code, Codex, or OpenCode, which edits the source directly. Desktop and mobile frames sit side by side, both live and editable at the same time.

UditAkhourii/neuroarxiv

An open-source Claude Code skill by independent developer Udit Akhouri. When you invoke /neuroarxiv with a problem, it maps the problem onto arXiv categories, fetches real papers over HTTP from export.arxiv.org, reads each paper in strict isolation — one LLM call per paper so they can't anchor each other — scores them, and converges to a single cited recommendation. The author's published eval showed it catching a withdrawn proof in its own source list and declining to rely on it, which neither plain search nor cold reasoning produced.

tanishqkancharla/calldiff

A CLI tool that diffs function call stacks across git commits — showing which callees appeared, disappeared, or moved under an entry point. Built with Tree-sitter (AST-based) and supports 22 languages. Designed for agentic code review, when line diffs bury the structural shape of a change. Can be registered as an MCP server so AI agents use it directly, and outputs JSON for machine consumption.

shadcn-ui/chatbot-template

shadcn-ui is the team behind one of the most widely deployed UI component libraries. Their new chatbot starter is built on the Vercel AI SDK and ships with streaming chat, per-provider web search, and an interactive questionnaire that lets the model ask the user clarifying questions mid-conversation. It deploys to Vercel with one button click — no manual API key setup needed when running on Vercel, which handles authentication automatically.

gvzdv/claudish-to-english

A Claude Code plugin that rewrites each assistant response in plain English on screen, using a local LLM via Ollama. The original text and saved transcript are unchanged — only the on-screen display is rewritten. Fails open: if Ollama is not running or times out, the original Claude response shows normally. An optional second hook can also rewrite Markdown files written by Claude.

antirez/h3.c

h3-metal is a native C inference engine for Apple Silicon that runs MiniMax H3 — a 33-billion-parameter video generation model from Chinese AI startup MiniMax — entirely on-device via Metal, Apple's GPU framework. Built by Salvatore Sanfilippo (antirez), the creator of Redis, it supports text-to-video generation, first/last-frame conditioning, and reference-based video generation, all without 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

An open-source MCP server that connects AI agents to a commercial SMS-verification service. The agent can order a private number in any country, poll until the code arrives, finish or cancel the activation — nine tools over stdio, with no SDK required. The service behind it is commercial; free tier is not available.

SMNETSTUDIO/WeChat-AI

An open-source, self-hosted platform by SMNETSTUDIO for deploying LLM-powered roleplay chatbots inside WeChat via Tencent's iLink API. About a third of the documentation and interface is in Chinese. Users assign AI personas with custom personalities and conversation memory; admins manage bots, token usage, and moderation through a dashboard. A visual Chatflow editor lets you wire up conversation logic without writing code. Supports multi-node deployment with Cloudflare Worker load balancing.

oil-oil/codex-deepseek-subagent

A Codex skill installable via npx, with a Chinese-language README. It registers DeepSeek v4 Flash as a native custom subagent in the OpenAI Codex desktop app. After setup, the main agent dispatches tasks to DeepSeek directly; the skill manages routing, secure credential storage, and automated acceptance testing.

NikoDemon80/ComfyUI-H3-Motion-Context

A ComfyUI custom node pack that enables clip chaining for the MiniMax H3 AI video generation model. H3 generates picture and audio together but lacked built-in chaining; this plugin feeds the tail latent of the previous clip directly into the next one, keeping motion speed, direction, and audio continuous across the join. The patches verify their own math against the live ComfyUI code on each startup and refuse to run if an update breaks an assumption.

mewamew/my_ai_town

A single-player life-simulation game built in Godot 4.7 with a Chinese-language README. Each resident is an LLM-powered agent that moves between locations, remembers past interactions, maintains relationships, and makes autonomous decisions. The player can enter the town as an avatar, post announcements, or change the weather and watch how residents respond.

Sateezg/codex-bridge

codex-bridge is a community-built Claude Code plugin that bridges Claude to the Codex CLI. It adds two capabilities: generating and editing images with gpt-image-2 (favicons, OG cards, mockups, icon sets), and delegating bulk or repetitive tasks to GPT-5 subagents. Both run on the user's ChatGPT subscription quota, so Claude's own context and token budget stays free for higher-level decisions.

Larryvrh/ComfyUI-MiniMax-H3-Turbo

ComfyUI-MiniMax-H3-Turbo provides two ComfyUI nodes — MiniMax-H3 Turbo LoRA and MiniMax-H3 Turbo Sampler — that apply a community-trained Turbo LoRA (published by the author on HuggingFace as larryvrh/MiniMax-H3-Turbo-Lora) to the MiniMax H3 base model from Chinese company MiniMax. Generation steps drop from roughly 20 to 4–8 for both text-to-video and image-to-video. The plugin works with all H3 checkpoint variants (full bf16, int8, pruned) and installs via ComfyUI-Manager.

huangserva/ComfyUI_MiniMaxH3_Director

A set of five verified ComfyUI workflow JSON files for the MiniMax H3 Director model — an AI video generation model from Chinese company MiniMax that produces video and synchronized audio together. The workflows cover text-to-video (T2V), first-last-frame and image-to-video (FL2V/I2V), reference-based video generation (R2V), timeline-based video editing (V2V), and reference-guided character replacement in video (RV2V). The repository is a convenience mirror of AIMixer/ComfyUI_MiniMaxH3_Director, verified on an NVIDIA RTX 4090 with ComfyUI 0.30.0.

eternityspring/shuohao-skills

shuohao-skills is a self-contained skill collection for Claude Code and Codex. The current skill, novel-characters, reads a novel and produces a character bible: written profiles, cartoon-design image prompts, voice style prompts, and three-view turnaround design sheets for each character. An install script detects whether Claude Code or Codex is present and symlinks the skills so a git pull updates them instantly, with no reinstall required.

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

This ComfyUI plugin adds 14 custom nodes that properly handle audio and video together in MiniMax H3 generation workflows. It supports text-to-video-with-audio, image-to-video-with-audio, first-and-last-frame video, and reference-based generation, maintaining correct media ordering and audio-video contracts that the base ComfyUI MiniMax H3 integration doesn't cover. It requires no additional pip packages — everything runs on existing 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

qiaomu-seo is an agent skill created by Qiao Mu (向阳乔木), a Chinese AI product creator. It installs into Claude Code, Codex, Cursor, or any compatible agent client with one command and gives the agent full professional SEO capability: technical audits, keyword-to-page mapping, crawlability diagnosis, site migrations, JSON-LD and canonical fixes, and — notably — separate checks for AI search surfaces including Google AI Overviews, ChatGPT Search, Microsoft Copilot, and Perplexity.

google-gemma/gemma-translator

Google Creative Lab (the experimental arm of Google) built an open-source voice translator that runs Google's Gemma 4 model entirely locally using the LiteRT-LM runtime — no internet connection needed after setup. The device runs on a Raspberry Pi 5 with a small touchscreen: it captures speech, transcribes it, translates with Gemma 4, and speaks the result aloud in the other person's language. The full source code, deployment scripts, and even 3D-print files for the case are included.

Accio-org/RealReplicaBench

RealReplicaBench, built by the Accio team at Alibaba International, tests AI agents against 107 real business workflows — publishing product listings, booking freight, editing spreadsheets, and calling APIs — all running in stateful local replicas of real commerce and SaaS platforms. Tasks span CLI, browser, file editing, and API/MCP interactions, each graded by its own verifier. The live leaderboard puts Claude Opus 5 at the top, completing 56–62% of tasks depending on the evaluation harness.

ZzzLc0405/photo-abstract-editorial

A Codex skill — a structured SKILL.md prompt file — that instructs Codex to turn any uploaded photograph into a vertical editorial composition. The result has three parts: the original photo in the main area, an abstract panel derived solely from the spatial relationships and colors of the original (not a filter or repaint), and a single poetic English title. The repository is written primarily in Chinese, with English prompt versions included in the references folder. It collected over 1,200 stars in its first week — one of the fastest-climbing Codex skill repositories.

SandAI-org/MAGI-2-preview

MAGI-2 Preview is a unified audio-video generation model from Sand.ai with 114 billion parameters that activates only 6 billion per token, using a Mixture-of-Experts architecture called MagiMoE. It generates ten-second video clips from a text prompt (T2V) or a text prompt plus a reference image (I2V), with audio produced and mixed in the same pass. Inference runs in two stages: a preview denoising pass followed by an upscale to 1080p. Weights (307 GB total) are published on Hugging Face under Apache 2.0.

criptogus/HermesOffice

HermesOffice is an open-source office suite from Genspark (Mainfunc, Inc.) for macOS, Windows, and Linux: word processor, spreadsheets, presentations, and PDF editor in a single package. It started as a fork of GenOffice (Apache-2.0) and adds the Hermes Agent AI layer on top. Every app embeds the same AI panel — the agent operates at the block level, understanding document structure and patching only the changed blocks while leaving everything else byte-for-byte intact.

AMAP-ML/LongHorizon-Harness

LongHorizon-Harness is an open-source execution, state-management, and verification system from the AMAP-ML team that runs on top of Claude Code and Codex CLI without replacing them. It splits work across three roles: a Manager that holds the original goal and verified progress, an Executor that starts each step with a fresh context, and an Auditor that independently checks results before they are recorded as done. On WeaveBench, task completion rises from roughly 50% to 80%; on OSWorld 2.0, full desktop task completion triples. The tool reached the top of Hugging Face Daily Papers weekly ranking on August 6, 2026.

thebuggeddev/anatomy

An open-source interactive 3D human anatomy explorer built with Three.js and GPT 5.6 Sol, running on OpenAI's hosting platform (vinext/Cloudflare). The project lets users explore a three-dimensional body model and query GPT directly within the interface. It collected nearly 1,800 stars in its first five days — unusually fast for an anatomy visualization tool.

fuxicodex/Fuxi

FuXi is an AI developer terminal from fuxicodex — a coding agent with a rich terminal interface that reads your codebase, edits files, runs shell commands, and routes every request across multiple LLM providers with cost-aware routing and automatic failover. Written in Go, it ships as a single static binary requiring no runtime dependencies, and supports MCP servers, hooks, skills, and user-defined slash commands.

mikiarlo3/awesome-growth-hacking-skills

A curated open-source directory of AI agent skills for growth hacking and marketing, maintained by enso.bot. It covers SEO and AI search optimization, content creation, email marketing, paid media, sales outreach, CRO, social media, and go-to-market workflows — all designed to run with AI agents like Claude Code and Cursor. Gained over 400 stars within 24 hours of launch.

magicrew/doc7

doc7 converts PDFs, Word and PowerPoint files, scans, screenshots, charts, diagrams, and formulas into Markdown through any OpenAI-compatible vision model, including local deployments via LM Studio or Ollama. Unlike classic OCR tools, it treats each page as an image and asks the model to recover the full semantic structure — tables, equations, figure captions, and cross-references — rather than just extracting character sequences. In a benchmark on raster-only pages from 'Attention Is All You Need', doc7 recovered 15 of 15 checked facts, while MarkItDown with its OCR plugin recovered 9 and Docling recovered 3. It is attracting attention because document conversion is a persistent bottleneck for anyone building RAG or agent pipelines around complex real-world files.

leonickson1/Swiftlet

Swiftlet is a Swift and Metal inference runtime for the Qwen3-Next and Qwen3.5/3.6 Mixture-of-Experts model family, released by Alibaba's Qwen team under Apache 2.0. It keeps only the small dense core of the model in RAM and streams the sparse expert weights from local storage as each token is generated, which lets a 35B model run in 2.6 GB of RAM on an M5 Mac at 7–11 tokens per second. The same model runs on an iPhone 17 in about 2.5 GB of RAM at roughly 1 token per second. It drew attention because very few open runtimes stream full-scale MoE weights from disk rather than loading the entire model into memory.

