GitHub's week: everyone rebuilt Jev in their garage

A browser agent from the Browser Use crew that drops the screenshot loop entirely. Instead of photographing the page and asking a vision model what to do, it keeps a live index of the page's elements and asks a typed-decision model two things in one request — which action, and which element — with a small text model waking up only when something actually has to be typed. The demo people kept resharing is a full Google Flights search finishing in about seven seconds. It ran away with the week because it is both far faster than the screenshot approach and small enough to read end to end.

A Claude Code plugin that replaces the built-in compaction step. Rather than summarising the old part of a session, it keeps the transcript verbatim and asks a decision model two yes/no questions about every tool call — keep the call, keep its result — then drops or truncates on the answers. The argument in its favour is that a summary is lossy in exactly the wrong way: the file path, the precise error text, the constraint you agreed on twenty turns ago are the first things to evaporate. It went straight up because everybody running long agent sessions has been burned by that at least once.

An open, Apache-2.0 typed-decision model from Convai Innovations, on Hugging Face and installable with pip. It answers three shapes of question about a piece of text — pick one of these labels, score it on this scale, how likely is this true — in a single pass with no generation at all, around 33ms on a GPU. There are three checkpoints: English on ModernBERT-large, multilingual on mmBERT-base covering 100+ languages, and one tuned for structured workflows, with a router that detects language and script and picks for you. By its own published benchmark it beats the closed alternative on accuracy and runs seven to eight times faster, which is what carried it this high.

A native Mac image editor aimed squarely at Photoshop, MIT-licensed, built by opening the Xcode project. It covers the load-bearing parts: layers with blend modes, masks and adjustment layers, non-destructive transforms, magic wand and content-aware fill, healing brush and clone stamp, filters and blurs, tabs, and JPEG, PNG, HEIC and TIFF. The author's stated reason for building it is that Photoshop costs too much and GIMP never felt familiar, so the keyboard shortcuts deliberately match Photoshop's. It needs macOS 26.

Z.ai's own coding-agent harness, published under Apache-2.0. You get an Electron desktop app, browser and terminal front-ends and an agent CLI, plus the ability to work on a remote machine over SSH or WSL. The interesting part is that it is the harness rather than a model: provider configuration is built in and you can point it wherever you like. Eight hundred forks in five days is the tell — a lot of people want a Claude Code-shaped tool whose insides they are allowed to change.

Semantic ifs: an if-statement whose condition is judged by a language model instead of a comparison operator. You hand it an unstructured situation, the criteria that matter right now, and the typed options; a frozen 4B model reads the probabilities for exactly those options out of one forward pass, so nothing is sampled and no answer is written. That lands at roughly a second per decision against five-plus for asking a model to emit JSON, and about twenty times the throughput when many decisions share the same context. It runs on a 3090 at home, has an MLX backend for Apple Silicon, and a WebGPU demo that needs no hardware at all.

Automation for ByteDance's Jianying — CapCut's professional edition — on macOS. You feed it media plus a structured editing plan and it writes a real, editable native project: multiple tracks, speed changes, volume, subtitles, titles, basic keyframes and native effects, then exports MP4 through the app's own engine. It ships an Agent Skill so an AI agent can drive the whole pipeline rather than a human clicking through it. Two things to note: the licence is personal and non-commercial, with no CapCut integration rights granted, and the repo carries 1,420 forks against 2,107 stars, which is an unusually forky profile for a week-old project.

A native MLX runtime for Laya's typed-decision models, so they run on Apple Silicon without PyTorch, without a server and without a cloud API. The reported medians are around 13ms end to end for a short English decision and about 7ms on the multilingual model on an M3 Max, in under a gigabyte of memory. There is a Snake demo where the local model drives the game in real time with no visible lag, which is a blunt but effective way to show what that latency feels like.

A market-making bot that takes one AI decision per Monad block — roughly every 300 milliseconds — on Kuru's MON-USDC pair. Each block it reads the order book over RPC, asks the decision model where the price goes over the next half-minute, and posts a limit order one tick inside the spread, so it earns the spread instead of paying it. The default mode is a dry run: real market data, simulated fills, no funds at risk. Add a private key and it trades for real.

A from-scratch miniature of Jev on a Qwen3-0.6B backbone, with parallel decisions, dynamic candidate sets and the entire training pipeline out in the open. It trains on roughly 18,760 decision questions across four game environments — ViZDoom aiming, ViZDoom moving targets, a 50x50 maze and Snake — blended by weight into a single checkpoint that plays all four. It is a nanoGPT-shaped teaching artefact: small enough to read in an evening, cheap enough to retrain on something of your own.