This week's list shows Jev-style direct decisions moving into browser agents, context compaction, trading bots, and local inference. ZCode and Compositor represent a complete agent workspace and a macOS image editor. I read each pinned README and kept the issue to documented capabilities, setup paths, and boundaries.
The observation window is 2026-09-14 through 2026-09-20. This is an independent review based on GitHub's public API, not GitHub's official Trending list; Stars measure attention, not code quality, security, or long-term maintainability. No listed repository was installed, executed, or audited here.
Top 10 projects
1. browser-use/jev-ultrafast
Core functionality: A browser agent with a dynamic action space. Each observation produces a numbered element table, then Jev chooses CLICK, TYPE_TEXT, SELECT, scrolling, waiting, done, or blocked plus a compatible target. Only TYPE_TEXT calls a small language model for text. The README demonstrates the loop with Google Flights.
Usage and boundaries: Run uv sync, copy .env.example to .env, set TYPESAFE_API_KEY and TEXT_MODEL_API_KEY, then run uv run jev and open http://127.0.0.1:8766; the Python Agent also supports custom URLs and goals. Chrome, Browser Harness, and a compatible text model are required. Model output is not turned into selectors, coordinates, shell commands, or JavaScript, and the executor rechecks freshness and occlusion. The MVP excludes shadow DOM, frames, canvas, uploads, popup tabs, and complex keyboard widgets, and the demo does not book a flight. Its benchmark is limited to one browser profile, not a general reliability claim. MIT.
2. tamaratran/fast-jev-compaction
Core functionality: An npm library and Claude Code function-hook plugin that asks Jev which tool calls and results matter, deleting or truncating stale tool records while keeping user and assistant text verbatim and ordered. It exports state-fitting, batching, decision, application, and custom JevAsker building blocks.
Usage and boundaries: Install with npm install fast-jev-compaction, set TYPESAFE_API_KEY, and call compactMessages; the Claude Code plugin requires 2.1.274+, CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1, marketplace installation, and reload. Only tool calls/results are candidates; text is not removed from output. Token sizes are character estimates, not tokenizer counts, and probability is not proof that deletion is safe. Errors or insufficient reduction fall back to the built-in summary, while long histories repeat the full state across requests. Unit tests use a fake Jev; the demo is the live network check. MIT.
3. NandhaKishorM/laya
Core functionality: A multilingual, non-autoregressive System 1 decision engine that returns choice probabilities, score distributions, or noul truth probabilities in one forward pass. The README provides English, multilingual, and typed-decisions checkpoints, a language/task Router, and prompt-guardrail, content-safety, and support-triage presets.
Usage and boundaries: Install with pip install laya, use Router(preload=True) with router.predict(state, questions), or load a Hugging Face checkpoint with laya.load(...). Preloading uses more memory but avoids rebuilding models on language switches. The checkpoints have different intended uses; the English model can fail badly outside English, and multilingual probabilities still need calibration. Base checkpoints are near chance on typed-decisions zero-shot; 0.766 comes from a fine-tuned benchmark. High-cardinality choices are constrained by head_max_len. Apache-2.0.
4. robbietilton/Compositor
Core functionality: A free, open-source macOS image editor for a Photoshop-style compositing and post-processing workflow. It lists layers, folders, blend modes, masks, adjustment layers, non-destructive transforms, selections, content-aware fill, painting, retouching, text, filters, tabbed projects, and JPEG/PNG/HEIC/TIFF import with JPEG export.
Usage and boundaries: Open Compositor.xcodeproj in Xcode 26 and run the Compositor scheme; macOS 26 is required. The release script additionally needs a Developer ID Application certificate, notarytool credentials, and create-dmg, then signs, notarizes, staples, and packages a DMG. The README makes no Windows or Linux support promise; its documented path is the current macOS build. MIT.
5. zai-org/ZCode
Core functionality: An AI coding workspace containing an Electron desktop app, Web/CLI distribution, terminal Agent CLI, backend, shared UI, TUI, and runtime. zcode starts TUI or Web mode and passes through Agent arguments; Web provides local workspaces, API/WebSocket routes, and SSH/WSL-related workflows.
Usage and boundaries: It requires Git, Node.js 24.14.0, and pnpm 10.33.2; run pnpm bootstrap. Use pnpm dev:desktop, pnpm dev:web, or pnpm --filter @zcode/cli dev; build the distribution with pnpm build:zcode and configure its download base URL. Web defaults to 127.0.0.1 without a token; with --host 0.0.0.0, use the generated token URL, and configure ZCODE_SERVER_AUTH_TOKEN for the generic server entry point. Desktop defaults to macOS arm64, although mac/win/linux and x64/arm64 options exist; actual packaging/signing still depend on target tools and local builds are unsigned. NOTICE.md covers execution, data-risk, and third-party boundaries.