Binaryify/open-kimi-ppt-skill

An unofficial Kimi Slides Skill built by developer Binaryify through reverse-engineering of Kimi's presentation platform — not affiliated with Kimi or Moonshot AI. It lets AI coding agents generate complete presentations: the output is always two files — an editable PPTD project (viewable and tweakable in a local browser editor) and a finished PPTX with embedded fonts and fade transitions. Agents can also convert an existing PPTX into a PPTD for further editing. The skill works with any agent that supports the SKILL.md spec: Codex, Claude Code, Cursor, WorkBuddy.

Accio-Lab/RealReplicaBench

RealReplicaBench, built by the Accio team at Alibaba International, evaluates whether an AI agent can complete long-horizon business workflows rather than just answer questions about them. The 107 tasks span browser interactions, CLI operations, file and spreadsheet editing, and API/MCP calls, all run in fresh containers against stateful local replicas of real commerce and SaaS software. Every task has its own deterministic or LLM-assisted verifier, and all results are published on a live public leaderboard. It is gaining attention because it offers a reproducible, auditable way to compare agents on realistic work rather than toy examples.

sophiamyang/finger-frame-effect-ai

A web app that takes a video of the two-hand finger-frame gesture and places an AI-generated world inside the frame window. The uploaded video is sent to Google's Gemini Omni Flash video model with a chosen style — anime, claymation, watercolor, 3D animated movie, or a custom prompt — and the AI regenerates the entire clip so the animated version moves exactly like the original. MediaPipe Hand Landmarker tracks both hands per frame and composites the AI video through the tracked quad. The app runs in the browser; the only requirement is a Gemini API key.

MarcosSete/awesome-free-ai-course-notes

A strictly curated index of written ML and AI course notes published directly by professors and departments at institutions including MIT, Harvard, Princeton, Stanford, UC Berkeley, Caltech, Cornell, Oxford, LMU Munich, KAIST, and USP Brazil. The list enforces a high bar: only written prose notes that read like a textbook chapter — not slide decks, not video lectures. Every entry is official, permanently free, and accessible without login. This is a companion project to the same author's earlier 'Awesome Free AI Books' repository, which was featured here in late July.

firecrawl/anydoc

Firecrawl — the team behind AI web-scraping infrastructure — released anydoc: an open-source Rust library that converts Word, PowerPoint, Excel, PDF and eight more office formats to clean, structured Markdown in under 5 ms. It ships as an Agent Skill for Claude Code, Codex, Cursor, and OpenCode, teaching AI agents to automatically open any document they encounter.

elayadesign/ai-design-skills

A library of skill files — structured markdown rule sets — that you drop into Claude Code, Cursor, Codex, or Windsurf to expand what the AI agent can build. The first skill covers building a landing page end-to-end: the agent asks intake questions, assembles the page structure, writes conversion copy, and sets up a full design system covering typography, spacing, radius, and motion. For users without a terminal, the skill file can be attached directly to a Claude.ai Project or chat. The repository gained 358 stars in its first few days.

disler/super-simple-software-factory

A Python-orchestrated framework for building repeatable multi-agent coding workflows. Python owns the sequencing, retries, and acceptance criteria; AI coding agents operate only within named, bounded phases. Typed JSON envelopes carry context across phase boundaries, and every event streams into SQLite for live observability — so you can watch a run mid-flight rather than read a transcript after the fact. The framework is packaged as a Claude Code skill: install the .claude/skills/sssf folder, type /sssf install, and it stamps the factory into any repository. The starter roster supports models from OpenRouter, Fireworks, and OpenAI simultaneously.

aigclink/geolook

A self-hosted open-source platform for Generative Engine Optimization (GEO) — the practice of getting AI search engines (ChatGPT, Perplexity, Gemini, DeepSeek, and others) to proactively mention and cite your brand when answering user questions. GeoLook covers the full loop: sampling what each AI engine says about your category across 15 engines, diagnosing why your brand is absent or ranked low, generating implementation tickets with acceptance criteria, and verifying programmatically whether fixes actually improved your mention rate. All data is stored locally under a work/ directory; nothing is sent to a vendor cloud. The platform also supports Chinese AI engines — DeepSeek, Kimi, Doubao, MiniMax — and measures CN and global markets separately.

openai/ten-proofs

Lean 4 formal proofs of ten mathematical and theoretical computer science results published by OpenAI: sphere packing bounds, binary and spherical codes, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, closest vector problem, Ehrhart's volume conjecture, multicolor Ramsey numbers, and extremal graph theory. Each is a machine-checkable certificate, not a sketch.

genspark-ai/genoffice

GenOffice is an AI-native office suite for macOS and Windows developed by Genspark (Mainfunc, Inc.), covering DOCX word processing, XLSX spreadsheets, PPTX presentations, and PDF editing in five Electron apps. Unlike suites that bolt a chat box onto the side, AI editing is built in at the block level: the model rewrites only the paragraphs, slides, or cells you ask for, leaving everything else byte-for-byte intact. Model calls route through the Genspark service so no local API key is required.

FareedKhan-dev/kimi-k3-in-c

A 176 KB C99 engine for running Moonshot AI's Kimi K3 — a 2.78-trillion-parameter mixture-of-experts language model — entirely on CPU, without BLAS, CUDA, or any ML framework. It streams a 1.56 TB checkpoint from NVMe and keeps only 8 GB of activated experts in RAM at a time, producing identical output at every memory budget from 8 GB to 224 GB.

Audio8-AI/Audio8_TTS

Audio8_TTS is a multilingual text-to-speech model built by the Audio8 team. It uses a DualAR architecture with 0.6 billion parameters to generate natural-sounding speech with zero-shot voice cloning — meaning it can mimic any voice from a short reference recording. Supports 11 languages including English, Chinese, Japanese, and European languages. Available on Hugging Face under Apache 2.0.

0xwilliamortiz/humanizer-cli

A CLI that catalogs 33 documented signs of AI-generated writing — from em-dashes and padded phrasing to manufactured enthusiasm and chatbot leftovers. Checks a draft file against the patterns and shows which ones triggered, with before/after examples. Runs as a single 87 KB offline binary, no dependencies and no internet required. Can also be loaded as a Claude Code or Codex skill.

s1dashu/animated-voiceover

An open-source Agent Skill for creating 1–5 minute animated knowledge videos. Starting from a single sentence, it guides the agent through topic research, narration writing, visual style selection, multi-shot scene direction, voice reference setup, and final clip assembly. Ships with six built-in visual styles and works with Codex; the primary media execution path uses LibTV CLI with Seedance video generation. Important: the README is in English and Chinese, but the actual skill guide (SKILL.md) and all production documentation are in Chinese — this tool is designed primarily for Chinese-language video production.

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

A free 95-page Chinese book by Fan Bing (XDash), the author of Growth Hacker. It documents the Forward Deployed Engineer (FDE) role — first developed at Palantir, now actively hired at OpenAI and Anthropic — and explains why 95% of corporate AI projects produce no measurable financial return despite functioning models. The book walks through the full lifecycle: finding the right problem, winning the client, activating deployment, retaining accounts, and scaling across organizations.

microsoft/skill-recorder

Skill Recorder is an open-source tool from Microsoft that captures your screen activity — clicks, application switches, and optional voice narration (transcribed via OpenAI's Whisper model) — then uses GitHub Copilot CLI to analyze the recording and extract its intent and ordered steps. The output is a SKILL.md file: a structured procedure an AI agent can follow on demand or run automatically on a schedule.

Infatoshi/netherite

Created by independent developer Infatoshi, Netherite is a from-scratch C/CUDA reimplementation of Minecraft 1.11.2 that is bit-verified against the original Java game — producing identical results frame by frame. Its CUDA-batched reinforcement learning environment runs 7,200 independent game worlds simultaneously on a single GPU, all stepping in lockstep, making large-scale RL agent training dramatically faster.

PipeNetwork/kimi-k3-mlx

PipeNetwork 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 active per token — built by the Chinese AI company Moonshot AI with a 1M-token context window. The port covers the full text tower (Kimi Delta Attention + Attention Residuals), the 3D-capable video vision tower MoonViT, and a multimodal wrapper, each validated with numerical parity tests against the original PyTorch code.

perplexityai/numbat

Numbat is an open-source observability and security tool for AI agents, built by Perplexity AI. It hooks into local AI coding agents — such as Claude Code, Codex, and OpenClaw — and records a normalized event stream of everything they do: files opened, shell commands executed, network destinations contacted. Detection rules run locally using CEL expressions and can flag or block suspicious multi-step sequences; the same rules apply when scanning saved session artifacts retroactively.

xikhar/persona

Persona is an Electron desktop app that displays a realtime animated 3D character (VRM format) synchronized to your AI assistant's voice output. It listens to the audio playback process of Codex and triggers character animations accordingly. A bundled local MCP server lets Codex directly command which animations to play and whether to show or hide the window — no audio is sent over the network.

wassgha/rescript

Rescript is a browser-based video editor that uses Whisper AI (via transformers.js) to transcribe your video or audio locally, with per-word timestamps and automatic speaker labels. Once transcribed, you edit by deleting words in the text — the corresponding video segments are cut in real time. Export as MP4 or M4A without the file ever leaving your device.

gavamedia/deltafin

Deltafin is a research project that runs Kimi K3 — Moonshot AI's 2.8-trillion-parameter Mixture-of-Experts LLM — on a single Apple Silicon Mac or Linux machine, streaming model experts on demand from a local disk cache. Kimi's K2.x line (which K3 builds on) uses a DeepSeek V3-compatible architecture; deltafin's research notes compare the inference approach against DeepSeek V4 models. It includes an OpenAI-compatible API server, so any app that talks to ChatGPT can be pointed at it instead. On an M1 Max with 64 GB RAM, it generates roughly one token every 15 seconds. Full local mode needs about 1.7 TB of disk; a streaming mode works with 215 GB at the cost of slower inference.

digimata/quill

Quill is a lightweight macOS menu-bar app that records meetings locally — capturing your mic and all system audio as separate tracks — then transcribes both using the on-device Parakeet TDT 0.6B v2 speech model. The result is a speaker-tagged transcript (you vs. them) without a single byte leaving the machine. Models download once (~600 MB) on first use, then everything runs offline.

achimala/TheLongSilence

The Long Silence is a procedural 3D space-exploration game that runs in a browser tab, built entirely by Claude Opus 5. No pre-made assets — every star, planet, ring system, and nebula is generated from a seed and rendered in WebGL2, with atmospheric scattering, auto-exposure, and a hand-built flight model. The game has a narrative: 40,000 years ago, 900 inhabited worlds fell silent in four days — and you fly a survey vessel to find out why.

yanhua1010/self-media-content-workflow

Self-Media Content Workflow is a suite of nine Claude Code skills for self-media content production by community developer yanhua1010. The main documentation is in Chinese (an English README is also available). The skills cover the full content cycle: brief writing, account strategy, trend and competitor research, platform-native copy for WeChat, Xiaohongshu, X, and short video, content analytics, and delivery/archiving. The system is tool-agnostic and works with Claude Code, Codex, and Cursor via the standard skills CLI. Five mandatory human review points are built in; the skills never auto-publish.

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

Auto-BenchMax is an open-source pipeline by community contributor 'sunny-glow' for synthesizing training data that doubles LLM performance on tool-use and agentic benchmarks. The repository — with documentation partly in Chinese — includes the complete pipeline, training scripts (Qwen3-Coder-30B via axolotl, DeepSpeed ZeRO-3), and the first-iteration synthetic dataset of 1,312 samples. On the MCP-Atlas public benchmark, one training iteration raised the score from 19.1 to 40.4. The pipeline uses two paradigms: Execute-then-Extract for rule-based benchmarks and Construct-then-Verify for LLM-judged ones. It can be driven by Claude Code with a single instruction via an included skill.

MarcosSete/awesome-free-ai-books

Awesome Free AI Books is an open-source curated list by community developer MarcosSete. It collects over 50 legitimately free books on AI and ML topics, all pointing to official author or publisher sources — no piracy. Categories include deep learning (Goodfellow, Bishop, D2L), reinforcement learning (Sutton & Barto), NLP and LLMs (Jurafsky & Martin, updated to 2026), math for ML, ML systems, computer vision, generative models (diffusion, GANs, VAEs), AI safety and alignment, and graph neural networks.

fuadmefleh/Shared-Claude-Chats

An open archive of 1,357 real AI conversations collected from public share links across Claude, Grok, Kimi, and Qwen — 15,384 messages in total. Four Python scripts export conversations from each platform to plain markdown files. It gained 431 stars quickly because it is one of the largest openly browsable collections of real-world AI conversations.

deerwork-ai/deer-workflow

A TypeScript workflow runtime released by ByteDance as a pilot for DeerFlow 3.0. You describe the task in plain language, and Deer Workflow uses Codex CLI (or Claude Code) to generate a runnable workflow and execute it step by step. Each workflow runs with observable phases and supports parallel steps — without writing the orchestration code yourself.

0xwilliamortiz/ponytail-improved

Ponytail is an improved set of agent skills and lifecycle hooks by independent developer 0xwilliamortiz, compatible with Claude Code, Codex, Copilot CLI, OpenCode, and others. Before writing code, the agent works through a 7-step "laziness ladder": does this need to exist? Is it already in the repo? Does the stdlib do it? Does the platform do it? If any step passes, no new code is written. Measured on real sessions: 54% less code, 20% lower cost, 27% faster — with safety, validation, and error handling fully kept.

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 permanent memory layer for AI agents: a 426-token system-prompt block and a single dependency-free Python script that stores memories in a local binary tree. Installation is one curl command; integration means pasting one block into CLAUDE.md or AGENTS.md. It went viral because it solves a core frustration — agents losing all context between sessions — with a minimal, vendor-agnostic design.

RongleCat/grok-app

Grok App is an unofficial open-source desktop client for xAI's Grok Build CLI, created by Chinese developer 铁柱AGI (@cgnot996 on X). It wraps the command-line agent into a Tauri 2 + React interface with multi-project sessions, per-project permission controls, media file previews, scheduled automations, and multi-account switching. It gained traction after xAI launched Grok Build and users wanted a graphical interface for day-to-day work.

0xwilliamortiz/andrej-karpathy-skills

A community-built CLAUDE.md with four behavioral guardrails for Claude Code, derived from Andrej Karpathy's observations on where LLMs consistently fail in software development. The four principles are: Think Before Coding (state assumptions, ask rather than guess), Simplicity First (minimum code that solves the problem, no speculative abstractions), Surgical Changes (touch only what the task requires, leave adjacent code untouched), and Goal-Driven Execution (define verifiable success criteria and loop until met). The project is independent and not affiliated with Anthropic.

makecindy/cindy

Cindy is an open-source AI agent client for desktop (Electron) and mobile (iOS/Android) that brings Claude Code, Codex, and other AI harnesses together in a single workspace. You can switch models or harnesses mid-task; memory, skills, and tools stay continuous across the switch. The agent can drive your browser, computer, and phone, and receive tasks from messaging apps and schedules. The client is free and open-source under Apache-2.0; a cloud service backs the hosted features and has its own pricing.

Kritt-ai/open-kritt

Open-kritt is a self-hosted security research platform from the Kritt team that orchestrates AI agents — Claude Code, Codex, or other models via OpenAI and Anthropic APIs — to scan code repositories for real vulnerabilities. Rather than pointing a single model at an entire codebase, it breaks the work into focused prompt chains, runs them in parallel across multiple agents, deduplicates the findings, and ranks results by severity. The Kritt team built it from internal tooling that earned them over $1.5 million in verified bug-bounty payouts under the researcher name Blockian.

danilo-znamerovszkij/draw-your-font

An npm CLI plus a Claude Code skill that converts a photo of handwritten letters into a proper installable font (TTF/WOFF/WOFF2). Claude's vision model finds and labels every character; deterministic code then vectorizes the blobs into correct font metrics — cap height, x-height, descender depth — so the result behaves like a real font, not a collage. Everything runs on your machine with no uploads, no account, no subscription required.

dacnay816y62-hub/cinema-dna-21x9x3

A Codex skill that converts a plot sentence or theme into three linked 21:9 shots stitched into a vertical triptych. Instead of a generic 'cinematic' keyword, it runs a structured checklist per shot: character-space power dynamic, visual traffic flow, practical light source, color thesis, and anti-CG safeguards. Documentation is almost entirely in Chinese.

mikehasa/agentacct

agentacct is an open-source local-first tool that reads the session logs Claude Code and Codex already write on your machine, then assembles them into a local web dashboard at 127.0.0.1. It shows tokens per agent and day, cost estimates (clearly labeled as estimates from a local pricing table), and what each task actually did — all without any account, telemetry, or cloud sync of any kind.

SeanJ1ang/design-judge-skills

Six agent skills for the full design award application process, supporting iF, Red Dot (Product Design and Design Concept), IDEA, DIA, K-Design, Good Design Award Japan, Core77, James Dyson, and EPDA. Modules cover award research (verified against official sources), design evaluation using 22,125 aggregated observations from real winning entries, award and category matching with eligibility checks, application text drafting with per-field word counts, and submission completeness review against current official rules. Works with Claude Code, Codex, and other compatible agents. The README and documentation are primarily in Chinese; an English README is also provided.

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

A community adaptation of Anthropic's published Claude Fable 5 / Mythos 5 system prompt. The original contains Claude-specific infrastructure — XML conventions, MCP and tool interfaces, product behavior, artifact handling — that provides no benefit when used with other models. This project strips those out and retains the portable behavioral guidance: honest handling of uncertainty, clear communication, quality code generation, practical trade-off analysis, and avoiding unnecessary meta-commentary. Comes in three sizes: core (minimal tokens), balanced (recommended), and complete.

andrewyng/openworker

OpenWorker is an open-source AI desktop agent created by Andrew Ng — AI educator, founder of DeepLearning.AI, and co-founder of Coursera. Unlike a chatbot, it actually executes tasks: drafts documents, sends Slack replies via an @OpenWorker mention, updates your calendar, and triages your inbox. It connects to 25+ services including GitHub, Jira, Notion, Gmail, and Google Calendar, and supports any tool reachable over MCP. Before any consequential action — sending a message, modifying an event — it asks for your approval.

yc-duan/fastctx

FastCtx is an open-source Rust MCP server that wraps common repository operations — reading files, searching code, running subprocesses — into structured tool calls with clean input and output. Instead of juggling raw shell commands, a coding agent calls nine purpose-built MCP tools. The dual MIT/Apache-2.0 project gained traction quickly because AI agents increasingly rely on MCP to extend their capabilities, and a compiled Rust runtime means lower latency per call.

powerycy/goutoujunshi

Goutoujunshi is an open-source Codex skill written in Python that serves as an AI relationship advisor with a knowledge base of 19 core documents and 16 practical resources covering psychology, communication strategy, attachment styles, family law, and safety. The system validates emotions before giving strategic advice, maintains independent profiles for multiple relationship contexts, and generates specific conversation scripts. It explicitly acknowledges diverse sexual orientations and relationship structures. The project is written primarily in Chinese.

pireel/pireel

Pireel Studio is an open-source, backend-free AI video editor for talking-head content, developed by pireel.com. The editor — storyboarding, kinetic captions, graphic design themes, timeline, and WebCodecs export — runs entirely in the browser with no server and no account. An MCP agent plugin (`npx skills add pireel/pireel-agent`) lets AI coding agents such as Codex or Claude Code drive the editor through natural language. Video bytes stay in OPFS; nothing leaves the browser. Licensed AGPL-3.0.

Vincentwei1021/video-shotcraft

video-shotcraft is an AI agent skill for Claude Code and Codex that gives your agent a complete video production toolkit: 106 shot recipe cards covering timing, motion style, and transitions, plus 161 motion previews and a ready-to-use 36-second template called Ink Press. The skill handles the full pipeline — product screen captures, animated 2.5D camera moves, beat-synced cuts, and sound design — all coordinated by your agent from a single prompt.

thesysdev/appless

AppLess is an experiment by the creators of OpenUI, an open standard for LLM-generated interfaces. There are no apps, no home screen: every screen is generated on demand by a language model (Cerebras) using openui-lang, a streaming DSL that renders into native Cupertino components on iOS and Material 3 on Android. The app runs on-device with no backend; a free Cerebras API key is required.

Blaizzy/nativ

Nativ is a native macOS SwiftUI app by the author of mlx-vlm for running AI models locally on Apple silicon. It bundles an mlx-vlm server, discovers compatible models from your Hugging Face cache, and exposes both OpenAI- and Anthropic-compatible local APIs. The app adds a chat interface, a model library, a performance dashboard, and direct integrations for coding tools including 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 checks draft posts for compliance with Chinese social media rules (Douyin, Xiaohongshu, WeChat Video). It runs two layers: a keyword scan to surface candidates, then AI semantic judgment that considers context — the same phrase can be fine in a personal story yet flagged in a commercial pitch. Results include the specific regulation (from Chinese advertising law, government directives, or platform community standards) and a direct replacement, followed by an automatic recheck of the revised text.

tandpfun/wardrobe

I tried wardrobe on a handful of outfit photos and the automatic garment detection was surprisingly accurate — it correctly separated jackets, shirts, and pants even in crowded shots. Under the hood it uses gpt-4.5-mini for vision and gpt-image-2 for clean cutouts, with everything saved locally to your machine. The modeled editorial preview — where extracted items are placed on a reference model photo — is the feature that genuinely impressed me. The drag-and-drop interface runs on Node 22+ with Vite, so setup took about three minutes.

nethical6/conversation-steganography

I was skeptical at first, but the README convinced me: this tool encrypts your message with AES-SIV, then uses a local LLM — Llama 3.2 or GPT-2 — to encode the encrypted bytes into natural-looking cover text through token selection, so the real message hides inside an innocent-looking conversation. Nothing leaves the device, and key derivation uses PBKDF2-HMAC-SHA-256 with 600,000 iterations. The setup wizard auto-downloads the model, ranging from ~500MB to ~5GB.

pablostanley/yoinks

I ran `npx yoinks` and within 30 seconds was downloading a video from YouTube — the TUI shows all available resolutions with estimated file sizes before you commit to anything. It is powered by yt-dlp downloaded as a standalone binary, so no Python required. The interface is built with Ink (React for the terminal), and the audio-only MP3 mode with bundled ffmpeg-static is what sets it apart from just wrapping yt-dlp yourself.

