AI Open Source Radar — November 5, 2026

Data sources: GitHub Trending dump bot issues — today huiyishalou/ai-trending-hub #210 (2026-11-04 16:01 UTC; star counts unlabeled, treated as stars today based on cross-checks) and kouweizhu/agents-radar #794 (2026-11-04 02:38 UTC), compared with agents-radar #780 (2026-11-03 03:40 UTC) and Rain1er/trending #425 (Nov 3). Rain1er #426 and happydog #624 could not be fetched. Primary metric is stars today; repos with 40k+ total stars, long-running chart leaders, slowing repos and non-AI repos are excluded. Descriptions come from each README; numbers come from the bot issues.

NVIDIA/auto-ontology — +3,767 ⭐ today

URL: https://github.com/NVIDIA/auto-ontology
Total: not listed by bots (about 4.4k on the GitHub page, about 200 a week ago — for reference) · vs. previous day: absent yesterday → +3,767 today (new entry) · Python backend + Next.js · Apache-2.0

An open context layer that helps AI agents give fewer wrong answers on enterprise data. It runs as a service next to your databases, compiles their metadata into an ontology (a semantic layer) and has agents query through it, so SQL is generated from what fields mean rather than from column names. Key components and features:

Auto-generates a draft ontology from database metadata, maps business terms such as Customer or Invoice Line to columns and infers table relationships
Stores reviewed SQL expressions for measures so agents compute them consistently; definitions can be reviewed, edited and certified in a web UI or via API
Checks whether a question is answerable from the data before querying and asks for clarification if not; every answer returns its SQL, steps and rows
Next.js frontend, FastAPI + LangChain backend, an ingestion service, a read-only FastMCP server and Postgres + pgvector
Ten sources including PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino and DuckDB; defaults to NVIDIA NIM (Nemotron) models but works with any OpenAI-compatible endpoint

Wrong numbers are the biggest risk in agentic analytics, and NVIDIA shipping a governed semantic layer with an MCP server and agent skills under Apache-2.0 drew thousands of stars in a single day.

Practical takeaway: Worth evaluating as a governance layer that enforces approved metric definitions when agents like Claude or Cursor query revenue and margin data directly from your warehouse.

#SemanticLayer #TextToSQL #KnowledgeGraph #DataAgents #MCPServer

vercel-labs/agent-desktop — +4,490 ⭐ today (2 days in a row)

URL: https://github.com/vercel-labs/agent-desktop
Total: 12,223+ (Rain1er, Nov 3) · vs. previous day: +2,275 → +3,791 (+67%, same bot: agents-radar), latest snapshot +4,490 (ai-trending-hub, 16:01 UTC) · 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 those refs (click, type, scroll) and re-snapshots to verify, so no screenshot pixel matching is needed. Main pieces:

npm-distributed agent-desktop binary (x64 and ARM64) plus a C FFI library for in-process calls
macOS, Windows and Linux (AT-SPI) adapters, headless by default with a –headed mode for real input
60 commands covering observation, ref actions, keyboard and mouse, windows, clipboard, waits and sessions, each returning an ok/error JSON envelope
Built-in, version-matched skill docs for agents

Daily gains grew after launch day, showing sustained interest in structured computer use that is faster and more deterministic than screenshot-based approaches.

Practical takeaway: A good execution layer for PoCs that automate legacy desktop work (ERP entry, repetitive internal tools) with agents instead of RPA.

#ComputerUse #AccessibilityTree #DesktopAutomation #AgentCLI #RustLang

max-sixty/worktrunk — +964 ⭐ today

URL: https://github.com/max-sixty/worktrunk
Total: 10,377+ (Rain1er, Nov 3) · vs. previous day: +437 (Rain1er #425) / +438 (agents-radar #794) → +964 (+120%) · Rust · MIT OR Apache-2.0

A Git worktree management CLI for developers running several AI coding agents at once. Each task gets its own working directory so parallel agents never overwrite each other, worktrees are addressed by branch name and paths come from a template, with the goal of making worktrees as easy as branches. Core commands and features:

wt switch -c <branch> -x claude — 'task': create a worktree and launch Claude Code, Codex or another agent in one step
wt list –full: changes, commits ahead of main, CI status and AI-generated summaries at a glance
wt merge: squash, rebase, merge into main and clean up at once; wt remove deletes the worktree and branch
Hooks on create/merge events, LLM commit messages from diffs, and wt switch pr:123 for PR checkout
Shared build caches (target/, node_modules/) on APFS, btrfs and XFS, plus a unique dev-server port per worktree via hash_port

As running agents in parallel becomes the norm, its daily gains more than doubled after sitting in the +400 range.

Practical takeaway: Teams that run one agent per feature in a single repo can use it right away to standardize branch, worktree and port conflict handling.

#GitWorktree #ParallelAgents #ClaudeCode #DeveloperTools #RustLang

Excluded (summary): Donchitos/Claude-Code-Game-Studios slowed from +1,379 to +1,182 (-14%, same bot). FlashML-org/FreeToken dropped off the list. anomalyco/hex (+410) kept slowing. FareedKhan-dev/kimi-k3-in-c (+653, about 7% of ~9.7k) was not a top non-excluded gainer and its claim of running a 2.78T model on a CPU in 8GB RAM lacks verification, so it is on hold. scenario-labs/skills (+448) fell short of the bar. claude-mem (112k), mattpocock/skills (311k), Scrapling (94k), colibri (47k) and Kronos (46k) exceed 40k stars; p-e-w/heretic is a long-running leader and uncensoring tool; floci is not AI-related.