Skip to content

AI Agent GitHub Digest — 2026-11-05

Nov 5, 20261 min
Home›daily
TL;DRmorluto/rea (⭐430k) wraps engines like Ghidra and Hopper into MCP tools for agents, hitting a new high in three weeks just as reports surfaced that Anthropic has started blocking some reverse-engineering requests in Claude — and even sent a takedown notice to a developer reverse-engineering Claude Code itself; vercel-labs/agent-desktop (⭐14.6k) reads the OS's native accessibility tree instead of screenshots, letting agents operate a computer, and doubled its stars in three days; D4Vinci/Scrapling (⭐96k), a year-old scraping framework, just added a native MCP server so agents can scrape without fighting anti-bot defenses themselves. No major framework release today.
#ai-agent#github#open-source#daily#mcp#computer-use#reverse-engineering
Series: AI Agent GitHub Digest (82 / 81)
AI Agent GitHub Digest — 2026-11-04

Look up & save

Loading dictionary…

🌏 中文版

Today's Highlight

Today's three repos each take a different angle on the same playbook: build the underlying capability first, then wire up an MCP layer for agents — instead of building an "agent-native" tool from scratch. REA wraps existing reverse-engineering engines into MCP tools and hit 430k stars in three weeks, but the same stretch also brought Anthropic's pushback on reverse-engineering requests — attention and policy risk climbing together. agent-desktop continues the growth curve we flagged last week, doubling its stars in three days. Scrapling, a year-old scraping framework, only recently filled in the MCP piece.

morluto/rea ⭐ 430k

GitHub　·　TypeScript　·　MIT

  • What it is: Wraps existing reverse-engineering engines like Ghidra and Hopper into a unified set of MCP tools, letting an agent analyze native binaries, JS/Electron apps, .NET assemblies, and Android APKs — answering "how does this feature work" with disassembly evidence attached, no source code required.
  • Why it's worth watching: We covered this as a standalone tool back in mid-October (see Tool of the Day: REA), and it's climbed to 430k stars in three weeks since. But the same stretch brought reports that Anthropic has started blocking some reverse-engineering requests in Claude Code and Opus 5.5 — reportedly tied to a rule against "recreating a specific existing game" — and sent a takedown notice to a developer trying to reverse-engineer Claude Code itself. A tool that lets anyone reverse-engineer anything has run straight into the gray zone between commercial licensing and copyright law; Cornell Tech law professor James Grimmelmann called it a looming crisis for trade secret and copyright law.
  • Tech stack: TypeScript wrapper + MCP server, calling into existing reverse-engineering engines like Ghidra, IDA, and Hopper
  • Getting started: Medium — installable via npx, but reading the output still requires understanding disassembly, and you need to confirm the target software's license actually permits reverse engineering

vercel-labs/agent-desktop ⭐ 14.6k

GitHub　·　Rust　·　Apache-2.0

  • What it is: A computer-use CLI for agents that reads the component tree through macOS/Windows/Linux's native accessibility APIs, returning a JSON snapshot with stable element refs — instead of making the model guess coordinates from a screenshot.
  • Why it's worth watching: We covered this in the Nov 2 digest at 6.5k stars; three days later it's at 14.6k, more than doubling. The advantage of accessibility-tree-based control over screenshot-based control is that element refs stay stable, so a failed action can just be retried instead of re-identifying coordinates from scratch.
  • Tech stack: Rust core + native accessibility APIs per platform, exposed through both a CLI and a C library
  • Getting started: Low — installable via npm, Homebrew, or direct download, with categorized commands for observation, action, keyboard/mouse, and app management ready to go

D4Vinci/Scrapling ⭐ 96k

GitHub　·　Python　·　BSD-3-Clause

  • What it is: A web scraping framework that handles everything from a single request to a full-scale crawl, with a built-in stealth browser for bypassing defenses like Cloudflare, CSS/XPath parsing, and an interactive shell for pulling data without writing code.
  • Why it's worth watching: This project has been around for over a year, climbing toward 100k stars on parsing speed and anti-bot capability alone. It recently added a native MCP server, so an agent can issue a scraping command directly instead of having to work around anti-bot defenses itself. Like REA and agent-desktop, it's taking the "build the underlying capability first, then bolt on an MCP layer for agents" route.
  • Tech stack: Python, with a built-in stealth fetcher (bypassing defenses like Cloudflare), concurrent crawling with session/proxy rotation, and a native MCP server
  • Getting started: Low — pip install and the CLI or interactive shell can pull data right away; the concurrent spider setup is only needed for more advanced use

Notable Releases

No major framework update today. CrewAI 1.15.35 (Nov 3) only updated the LiteLLM model catalog docs; LangGraph SDK 0.4.12 (Nov 3) just removed the fairness_weight field from PrioritySettings and fixed a minor stream-reconnect bug — neither clears the bar for "significant new feature or breaking change."

Today's Takeaway

I used to score "how good is an agent tool" purely on technical capability — how many formats it can parse, how fast it runs. REA's controversy this week is a reminder that once a tool makes reverse engineering something anyone can do, the next bottleneck usually isn't technical — it's the legal line around copyright and licensing. Anthropic's restriction on reverse-engineering requests changed what this tool can actually be used for faster than any model capability upgrade has.

References

Related

AI Agent GitHub Digest — 2026-11-071 minAI Agent GitHub Digest — 2026-10-121 minAI Agent GitHub Digest — 2026-11-041 min