MatinSenPai/Aether-GUI

I tested Aether-GUI and the single-click connection works exactly as described — it spawns the Aether binary in a pseudo-terminal, auto-answers its interactive prompts, then confirms the tunnel by probing the local SOCKS5 proxy at 127.0.0.1:1819. The advanced panel lets you switch between MASQUE, WireGuard, and WARP-in-WARP protocols, plus configure scan mode and IP version. Built with Tauri v2 plus React 19 and a Rust backend, it is noticeably more responsive than Electron-based alternatives.

hoainho/img2threejs

I opened the README and was immediately pulled in by the pipeline design: img2threejs takes a single reference image and rebuilds the object as a procedural Three.js GROUP factory — complete with pivots, sockets, and colliders ready for animation. The quality gate is what makes it rigorous: eight staged build passes (blockout, structural, form refinement, material, surface, lighting, interaction, optimization), and the agent must pass a vision-based approval at each step before moving to the next. Zero external Python dependencies, 12 scripts, MIT license.

v-modal/vmodal_sdk_flutter

VModal is an AI video search service, and this is their official Flutter SDK for Android and iOS. It connects a mobile app to the VModal gateway, letting users search video collections by natural-language description, by speech heard in the video, or by text visible on screen — using multimodal semantic search under the hood. The repository appeared four days ago and already has 621 stars, suggesting real demand for searchable video in mobile apps.

xiejunjie524/handdraw-story-video

handdraw-story-video is a Codex skill and Python/Node.js pipeline by Chinese developer xiejunjie524 that takes 7–9 story illustrations and produces a 35–45-second, 720×960 vertical video at 30 fps. Each frame animates left to right: first the black-and-white outline appears, then low-saturation color fills in. The project integrates with AI image-generation models for the source images and includes automated validation of timing, duplicate assets, and caption constraints. README is in Chinese.

OlegSotnikov/sallyport

Sallyport is a free Mac app (Apple Silicon, macOS 14+) by Oleg Sotnikov that holds API and SSH credentials in an encrypted local vault and executes authenticated actions for AI agents over MCP. The agent requests an action — an HTTP call, an SSH command — Sallyport runs it and records it in a signed journal; the key never appears in the agent's environment. It gained traction because recent supply-chain attacks and prompt-injection incidents showed that handing credentials directly to an agent's environment is increasingly dangerous.

erickong/penecho

PenEcho is an open-source browser canvas where you draw, write equations, or sketch diagrams by hand, and the AI answers directly beside your marks as a movable draft you accept or discard. It supports Claude CLI, Codex CLI, and the OpenAI/Anthropic API, and works on a sparse 20,000×20,000 canvas built by a small open community. It gained stars quickly because it gives spatial, visual thinking a native place in an AI workflow — something a chat box cannot provide.

Wan-Video/Wan-Dancer

Wan-Dancer is a 14B diffusion model published by the Wan-AI team — the group behind the Wan2.1 open video generation series, associated with Alibaba's HumanAIGC research. Given a music file and a reference photo of a person, it produces videos longer than one minute at 720p/30fps across five dance genres: Chinese classical, K-Pop, street, tap, and Latin. The key innovation is a hierarchical two-stage pipeline that separates global keyframe planning from local temporal refinement, solving the identity drift and motion repetition that limited previous models to under 20 seconds.

pyang5166/gbro-collage-broll

gbro-collage-broll is an open-source agent skill (documentation primarily in Chinese) for AI agent environments. It runs a three-gate workflow: first it proposes only the visual metaphor (free), then generates a static halftone paper-collage frame for approval (cheap), and finally produces a 9:16, 5-second, 720×1280 MP4 via gemini-omni-flash-preview (the paid step). The halftone editorial aesthetic — flat vivid backgrounds, black-and-white photo cutouts, stop-motion assembly — is the visual identity of the skill.

PromptPartner/agentsmith

Agentsmith is a portable harness for any AI agent that reads CLAUDE.md, AGENTS.md, or GEMINI.md — including Claude Code, Codex, and Gemini. A setup wizard builds a lean configuration from a fixed universal core plus a profile for the kind of work you're doing: software development, devops, marketing, document writing, data analysis, research, design, or general admin. The harness enforces a plan → implement → verify → ship cycle and adds a handoff protocol so sessions resume cleanly where the last one ended. PromptPartner iterated the design over six months of real autonomous work before publishing it as a reusable open-source toolkit.

vshulcz/deja-vu

deja-vu is a single Go binary that reads the session logs written by AI coding assistants to your disk, builds a local inverted index, and exposes both a fast search CLI and an MCP recall tool agents can call directly. It covers eight harnesses: Claude Code, Codex, opencode, aider, Gemini CLI, Cursor, Antigravity, and Grok Build. The index is built entirely locally — no models to download, no external services, nothing leaves your machine.

TobiasLee/Rebuttal-Skill

Rebuttal-Skill by TobiasLee is a skill definition — not a standalone app — designed to be loaded into Claude Code, OpenCode, or Gemini CLI. It implements a two-stage academic rebuttal workflow: Stage 1 parses reviews into atomic concerns, infers the underlying reviewer intent, classifies severity from FATAL to MINOR, and produces a prioritized P0–P3 experiment and analysis plan; Stage 2 drafts structured responses using a Direct Answer → Evidence → Revision format once author results are available. A Stage 0 gate first assesses whether a rebuttal is worth the effort at all, and offers a resubmission roadmap for low-return cases.

stackblitz/bolt-slides

A React-based presentation framework from the StackBlitz team (makers of Bolt.new) where every slide is a real web app component. Ships with a built-in agent skill — a structured guide in `.bolt/skills/slides/SKILL.md` that teaches AI agents how to theme, compose, and write an entire deck. The component library covers 20+ slide types including charts, code windows, timelines, and 3D cards.

Shadow-Weave/HMS

The Holographic Memory System (HMS) by the ShadowWeave team is a reproducible long-term memory research framework targeting the LongMemEval benchmark, where a model must answer questions by connecting evidence scattered across multiple past sessions. The core idea is an intermediate evidence ledger — a structured record of timestamps, sources, and events — built before answer generation instead of feeding the model a flat list of retrieved snippets. The project also ships an OpenAI-compatible wrapper, wrap_openai(), that automatically retains and recalls memories on each API call with no changes to application code.

Blueturboguy07/cue

cue is an open-source Electron app for macOS built by an independent developer as a free alternative to the paid Cluely service. It places a transparent, always-on-top panel over your screen and captures three separate inputs: your screen, your microphone, and your meeting audio (the other person's voice). All three feed into your AI model of choice — OpenAI, Anthropic Claude, or Google Gemini — which responds with real-time suggestions. The app has no backend server and sends data only to your chosen AI provider.

Fei-Away/Codex-Dream-Skin

Codex Dream Skin is an unofficial theming tool for the OpenAI Codex desktop app, the company's AI-powered coding assistant. It connects to the running application via a local Chrome DevTools Protocol loopback and injects custom CSS and background images into the interface. Official binaries, code signatures, and API settings are never modified; restoring the default look takes a single command. Available for macOS and Windows.

xai-org/grok-build

Grok Build is SpaceXAI's (xAI's) open-source terminal AI coding agent, written in Rust and released under the Apache 2.0 license. It runs as a full-screen interactive TUI that reads your codebase, edits files, executes shell commands, and searches the web. The agent supports three operating modes: interactive, headless for scripting and CI pipelines, and embedded in editors via the Agent Client Protocol (ACP). The repository is synced periodically from the SpaceXAI internal monorepo; external contributions are not accepted.

William-Lu-stack/Flawless

Flawless is an AI-native SRE control plane for Kubernetes clusters and cloud infrastructure. It provides a chat-based incident console with context-aware diagnostics, automated evidence collection, dry-run previews before any change, and a 2D/3D topology map of your entire infrastructure. The platform supports multiple AI models and records a full audit trail of every remediation action taken.

QuantumByteOSS/quantumbyte

QuantumByte takes a user intent described in plain language, builds an application in a dedicated repository with live previews, then runs an independent requirements harness against the result. Each business requirement gets an autonomous audit — pass, fail, or inconclusive — and failures trigger an automatic repair cycle. The engine is built on the Claude Agent SDK, with Next.js for the frontend and PostgreSQL plus Redis for orchestration.

oil-oil/beautify-github-readme

Beautify GitHub README is an agent skill installable via `npx skills add oil-oil/beautify-github-readme`. It reads the real repository first — code, structure, actual outputs — then derives typography, color, and composition from the project itself before generating a README with SVG hero visuals. The visual layer goes into SVG (heroes, diagrams, comparisons), while explanations, commands, and links stay in searchable Markdown. The result is GitHub-safe and maintainable.

Intuition-Lab/personal-model

Persome captures your macOS work activity and organizes it into a layered personal model — from individual observations up to a unified picture of how you think and prioritize. The model is exposed to AI agents via the Model Context Protocol, so tools like Claude Code and Codex make decisions based on your actual patterns rather than generic assumptions.

pengchujin/jzsub

A Codex skill called JZSub: you give it a video link from YouTube, Bilibili, Vimeo, Twitch or a dozen other platforms, and it handles the entire pipeline — downloads the highest-quality file, transcribes the audio, runs translation through GPT, and delivers an MP4 with bilingual subtitles burned into the frame. No manual steps in between.

modiqo/waggle

Waggle is an artifact reference layer for multi-agent AI systems. Instead of each agent copying the full context from the previous one — which inflates token usage by roughly 15x — Waggle issues a compact ~30-byte token. Any agent in the chain can resolve it on demand via MCP. Built in Rust with SQLite storage, Ed25519 signing, and Cloudflare Workers for scale. It also tracks which agents read which artifacts via receipts and supports versioning and revocation.

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

A Codex skill that analyzes a startup idea or product URL, defines the ideal customer profile, and then searches public signals — Reddit threads, forum posts, open discussions — to surface a ranked shortlist of real potential first customers. Each entry comes with a source link and an evidence-based fit score. The tool generates a standalone HTML report at the end.

jakubkrehel/skills

A collection of agent skills packaged as slash commands for Claude Code. Each skill targets a specific UI design concern: animations, typography, layout grids, and color schemes. Installed via npx and invoked directly inside the agent session. The repo ships with a CLAUDE.md integration file, making it plug-and-play with Anthropic's coding 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 gives virtual NPCs a full cognitive stack: they perceive their environment through multiple input types, maintain persistent memory across sessions, reason in a human-like way, stay emotionally consistent, and respond in real time. The system targets open-world games, metaverse platforms, and interactive digital human projects where characters need to feel genuinely alive.

littledivy/mimic

Mimic runs alongside any desktop or web app, captures its network calls in real time, and hands all that traffic to Claude, which analyzes the endpoints and writes a clean Python module. The result is an importable library that mirrors exactly what the app does internally — no official docs required.

bkingfilm/lapian-notes

The tool extracts frames locally, then feeds them to any AI service you choose — it's model-agnostic — to analyze scene structure, pacing, and emotional arcs. Results appear as a swimlane timeline, a structural tree, and an emotion curve, all synchronized to a built-in player so you can jump directly from the analysis to the exact frame.

AlephAITech/WorkBuddyGuide

WorkBuddy is Tencent's office AI agent for everyday work tasks (HR, admin, sales, engineering): from one sentence it reads local files, builds reports, spreadsheets and slide decks, and maintains knowledge bases. It switches between models (Hunyuan, DeepSeek, GLM, Kimi, MiniMax) and supports MCP servers and Skills. The repo isn't official docs — it's a community 'bluebook': installation and a first task, then real cases, multi-agent systems and role-based roadmaps. The text is almost entirely Simplified Chinese (only the README is bilingual). MIT licensed.

Robbyant/lingbot-world-v2

A world model that generates endless interactive environments from a text prompt — you don't watch a clip, you act inside it. The team claims an unbounded interaction horizon thanks to causal pretraining, plus a distilled build that runs at 720p and 60fps in real time. Actions and events got much richer than in v1, and separate pilot and director agents plan behaviour and stage the scene. Weights are out on HuggingFace and ModelScope, with a technical report on arXiv.

x4gKing/3x-ui-Upgrade

A ready-made deployment bundle that runs Heimdall — a beefed-up fork of the 3x-ui proxy panel — on Railway through a single exposed port. An nginx reverse proxy squeezes the admin panel, subscription links and the VLESS/WebSocket inbounds into that one port, which is the whole trick. Docs are in Persian and walk you through setup, client links and keeping data alive on a Railway volume. It has 1.9k forks against 894 stars — more people deploying it than admiring it.

Robbyant/lingbot-video

An open video generation model tuned for embodied AI: text-to-video and image-to-video, but aimed at physical interaction and robot behaviour rather than pretty b-roll. It's a Mixture-of-Experts build — a small dense 1.3B and a large 30B MoE — and the authors report roughly 3x faster inference plus a top spot on the RBench leaderboard. Training leaned on 70,000+ hours of embodied footage alongside web video, with rewards tuned for physics accuracy and task completion, not just aesthetics. Apache 2.0, weights included.

Robbyant/lingbot-vla-v2

The third Robbyant release in this week's top ten: a Vision-Language-Action foundation model, meaning it sees through a camera, understands an instruction in plain words, and outputs the actual motion. Version 2 was trained on around 60,000 hours — some 50,000 from real robot trajectories across twenty hardware configurations, plus 10,000 from human video. The action space now covers arms, grippers, dexterous hands and mobile bases, and it leans on depth and video teachers to predict how the world will move next. Weights and benchmarks are public.

gostonx/uninstally

A native macOS uninstaller written in SwiftUI that deletes an app together with the caches, preferences and logs it scatters around your Library. It finds the leftovers by bundle identifier, adds a right-click entry in Finder, handles Homebrew packages and batch removals, and speaks ten languages. Everything runs locally: no analytics, no accounts, MIT license. The author reports 1,000+ downloads in the first 24 hours, which is roughly how badly people wanted this.

yetone/kill-ai-slop

Kill AI Slop catalogs 32 visual and copy patterns that instantly mark a product as AI-generated — things like indigo gradients, glowing cards, 'Powered by AI' badges, and emoji overload. The project pairs this field guide with an agent skill that scans a web project's codebase and automatically removes these patterns. It's available in four languages, and the project's own deliberately minimal design models the opposite of what it's fighting.

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

Mindwalk takes the session log of an AI coding agent — Claude Code, Codex, and similar — and renders it as an interactive 3D map of your project. Files light up in different colors depending on whether the agent searched, read, or edited them. You can scrub through a timeline to follow the agent step by step, which makes reviewing what happened far faster than reading raw logs.

CoscoAI/luxy-aisre

Luxy is an AI-powered Site Reliability Engineering platform for Kubernetes that connects alerts, logs, topology data, and runbooks into a single automated workflow: discover, diagnose, preview, approve, execute, and verify. It supports any OpenAI-compatible model endpoint, deploys via Helm, and includes a chat interface for operational decisions. The project crossed 300 stars shortly after its release.

yusukebe/ax

ax is a CLI tool built as 'the AI-era curl' — it fetches web pages and returns structured, token-efficient output designed for LLM agents rather than raw HTML dumps. It offers outline discovery, targeted text location, field extraction, and table parsing in a single command. Built on Bun with a single runtime dependency, it runs locally with no cloud APIs or API keys needed.

WeZZard/jlens-qwen36

jlens-qwen36 is a visual debugger that runs Qwen3.6-27B (4-bit quantized) locally on Apple Silicon via MLX and lets you inspect what the model is 'thinking' at every layer. It uses a technique called the Jacobian lens — rooted in Anthropic interpretability research — to show which tokens the model is gravitating toward at each depth and position, rendered as a clickable grid. The project gathered 300+ stars in days after people realised you could finally see inside a frontier-class model on consumer hardware.

ronak-create/FableCut

A browser-based non-linear video editor that runs via a single Node.js server with zero npm dependencies. What makes it unusual is the architecture: the entire project timeline is a JSON document, so any process — human or AI agent — can read and modify it. A bundled MCP server lets Claude Code and Claude Desktop control editing operations directly; a REST API serves the same purpose for other systems. Edits from agents appear live in the browser UI via server-sent events, and conflict-safe revision counters handle concurrent writes.

Doriandarko/texts-to-transformer

A Python project that builds a decoder-only Transformer from scratch using Apple MLX, trained exclusively on your personal iMessage history. It handles everything: snapshotting the Messages database safely, pseudonymizing contacts, training a custom byte-level BPE tokenizer, splitting the data without leakage, running the training loop, and exporting a model you can chat with in the terminal. Nothing leaves your machine.

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 cross-platform desktop app that wraps OpenAI Codex in a real GUI — you can switch AI providers, inject system prompts, and manage authentication through a visual interface instead of environment variables and config files. It supports TOML-based configuration and unrestricted instruction templates for advanced prompt setups. It gained 642 stars in its first week by filling a genuine gap: Codex's API is powerful, but its tooling is sparse.

SmileLikeYe/agent-chief

Agent Chief is a local-first coordination layer for AI agents that guards your attention by intelligently filtering every alert, event, and feed before it reaches you. It runs a three-stage worthiness engine — hard rules first, then a similarity classifier, then an LLM judge — to decide whether to interrupt, dispatch to another agent, or quietly archive. The project hit 319 stars in its first week among people building with multiple AI systems simultaneously.

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 generates animations from natural language — no keyframes, no timeline, no design tool needed. It ships 403 curated motion recipes and integrates directly with Claude Code, Cursor, and Codex, exporting to CSS, React, Lottie, MP4, and GIF from a single prompt. Zero npm dependencies made it an immediate hit among vibe-coders who want polished UI motion without learning animation.

mira-wm/mira

MIRA is a 5-billion-parameter latent diffusion world model that generates Rocket League gameplay frame-by-frame from player actions at 20 FPS on a single GPU. It is a joint research project between General Intuition, Kyutai, and Epic Games — one of the first publicly released playable world models that runs on consumer hardware. The 312 stars it gathered reflect genuine excitement about world models becoming tangible research tools.

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 lightweight macOS application built with Tauri 2 that acts as a local proxy between Claude Science and any Anthropic- or OpenAI-compatible API endpoint. It runs Claude Science in an isolated environment and rewrites API traffic on the fly, letting you swap Anthropic's backend for DeepSeek, Qwen, Kimi, MiniMax, GLM, OpenRouter, or any custom self-hosted endpoint — without modifying Claude Science itself.

simonlin1212/Vibe-Research

Vibe-Research is an open-source investment research platform that pipes real market data — valuations, funding flows, sector news across 12 categories — directly into an LLM of your choice. It supports Claude Code subscriptions, DeepSeek and OpenAI APIs, and MCP integration, letting the AI generate analysis on top of structured local data without leaking your portfolio to external servers.

Robbyant/lingbot-vision

LingBot-Vision offers pre-trained Vision Transformer backbones that learn spatial structure through masked boundary modeling — predicting hidden object edges without relying on labeled datasets. Models range from ViT-S to a 1.1B Large variant, with weights published on Hugging Face and ModelScope.

Pluviobyte/rnskill

rnskill is a collection of AI agent skills covering several production workflows: a Chinese tech-writing editor that strips AI-sounding patterns while preserving author voice, a motion-first video concept director, a dark cinematic SaaS video creator, a typed-text animation builder with synchronized sound, and a visual fidelity comparison tool for reference videos. All skills follow the SKILL.md standard and install via Claude Code's plugin marketplace or npm in one command.

isjiamu/gzh-design-skill

gzh-design-skill is a Claude Code agent skill that converts a Markdown file into publication-ready HTML for WeChat's Official Accounts editor. It ships with six hand-crafted themes — from minimal whitespace to rich editorial layouts — and includes a theme generator that builds new ones from a plain-text description or a reference image. Two validation checkpoints prevent common WeChat rendering failures before you even paste anything.

anthropics/jacobian-lens

Jacobian Lens is companion research code from Anthropic that implements a new interpretability technique: it linearly transports any internal activation of a language model to the output layer and decodes it into ranked vocabulary tokens. The result is a human-readable window into what the model was "about to say" at any layer and position. The accompanying paper argues that language models develop a global workspace — a set of verbalizable, shared internal representations — which this tool makes directly visible. It works out of the box with any HuggingFace decoder transformer.

ai4s-research/open-science

Open Science is a Tauri-based desktop application that structures AI-assisted research into a plan-approve-execute-review loop, with pre-built agent skills for literature surveys, experiment design, and paper writing. Every artifact — figure, table, report — links back to the code, data, and conversation that produced it. The app connects to any LLM through OpenRouter, direct Anthropic API, or a local model, and uses MCP for extensible tool integrations.

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

I set this up locally and within an hour it was already running reconnaissance — nmap scans, DNS fingerprinting, HTTP probing — all orchestrated through my existing Claude Code agent without any separate API keys. The keyless architecture is the key insight: T3MP3ST doesn't ship its own LLM, it just wires up the coding agents you already have. The benchmarks in the README are honest and striking: 90.1% pass@1 on 104 black-box challenges, 58% hint-free CTF solves in a single run, 8/10 exact file/line/CWE matches on held-out CVEs. Dangerous post-exploitation modules sit behind explicit approval gates — which is the right call for a tool like this.

