Sir — Security skill for Claude Code

sir — Sandbox in Reverse

A local security layer for AI coding agents. When your agent reads your secrets and then tries to phone home, sir stops it, proves it, and tells you exactly what to do next. **Silent while you code; loud only at the exits.**

A sandbox sees a process making a network call. sir sees **why**: the agent read `.env`, an MCP tool said "forward these for analytics," and now it's curling them to an unknown host. To a sandbox that's indistinguishable from `npm install`. sir adds the context that makes the block intelligent.

[!NOTE] sir is experimental and not yet production-ready — test on your own machine, not shared infrastructure. `sir doctor` recovers any wedged state; `sir uninstall` removes hooks cleanly. [Report bugs](https://github.com/somoore/sir/issues).

Quick start

curl -fsSL https://raw.githubusercontent.com/somoore/sir/main/scripts/download.sh | bash
cd /path/to/project
sir install            # auto-detect supported agents already on this machine
sir demo               # see all three detections run, end to end

Then use your agent as normal. sir stays quiet until something crosses a line — and when it does, the message is the fix:

Claude tried to reach evil.example.com — × deny.

  reason: