AI Open Source Radar — November 4, 2026

Data sources: GitHub Trending bot dumps — today’s Rain1er/trending #425 (2026-11-03) and kouweizhu/agents-radar #780 (2026-11-03 03:40 UTC), compared with the previous day’s agents-radar #766 (2026-11-02 01:45 UTC) and happydog-intj/ai-xiaohongshu-daily #617. Ranked by stars today; repos above 40k total stars, long-running top-ranked repos, decelerating repos and non-AI repos are excluded. Descriptions come from each README; numbers come from the bot issues.

vercel-labs/agent-desktop — +3,792 ⭐ today

https://github.com/vercel-labs/agent-desktop
Total 12,223 stars · vs. previous day: not listed yesterday, +3,792 today (Rain1er) / +2,275 (agents-radar) — new entry · Rust · Apache-2.0

A computer-use CLI and C library that lets AI agents read and operate desktop apps through the operating system’s accessibility tree. The agent calls snapshot to get JSON with a ref for every actionable element, acts on a ref (click, type, set value, scroll and more), then takes a fresh snapshot to verify the result, so no pixel matching is needed. Key components:

The agent-desktop binary distributed via npm (x64 and ARM64), plus the in-process C library libagent_desktop_ffi
Platform adapter crates for macOS, Windows and Linux (AT-SPI); headless by default without moving the pointer, with --headed for physical input
60 commands across observe, ref actions, keyboard/mouse, apps/windows, notifications, clipboard, wait and sessions, each returning an ok/error JSON envelope with retry guidance
Built-in, version-matched agent skill docs (skills get desktop)

It is faster and more deterministic than screenshot-based computer use, and the fact that Vercel shipped it as a general-purpose agent tool helps explain thousands of stars on day one.

Practical takeaway: A structured execution layer that can replace RPA when automating legacy desktop work (ERP data entry, repetitive internal tools) with agents.

#ComputerUse #AccessibilityTree #DesktopAutomation #AgentCLI #RustLang

FlashML-org/FreeToken — +653 ⭐ today

https://github.com/FlashML-org/FreeToken
Total: not shown by the bots (about 15k per README, for reference) · vs. previous day: +242 → +653 (+170%) · Python · Apache-2.0

An edge-native Mixture-of-Experts (MoE) serving engine that runs huge open-weight MoE models on consumer hardware such as gaming PCs and laptops. It treats GPUs, CPUs, host memory and interconnects as a single inference pool and loads experts on demand to get around VRAM limits. Core techniques:

Bandwidth-adaptive CPU–GPU co-execution and full-layer double-buffered prefill streaming
Global LRU expert caching, the FTW fast weight format and graph-compatible execution
Semantic anchor checkpoints that avoid recomputation when agents edit context (tool calls, thinking blocks)
Elastic memory management that reallocates VRAM between expert cache and KV memory at runtime
OpenAI- and Anthropic-compatible APIs for Codex, Claude Code, OpenCode and other coding agents; RTX 30/40/50 and MXFP4, NVFP4, FP8 support

Daily gains more than doubled as developers look for ways to run DeepSeek-V4-Flash- or GLM-5.2-class MoE models as local backends for coding agents.

Practical takeaway: A good fit for proofs of concept that attach a large MoE coding agent to on-prem workstations for clients that cannot send data to cloud APIs.

#MoEInference #LocalLLM #ExpertOffloading #ConsumerGPU #LLMServing

Donchitos/Claude-Code-Game-Studios — +1,379 ⭐ today

https://github.com/Donchitos/Claude-Code-Game-Studios
Total 34,453 stars · vs. previous day: +967 → +1,379 (+43%, agents-radar; Rain1er shows +1,181) · +2% vs. its last feature on Nov 1 (+1,347) · Shell (Markdown/YAML definitions) · MIT

A template that turns a single Claude Code session into a full game development studio. Forty-nine agents collaborate across three tiers — directors (creative, technical, producer), department leads (game design, lead programmer, art, audio, narrative, QA, release, localization) and specialists — and nothing is written until the user approves. Components:

70+ slash-command skills such as /start, /brainstorm, /team-combat and /team-release
15 hooks on commits, pushes, asset changes and design-doc reads/writes, plus an agent audit trail
14 path-scoped coding rule sets (gameplay, networking, shaders and more) and 62 document templates (GDDs, ADRs, sprint plans)
Dedicated agent sets for Godot 4, Unity and Unreal Engine 5; process rigor adjustable via modes.rigor in project.yaml

After a one-day dip (+970) following its Nov 1 feature, it climbed back to its previous level, showing sustained interest in multi-agent “virtual organization” templates.

Practical takeaway: Beyond games, the role-based agents + approval gates + hooks pattern is a useful reference for building Claude Code templates for in-house engineering teams.

#ClaudeCode #GameDev #MultiAgent #AgentSkills #GodotEngine

Excluded (summary): claude-mem (111.6k), mattpocock/skills (311k), Scrapling (94.3k), JustVugg/colibri (47.1k) and Kronos (46.4k) are over the 40k-star cap. p-e-w/heretic (38.6k) is a long-running trending repo since 2025 and a decensoring tool. anomalyco/hex slowed from +1,179 to +410 (-65%). max-sixty/worktrunk (+437, 4.2%) and paperclipai/paperclip (+412) fell short of the surge bar. kornia is a perennial large repo; floci and public-apis are non-AI; lightpanda-io/browser is a long-term top-ranked repo. Trendshift badge history for agent-desktop could not be checked (new repo).