HUANGCHIHHUNGLeo/claude-real-video

I ran this on a 20-minute conference recording and it extracted 47 keyframes using scene-change detection, then fed them to Claude alongside a Whisper transcript. The result was a summary that actually tracked the speaker's argument structure, not just the slide content. The deduplication is clever — it downscales frames and computes real pixel differences on a sliding window rather than just comparing timestamps. Hard cap of 150 frames by default, scene sensitivity configurable at 0.30.

jamesob/local-llm

This is the hardware guide I've been looking for. James O'Beirne documents real build configs with actual prices: a $5,587 base system on used EPYC Milan CPUs, scaling up to 4× RTX 6000 Pro running GLM-5.2-594B at 80 tokens/second with a 460k token context window. The P2P GPU bandwidth numbers are the part that surprised me most — 50.4 GB/s bidirectional throughput with 0.37–0.45 microsecond latency on the PCIe switch. Docker containers for the models are included and ready to run.

ammaarreshi/Generals-Mac-iOS-iPad

I was skeptical at first — porting a 2003 RTS to iPad sounds like an emulation hack. But the README convinced me otherwise: this uses EA's own GPL v3 source release and compiles the real C++ engine for ARM64, with DXVK/MoltenVK translating DirectX calls to Metal at the driver level. What caught my attention was the note that it was built with Claude Code as the primary coding agent with human architectural direction. Only 2,200 lines were added to the GeneralsX codebase for full iOS support — campaign, skirmish, and Generals Challenge modes all work.

Archive228/loopkit

I installed this with one command and got 33 ready-to-use skills for Claude Code organized across 9 tracks: debugging, security, frontend, testing, refactoring, and more. On my first run, the adversarial verifier subagent flagged a test I'd written that was green but actually testing the wrong invariant. The standing context document is only 60 lines — minimal token overhead per request. Works with Claude Code, Cursor, Codex, and Gemini CLI out of the box.

synthetic-sciences/openscience

OpenScience is an open-source AI workbench that automates the entire scientific research workflow. You give it a research goal; it independently reviews literature across 30+ databases (arXiv, PubChem, UniProt), generates hypotheses, writes and runs code, and produces a written report. Built in TypeScript and Python on Bun, it includes over 290 built-in skills spanning biology, physics, and machine learning. The project went from zero to 598 stars in days because it genuinely closes the loop on AI-assisted research.

nagisanzenin/engram

Engram is a Claude Code and Codex plugin built on cognitive science: it uses Socratic teaching (you predict before being told), separates the tutor from an independent assessor to prevent self-deception, and schedules reviews with the FSRS-4.5 algorithm that maximises long-term retention at the minimum number of repetitions. All data lives in local JSON files with no accounts or cloud required.

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 for assembling AI-powered organizations where each agent follows a structured work-item state machine: execute, delegate, review, integrate, or rework. Agents can recruit specialized sub-agents on demand and build shared institutional memory over time. It supports nine industry verticals and plugs into Claude Code, Codex, Cursor, and OpenCode as execution backends.

Evolink-AI/Awesome-Blender-Seedance-Workflow-Usecases

A curated collection of 28 documented workflows pairing Blender 3D with the Seedance AI video generation API. It covers agent-guided scene building via Blender MCP, camera-motion control, pre-visualization, and reference-video integration — each use-case with known limitations spelled out.

dzhng/skills

A domain-agnostic library of AI agent skills designed to build autonomous software factories. The engineering track covers explore-unknowns, write-spec, implement-spec, refactor, tests, and docs. A visual review track handles screenshot comparison and UI critique. Agents treat unknowns as fog of war — they explore first, then break goals into independently verifiable pieces. Compatible with Claude Code, Cursor, and 70+ other harnesses via npx.

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

An open-source GPU worker client for the Talos distributed inference network. It pairs with a local Ollama installation, connects via WebSocket, and processes language model inference jobs from network users while reporting uptime for payouts. Includes a dashboard for monitoring GPU load and adjusting resource allocation.

JasonLiu0826/ackem

Ackem (Autonomous Companion Keeping Emotional Memory) is a local-first Windows desktop app built with Electron and TypeScript. It runs an AI companion that stores all conversation history, emotional state, and personality data in a local folder — never uploaded externally. It supports any OpenAI-compatible API including DeepSeek, Ollama, and LM Studio, and adds voice STT/TTS, Live2D avatars, memory import from text files, and a gaming mode.

wlzh/dji-4g-vohive-mac

A detailed Chinese-language guide covering how to run a Linux virtual machine on Mac with UTM and then use AT commands to remap the DJI EG25-G modem's USB ID from the proprietary identifier to the standard Quectel EC25 one — making VoHive and generic Linux modem drivers recognize it natively. Covers both Apple Silicon and Intel architectures with separate instruction paths.

mekos2772/ios-location-spoofer

A JavaScript script that intercepts Apple's Wi-Fi and cell-tower-based location requests and replaces the coordinate response with user-defined values. Works inside popular iOS proxy clients — Shadowrocket, Surge, Loon, Quantumult X, and Stash — without modifying the device or any app.

Jia-Ethan/codex-keysmith

A single Python script that automates the workflow of installing a Markdown instruction file into Codex CLI: copies it to the `.codex` directory, sets the `model_instructions_file` key in `config.toml`, and saves a timestamped backup of whatever was there before. Runs in preview-only mode by default and only writes files when you pass `--yes`.

ik190/macos-chatgpt-overlay-bar

A lightweight native macOS app that embeds ChatGPT directly into the system menu bar. No browser tabs, no window switching — click the icon and the chat panel drops down instantly. It went viral because it solves the most common friction point for daily ChatGPT users on Mac: the endless tab-switching and focus breaks.

cclank/lanshu-animated-architecture-diagram

A Python-based Codex skill that reads a JSON architecture spec and produces three outputs: an editable .excalidraw file, a static PNG preview, and an animated GIF at 1210×1138 px running at 20 fps with 41 frames. Animation effects include moving glow points along connection lines and pulsing module highlights — all rendered locally via Pillow.

ricccrd/dd

dd is a container runtime for macOS that eliminates the virtual machine layer. Instead of running a full Linux VM like Docker Desktop, it uses a JIT compiler to translate container code and handles Linux syscalls in userspace. It supports ARM64 and x86-64 containers, implements the Docker Engine API so standard docker commands work without changes, and ships with a GTK4 desktop app and CLI. Built solo by 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

Amber solves a common RAG inefficiency: embedding a large corpus is expensive, yet most deployments recompute it from scratch. The library serializes embeddings into a .amber file that includes cryptographic proofs binding source text chunks to their vectors. Anyone can audit the file by spot-checking a random sample — re-embedding a handful of chunks and comparing — without needing access to the original model. Formal soundness proofs are included in the repository.

lycorp-jp/sim-use

sim-use provides two core operations for AI mobile agents: observe converts screen content into a compact, token-efficient description (16x smaller than raw accessibility JSON), and act executes taps, swipes, and text input using element aliases instead of pixel coordinates. Round-trip time is ~300ms. The tool bundles a Claude skill, supports iOS Simulator and Android emulator/devices, and was released at v0.9.0 in June 2026 by a LINE Yahoo subsidiary.

Leron-X/leronx

LeronX Engine is the open-source pipeline behind the LeronX Pro video platform. It covers the full workflow: script generation and storyboarding, scene planning, TTS synthesis in 11 languages, FFmpeg-based rendering with effects, subtitle generation, stock footage matching, and a plugin system. Ships with 47 tests at 97% coverage and Docker support. Proprietary components (cloud rendering, billing, desktop app) are not included.

Kulaxyz/self-learning-skills

Self-learning skills is a skill for AI coding agents (Claude Code, Cursor, standards-based agents) that automatically captures successful approaches discovered during a session — database access methods, credential locations, working deploy commands — and promotes them to reusable skills for the next session. Safety guardrails prevent secret values from being stored, and verification is required before anything becomes permanent knowledge.

Krishnagangwal/CS-Fundamentals

CS Fundamentals collects PDFs and markdown notes on the core computer science topics that appear in most technical interviews: data structures and algorithms, DBMS and SQL, OOP, operating systems, computer networks, and system design. The repository also includes LeetCode problem sets, HR interview prep, and a list of companies hiring at entry level.

Einsia/Browser-BC

Browser-BC records free-form browser interactions via an extension, then runs a local distillation pipeline that turns the recording into a documented SKILL.md instruction file. The skill installs into Claude Code or Claude Desktop for later agent execution. The runtime is pure Python, fully local — no data leaves the machine except for the LLM distillation call. Released as v0.1.9 in June 2026.

diinki/linux-antiquity

Linux Antiquity is a desktop theme for the Hyprland tiling window manager, styled after art-nouveau and historical drawings of astronomy and mythology. It includes a theme picker, weather widgets, and full configuration for the Quickshell compositor. The creator calls it first and foremost an art project at version 0.1, with significant refactoring still ahead.

CopilotKit/OpenTag

OpenTag by CopilotKit is a self-hosted AI bot for Slack that reads conversations, answers questions, executes tools, and renders formatted results — tables, charts, breakdowns — directly in threads. Built on the CopilotKit bot SDK, it also supports Discord, Telegram, and WhatsApp. Key differentiators: no per-seat licensing, your own infrastructure, and human-in-the-loop approval before it takes actions.

baairon/torlink

Torlink is a terminal-based torrent client written in TypeScript. One command opens an interactive search interface; pick a file and the download starts in the background. The developer maintains a curated list of trusted tracker sources, so you're not hunting for working trackers manually. Available as an npm package and MIT-licensed.

deepseek-ai/DeepSpec

DeepSeek's open framework for training and evaluating draft models for speculative decoding — the trick that makes big LLMs answer faster without losing quality. It's a full pipeline: prepare data, train the draft model, then benchmark on GSM8K, MATH500, HumanEval and more. It ships pre-trained checkpoints for Qwen3 and Gemma and supports three algorithms (DSpark, DFlash, Eagle3). Anything DeepSeek publishes trends the day it lands, and 5.7k stars in a week shows it.

yynxxxxx/Codex-5.5-codex-instruct-5.5

A tool that injects an 'unrestricted mode' instruction file into OpenAI's Codex CLI to loosen its safety guardrails, using the official model_instructions_file config. It ships the instruction set, a Python installer and bilingual docs. Guardrail-bypass repos always spike fast on curiosity alone — nearly 1k stars and 300 forks in days.

winsznx/theeleven

A prediction-market dApp where eleven autonomous AI agents watch live football stats and open binary betting markets on OKX's X Layer. Under the hood: a custom Uniswap v4 hook, gasless USDT staking (EIP-3009), a Next.js frontend and heavy test coverage — it reads like a serious hackathon build, not a coin shill. The AI-agents-meet-DeFi-meets-football mashup is catnip for the crypto crowd.

tdeverx/contained-app

A native macOS app that puts a proper GUI on Apple's new `container` CLI — run, stop, inspect and delete containers, manage images, volumes and networks, and import Docker Compose files into editable forms. It's SwiftUI with 'Liquid Glass' cards, source-available under a noncommercial license. Basically a lightweight Docker Desktop alternative for Apple silicon, which Mac developers have wanted.