6. TheoLeeCJ/SemIf
Core functionality: SemIf, formerly OpenJev, reproduces a semantic-if interface pattern with open models. It reads runtime state, questions, and options directly from logits and returns option probabilities without answer sentences, JSON repair, or a decoding loop. It includes a WebGPU demo, shared-state reuse, an Apple Silicon MLX backend, and benchmark material.
Usage and boundaries: With Python 3.10+, CUDA, and a GPU that can hold a 4B BF16 model, create an environment, install pip install -e '.[test]', and run semif-score --mode direct --model ... --input ... --output ...; Apple Silicon can install .[test,mlx] and add --backend mlx. Results include probabilities, timing, model revision, and prompt hash. It is independent and not affiliated with Jev or TypeSafe; model weights and third-party sources are not bundled. Probabilities are conditional on supplied options and must be calibrated on the real workload. Shared-state and BF16 fast paths remain experimental and have reported argmax differences from fresh scoring. MIT.
7. mcncarl/jianying-headless
Core functionality: A local automation tool for Jianying Professional on macOS. It turns structured plans into editable drafts, edits multi-track projects in an isolated copy, and uses the local engine to export H.264/AAC MP4. The documented scope includes clips, multiple tracks, speed, volume, picture-in-picture, subtitles, titles, local fonts, and basic keyframes.
Usage and boundaries: It requires Apple Silicon, macOS 26.0+, Jianying 11.5.0 (compatible with 11.4.2), Python 3.9+, FFmpeg/ffprobe, and Xcode Command Line Tools. Build the bridge, run headless_draft.py doctor, then use a JSON plan with build, verify-build, publish, and export; publish registers a draft on the local home page, not online. It is not an official SDK and does not support arbitrary versions, online templates, cloud projects, or account entitlements; compound clips and some image/GIF cases remain limited. The license permits personal learning and non-commercial use; commercial use requires written permission, and no standard SPDX license is declared.
8. mizorewww/laya-mlx
Core functionality: An MLX port of Laya typed-decision inference for Apple Silicon, running choice, score, and noul locally without PyTorch, Transformers, or a cloud API. It provides a Python API, CLI, routing, Snake demo, and checkpoint conversion.
Usage and boundaries: Install with pip install laya-mlx, load aac6fef/laya-mlx, and call agent.predict; the first load downloads the checkpoint and later inference is local. Requirements are Apple Silicon, Python 3.11+, and macOS 14+; development uses uv sync --extra demo, while compilation, prefix caching, and larger batches are optional. The README calls it an independent MLX port, not an official Convai Innovations release; training and fine-tuning remain upstream. FP16 and FP32 may differ, BF16 is outside the published validation matrix, and full GPU benchmarks are not in hosted CI. Apache-2.0.
9. jarrodwatts/jev-trader
Core functionality: A trading bot for the Monad Kuru MON-USDC order book. Each block, Jev or a mock model chooses buy or sell, posts a post-only limit order one tick inside the touch, replaces the previous order, and streams block, order, fill, gas, position, and P&L events through SSE. It supports dry-run simulated fills and /history and /events endpoints.
Usage and boundaries: Copy .env.example, run bun install, then bun run start; without PRIVATE_KEY it uses the real book with simulated fills. Set MODEL=jev and TYPESAFE_AI_API_KEY to use Jev; the default mock is a momentum heuristic. Live trading adds private keys, RPC, orders, gas, and funds. The README says live sends are fire-and-forget: a block event records intent and may become placed, reverted, or lost; late decisions, position caps, or margin can change the quote. It is not a low-risk investment tool; verify contracts, key handling, RPCs, and fund controls independently. MIT.
10. TianyuCodings/NanoJev
Core functionality: A 0.6B parallel decision model that accepts state, questions, and candidates and returns full probability distributions without output-token generation. The README demonstrates one checkpoint across ViZDoom Basic, Predict Position, Maze, and Snake, with Choice, Boolean, and ordered Score heads.
Usage and boundaries: Install requirements-toy.txt and huggingface_hub, download the unified-games-v1 checkpoint and dataset from Hugging Face, start the CUDA service with scripts/serve_decisions.py, and POST batches to 127.0.0.1:8765/api/evaluate; replays use http.server. Results use fixed test sets, observation interfaces, candidate actions, and seeded controllers; Predict Position is much weaker than Basic, and the interactive site currently requires access. Deployment still needs CUDA, weights, data, and the input contract, and game results do not establish general-agent or production decision quality. The pinned README does not clearly state a project license, so confirm permission before adoption.
Data limitations
Stars, forks, and issues/PRs are collection-time snapshots. Descriptions and performance figures come from the projects and were not independently reproduced. No listed code was cloned, installed, or executed. Before adoption, inspect licenses, dependencies, model provenance, credential handling, network permissions, and the latest revision.
Source: GitHub Search API · Collected 2026-09-21 (Asia/Singapore)