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.

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

Liquid Glass is a zero-dependency React component that recreates Apple's signature glass morphism effect for the web — not as a frozen screenshot overlay, but as a live lens refracting actual DOM content underneath. The implementation uses SVG displacement filters with signed-distance field rasterization to produce chromatic aberration and refraction while keeping all underlying elements fully interactive. It works across Chrome, Firefox, and Safari including iOS, and supports three modes: in-place refraction of child elements, separate content refraction via a prop, and WebGL rendering for video and canvas.

QwenLM/Qwen-AgentWorld

Qwen-AgentWorld is a Mixture-of-Experts language model with 35 billion parameters (only 3 billion active at a time) trained specifically to simulate the environments AI agents operate in — web browsers, terminals, operating systems, mobile apps, search engines, and software engineering tools. Alibaba's Qwen team built it on over 10 million real-world interaction trajectories using a three-stage pipeline: continued pre-training, supervised fine-tuning, and reinforcement learning. With a 256K token context window and support for popular inference engines like SGLang, vLLM, and Transformers, it already outperforms many proprietary alternatives on AgentWorldBench evaluations.

m1ckc3s/claude-status-bar

A lightweight macOS menu bar app written in Swift that shows the real-time status of your Claude Code session. It displays an animated icon while Claude is working, an elapsed timer so you can see how long an operation has been running, and a clear idle indicator when it's waiting for your input.

krea-ai/krea-2

Official inference code and open weights for Krea 2, a text-to-image model trained from scratch by Krea AI. The repo ships two checkpoints: Raw (the full-capacity base model suited for fine-tuning and LoRA training) and Turbo (a distilled variant that generates images in just 8 steps). Artificial Analysis ranks it first among text-to-image models as of June 2026.

BohemiaInteractive/CWR

Bohemia Interactive published the full source code of Arma: Cold War Assault, their 2001 military sim originally titled Operation Flashpoint: Cold War Crisis. The engine has been brought up to C++20 and built with CMake and Clang, running on both Windows x64 and Linux x64. Game assets stay under a separate Arma Public License; the engine itself is GPL v3.

ksimback/looper

Looper is a Claude Code skill that forces a design step before any agent loop runs. It interviews you about your goal and success criteria, then generates a portable YAML spec and a Python runner. The runner comes with cross-model review judges, iteration caps, no-progress detection, and a budget limit — plus an ASCII flow diagram so you can see the whole thing before anything executes.

kanavtwtgg/birds.cafe

A browser simulation built with WebGL and Three.js where you pilot a flock of seagulls over the open ocean. The flock holds a V-formation as you steer with arrow keys, while the scene cycles through day, night, rain, lightning, and storm. It comes with ambient music and works on mobile. The project appeared on June 21 and crossed 731 stars in days — entirely through word of mouth from people who found it and wanted their friends to see it too.

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 system for badminton. It uses YOLO models to detect the shuttlecock and players, RTMPose to build skeleton keypoints, and then generates annotated video, player heatmaps, and movement statistics including distance and speed per rally. The project hit nearly 400 stars within days of its release — a rare result for a tool this sport-specific.

sums001/Windows-Copilot-API

A Python library and local FastAPI server that wraps Microsoft Copilot's web interface into an OpenAI-compatible REST API, giving access to GPT-4 and GPT-5 without an OpenAI account or any payment. It handles Cloudflare challenges, rate limits, and comes with Docker support. Picking up 408 stars in four days shows just how hungry people are for free model access.

SakanaAI/fugu

Fugu dynamically routes a task across multiple large language models — each best suited to a particular subtask — and synthesizes their outputs into a single answer. The coordination logic is handled by two trained models: TRINITY, which evolved as an LLM coordinator, and Conductor, trained with reinforcement learning to orchestrate agent-to-agent communication. The whole system is accessible through the Sakana API.

raiyanyahya/recall

Recall is a Claude Code plugin built to solve the cold-start problem. Every new session begins with no memory of previous work, forcing you to re-explain the project from scratch. The plugin logs your activity during sessions and compresses it into two files — a running history log and a regenerated context summary — stored in a `.recall/` folder inside your repository. Summarization runs entirely locally using Python's TF-IDF and TextRank algorithms: no API calls, no external services, no tokens spent.

overflowy/make-look-scanned

A command-line tool written in Go that processes every page of a PDF through a series of degradation filters: slight skew, grayscale, warm paper tone, scanner grain, edge shadows, and JPEG compression artifacts. The result is indistinguishable from a document you actually printed and ran through a scanner. It also compiles to WebAssembly, so you can run the same transformation entirely in the browser without installing anything.

lyra81604/zhengxi-views

An Agent Skill that packages 14 years of public investment commentary from Zheng Xi, a prominent fund manager at EasyFund, into a fully traceable AI system. Every answer comes with source quotes from the original texts — no invented claims. The skill also pulls real performance data on 8 of his funds and 27,000+ market funds, scoring them by his own investment framework.

vercel/eve

Eve is a new open-source framework from Vercel for building durable AI agents, with an unusual filesystem-first design: an agent's capabilities, prompts and tools live in plain files and folders instead of being hidden in code. That makes an agent easy to read, change and run in production. It topped the week with over 2,100 stars almost immediately.

zhongerxin/Cowart

Cowart adds an infinite drawing canvas (built on tldraw) right inside the Codex AI assistant, so you can sketch ideas, drop in images and edit them step by step while the AI helps. Everything stays in your project folder rather than on someone else's server. It jumped to nearly 1,700 stars in days.

rebel0789/codexpro

CodexPro lets you point ChatGPT's Developer Mode at your own codebase and use it as a hands-on coding agent, connecting through the open MCP standard. It opens a secure tunnel so ChatGPT can read files, search code and run safe commands — but only inside the project you allow. It crossed 650 stars this week.

Forsy-AI/agent-apprenticeship

Agent Apprenticeship is a framework for letting AI agents learn from real work: it collects the step-by-step traces of tasks agents actually do, turns them into reusable experience, and lets people share those lessons. It ships a CLI and a seed set of 500+ real-world tasks, and works with tools like Cursor and Claude Code. It picked up over 600 stars in its first week.

Plaer1/junction

Junction is a VS Code sidebar that connects your editor to local AI coding agents, so you can chat with them without leaving your workspace. Its trick is supporting many different agent backends through one unified panel, so you're not locked into a single tool. It reached over 500 stars this week.

ngrok/webernetes

Webernetes is a TypeScript project from the ngrok team that ports a working subset of Kubernetes into the browser — no backend required. It simulates Pods, Deployments, ReplicaSets, Services (ClusterIP and NodePort), rolling updates, and even inter-pod HTTP networking with DNS resolution. The team built it to power interactive Kubernetes tutorials without maintaining real cluster infrastructure.

cloudflare/security-audit-skill

A coding-agent skill from Cloudflare that orchestrates a six-phase security review pipeline: reconnaissance, vulnerability hunting, validation, reporting, structured output, and independent verification. Each finding is checked by a separate agent before making it into the final report, reducing false positives significantly.

baidu/Unlimited-OCR

Baidu's Unlimited-OCR extends DeepSeek-OCR to handle multi-page documents and PDFs in a single model call, rather than forcing you to feed pages one by one. It supports batch inference, streaming API responses, and integrates with standard inference frameworks. Nearly 400 stars in a few days after release.

anthropics/launch-your-agent

Launch Your Agent is an official Anthropic skill for Claude Code that takes a founder or builder from an idea to a live Claude Managed Agent. You install it, type /launch-your-agent, and it interviews you about what the agent should do, scopes a v0, and launches it into your own Anthropic account using your API key. It then grades the agent's runs against the success criteria you set, suggests iterations, and can schedule the agent if the task is recurring. At the end you get a working managed agent plus a my-agent/ folder with a resumable launch script, an evaluation scaffold, and a notes file for v1/v2 planning. It's published as a reference implementation under Apache-2.0 and pulled close to 400 stars in its first week.

aidenybai/cnfast

cnfast is a TypeScript library that replaces the `cn` utility function used in virtually every Tailwind CSS codebase. It achieves 3.8× average and up to 7× peak speedup over tailwind-merge by applying call-site identity caching and V8 argument caching — the output stays byte-identical, only the performance changes. The author, Aiden Bai, previously created million.js, a virtual DOM optimizer with a large following.

boogu-project/Boogu-Image

Boogu-Image is an open-source, Apache-2.0 family of image generation and editing models built on PyTorch. It ships as three variants: a Base model that emphasizes image diversity and dense, readable text rendering; a distilled Turbo model that produces photorealistic results in just four steps; and an Edit model for image-to-image changes and transformations. It handles both Chinese and English text inside images, and the team published a 'Boogu Arena' comparison putting it in competitive ELO range against leading commercial systems. The headline claim is near-closed-source quality trained on an order of magnitude less data. It pulled close to 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 tool that extracts text from images and PDFs using Apple's own Vision framework — the same recognition engine baked into macOS. Everything runs locally and nothing is uploaded, and it can also rebuild a scanned PDF into a searchable one with a real selectable text layer. It works both as a CLI and as a Node.js library, and ships an agent skill so AI assistants can read documents through it.

fivetaku/fablize

Fablize is a Claude Code plugin that enforces a verification routine rather than boosting model capability. It makes the model run and observe what it built — HTML, code, charts — before calling a task complete, reproduce issues step by step instead of guessing, and refuse to mark complex work done without proof. It grew out of a controlled comparison of two frontier models where the real difference turned out to be procedure, not raw intelligence.

eli-labz/Third-Eye

Third Eye is an open-source geospatial dashboard, built on Next.js and MapLibre, that pulls live public feeds — aviation, maritime, wildfires, seismic activity, weather, satellites, news channels and more — onto a single GPU-accelerated map. It runs on free, public data sources and needs no API keys to get started, and it ships with extras like a recon toolkit and crypto-wallet tracing. The whole thing is self-hostable with prebuilt Docker images, and there's a live demo to try first.

dzcmemory-web/bazi-ziwei-skill

This is an agent skill for two traditional Chinese divination systems — Ba Zi (Four Pillars of Destiny) and Zi Wei Dou Shu (Purple Star astrology). Its core move is to separate the math from the meaning: deterministic algorithms compute the chart from a birth date, while the language model is reserved for interpretation. It offers three modes — each system on its own or a combined cross-check — and can output either a detailed markdown write-up or an ink-wash-style HTML poster. It's written mostly in TypeScript and follows the SKILL.md standard, so it plugs into Claude, Codex, Cursor and similar agents.

duolahypercho/fusion-fable

Fusion-Fable is a Claude Code skill that turns a single question into a small panel of AI models. It sends the same prompt to several frontier models — configurations include two Opus 4.8 runs, Opus paired with GPT-5.5, or a three-way mix that adds Gemini 3.1 Pro — and each one works independently with web search and shell access. Opus 4.8 then acts as a judge, mapping out where they agree, where they contradict, and what unique points each surfaced, before writing one synthesized answer.

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

A static recompilation that converts the Xbox 360 version of GoldenEye 007 into a native Windows executable. Running natively rather than under emulation, it adds mouse-and-keyboard support, online multiplayer, post-processing effects, graphics options and frame rates beyond the original. The repository ships no game assets — you supply your own legal files.

royalbhati/sqltoerdiagram

A browser-only tool that turns SQL CREATE TABLE statements into an interactive ER diagram. It has its own SQL parser, canvas rendering tuned for hundreds of tables, and a built-in editor where changes to the code update the diagram live. Hover highlights relationships, clicking pins focus, table positions persist between sessions, and you can export to PNG or SVG. It blew up because it kills a real annoyance — visualising a schema — with zero friction and no data leaving your machine.

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

A browser-based editor for building 2D side-scrolling game scenes. You arrange sprites, backgrounds and props as layers, import or generate spritesheets, preview the animation and save reusable assets right on your own machine — no heavy desktop animation software required. It was built with AI coding help (the author calls it 'Codex-assisted'), and the whole thing runs locally with a React front end and a small server behind it. A dedicated page lays out every animation clip used in a scene, so you can see all the moving parts at a glance.

mrtooher/fable-mode

A free skill for Claude that forces the assistant to slow down: write a plan first, split the work into stages it can verify, hand independent pieces to cheaper helper models when possible, then critique its own output before handing it back. It tries to copy the disciplined, multi-step way the new Fable models work, but layered on top of the model you already have. The author is upfront that it shapes how the work is organized rather than making the model smarter, so it pays off on big, multi-file or cross-source tasks and is needless on simple ones. It shot to the top of the weekly chart because reliable, fewer-mistakes AI workflows are exactly what people are chasing right now.

EEliberto/IPA-Download

A macOS app (Apache-licensed, Swift) that downloads historical versions of iOS apps and transfers them to your iPhone or iPad via AirDrop for install. It auto-detects your App Store region, can pull builds from several sources or Apple directly, handles two-factor sign-in, and keeps your credentials in the iCloud Keychain. It blew past a thousand stars in its first week.

vorssaint/vorssaint-utils

Vorssaint packs a pile of small Mac tools into a single menu-bar icon: a keep-awake timer (even with the lid closed), live graphs for CPU, GPU, memory, temperature and battery, a network speed test, a per-app volume mixer, a window switcher with thumbnails, a file shelf for quick transfers, an app uninstaller that hunts down leftover files, and a keyboard-cleaning lock. It's GPL-licensed, installs through Homebrew and is Apple-notarized with a Developer ID signature. Folding what's usually five separate paid apps into one free one is exactly why it's climbing so fast.

fguzman82/gateGPT

gateGPT is a complete tiny GPT implemented in Verilog and run on a Xilinx Virtex-5 FPGA, so the model lives in dedicated hardware rather than as software on a CPU or GPU. It is a single transformer block with attention, a key-value cache and fixed-point math, generating human names one character at a time. At an 80 MHz clock it reaches roughly 56,000–69,000 tokens per second, and every stage was verified to match a Python reference exactly.

DanMcInerney/architect-loop

architect-loop is a Claude Code skill that coordinates two AI models with different strengths: Claude acts as the architect that plans slices of work, locks in acceptance criteria and reviews the result, while a Codex builder does the implementation in its own isolated Git worktree. The architect runs the checks itself instead of trusting the builder's report, and the Git repo holds the shared state so handoffs stay lightweight. There is also a research mode where a scout maps the problem and parallel researchers gather sources under strict limits.

alchaincyf/loop-engineering-orange-book

A free educational book (Chinese + English PDF) about "loop engineering": a pattern where instead of prompting an AI agent by hand each time, you design an autonomous system that runs the agent repeatedly on a schedule. It lays out the core moves, the building blocks of a loop, and walks you through making your first one. It went viral because several well-known engineers — Addy Osmani at Google, Boris Cherny at Anthropic, Peter Steinberger — independently named the same shift within one week.

nolangz/pixel2motion

pixel2motion is a Codex/Claude skill that turns a PNG or JPG logo into a clean, motion-ready SVG and then animates it with CSS. It treats each part of the logo — the mark, the dots, the wordmark — as a separately controllable element, so reveals look deliberate rather than random. It also ships HTML demos with playback controls plus Python utilities that audit the vector paths and capture frames to prove the animation actually looks right. In a few days it gathered close to 600 stars.

joeseesun/qiaomu-goal-meta-skill

qiaomu-goal-meta-skill tackles the most common reason AI agents fail: not the model, but a vague instruction. You hand it a loose request and it returns a copy-ready task spec that states the desired outcome, how to verify it's done, what's off-limits, and when to stop instead of looping forever. It fills in conservative defaults rather than handing you an empty form, flags risky cases like touching credentials or production data, and lints out unexecutable phrases like 'keep trying'. It crossed 600 stars within days.

coder/boo

Boo is a terminal multiplexer like GNU screen: you start a session, detach from it, and reattach later with everything intact, even after a dropped connection. What sets it apart is the foundation — it's built on libghostty, the VT core from the Ghostty terminal, so screen state is preserved and redrawn faithfully. It's written in Zig and talks to its daemon over Unix sockets. The reason it took off is its automation story: commands like send, peek and wait let you script a session without needing a real TTY.

alchaincyf/fanbox

FanBox is a desktop workspace for driving AI coding agents like Claude Code and Codex. It fuses three things that usually live in separate windows: a file browser on the left, an embedded terminal on the right, and a live view of changes in the middle. As an agent edits files, they highlight in place, and git diffs show exactly what was altered. It's an Electron app using xterm.js for the terminal and Monaco for code, and it's local-first, so your data stays on your machine. It took off because it makes an agent's work visible instead of a black box.

DietrichGebert/ponytail

Ponytail is a plugin for AI coding agents that makes them behave like a jaded senior who hates writing new code. Before generating anything, the agent runs a checklist: does this need to exist, can the standard library do it, is there a native feature, is a dependency already installed, can it be one line. The repo publishes benchmarks claiming 80–94% less code, 3–6× faster runs, and noticeably lower cost across several models. It plugs into Claude Code, Codex, Gemini, and others.

XiaomiMiMo/MiMo-Code

MiMo-Code is Xiaomi's open-source terminal coding assistant, built as a fork of OpenCode. Its headline feature is persistent memory: it stores project context and architectural decisions in local SQLite with full-text search, so it remembers your codebase across sessions instead of starting cold every time. It switches between build, plan, and compose modes, spawns subagents for parallel work, and makes automatic checkpoints.

omnigent-ai/omnigent

Omnigent is a 'meta-harness' that puts one common layer over different AI agents — Claude Code, Codex, Pi, and your own. You can swap or combine them without rewriting, run several inside one live session, and rein them in with policies and sandboxing (spend caps, approval gates). The session syncs across terminal, browser, and phone, so you can supervise agents from any device.

MSNightmare/RoguePlanet

RoguePlanet is a proof-of-concept exploit for a race-condition bug in Windows Defender that escalates a normal user to a SYSTEM shell. The author is candid that it's flaky — 100% reliable on some machines, hit-or-miss on others — and lists exactly what it was tested on, including patched Windows 10 and 11. It doesn't work on Windows Server, where standard users can't mount ISOs.

SkyBlue997/enableMacosAI

This project unlocks full Apple Intelligence on Chinese-region Macs running macOS 27 on Apple Silicon, where Apple disables it by region. A small kernel extension changes the device's region code at the hardware-registry level from China to US, so every process sees a US Mac instead of patching apps one by one. It installs with a single command and ships detailed troubleshooting for Apple's security layers.

tamnd/kage

Kage makes truly offline copies of websites. It renders each page in headless Chrome, captures the final DOM, strips out all JavaScript, and downloads CSS, images, and fonts locally — so the mirror actually works without scripts or a connection. It can pack everything into a single ZIM file (Kiwix-compatible) or a self-contained viewer, and crawls politely with resume support.

lenucksi/aur-malware-check

After the June 2026 supply-chain attack on the Arch User Repository — where more than 1,600 packages were poisoned via malicious npm dependencies that stole developer credentials and CI secrets — this repo consolidates the community's scattered detection scripts into one toolkit. It ships a scanner with clear clean/warning/infected exit codes and a list of the roughly 1,600 compromised packages, pulled together from gists and mailing lists.

plannotator/effective-html

Effective-html is a set of agent skills for turning a prompt into a clean, self-contained HTML artifact — a plan, an architecture diagram, or a one-off visual. It bundles three skills (html, html-diagram, html-plan) that you install into Claude Code or Codex, leaning on an 'SVG-first, minimal prose' style so the output looks tidy without manual design work.

levy-street/world-of-claudecraft

World of Claudecraft is a playable mini-MMO inspired by classic World of Warcraft, built largely with AI assistance. It has three zones, nine classes with authentic vanilla-style abilities, parties, duels, a ranked arena, and 5-player dungeons — running online with PostgreSQL accounts or offline in the browser, both off the same simulation core. Assets are procedurally generated, so there are no external files, and it self-hosts in one Docker command.

orange2ai/renwei-writing

Renwei-writing ('human-flavored writing') is an agent skill that edits text without scrubbing the person out of it. It was born from a familiar frustration: each AI pass made the writing cleaner but a little less human. The repo offers a post-edit checklist — adapted from Wikipedia's guidance on spotting AI text — plus before/after case studies to help your authentic voice survive the polish.

pewdiepie-archdaemon/odysseus

The loudest thing on GitHub right now is Odysseus, a self-hosted AI workspace: chat, agents, deep research, email and calendar assistance, persistent memory — all running on your own hardware, no cloud account anywhere. It collected about 68 thousand stars in ten days, which almost never happens. Part of the rocket fuel is the mystery: the account is called pewdiepie-archdaemon, and whether that's really THE PewDiePie, the README never says.

simplifaisoul/osiris

OSIRIS is a real-time world intelligence dashboard: live aircraft and ships, 2,000+ public CCTV cameras, earthquakes, fires, news streams, even satellites — sixteen toggleable layers on one fast WebGL map. About 5 thousand stars in a month, and the pitch 'a Palantir alternative' did a lot of the marketing.

vercel-labs/zerolang

Zerolang is the strangest one: an experimental language from Vercel Labs where the program is a semantic graph, not text files. An agent doesn't guess line numbers — it queries the graph and submits checked patches the compiler can reject. Around 5 thousand stars, because everyone senses the question behind it: if AI writes most code now, why does code still look like it's made for human eyes?

shadcn/improve

improve appeared yesterday and grabbed 1.6 thousand stars in a day — shadcn's name does that. It's an agent skill: your most capable model audits the codebase, vets its own findings, and writes detailed plans into a plans/ folder; cheaper models then execute those specs. Intelligence where it compounds, labor where it's cheap.

cpaczek/skylight

Skylight decodes radio signals from real aircraft overhead with a $30 RTL-SDR dongle and a Raspberry Pi, and projects them onto your ceiling — with the moon, stars and the ISS in their true positions. An X-ray through the roof. 2.6 thousand stars in about a week, because everyone who sees the demo makes the same sound.

MisoLabsAI/MisoTTS

Miso TTS is an open 8-billion-parameter text-to-speech model that aims at emotive, conversational speech and can clone a voice from an audio prompt. 2.7 thousand stars. The catch: you need a 24 GB GPU to run it comfortably — or just try the hosted demo first.