Compass halts AI agents at spending cap—no external calls

What happened

Compass is a tool that sits between Claude Code, Codex, or Gemini and your codebase. It enforces a hard spending limit (set via COMPASS_MAX_USD), blocks commands deemed catastrophic before they run, and automatically reviews and fixes pull requests in a loop until tests pass—then stops at you to merge.

Why it matters

Most AI-agent configurations claim to be safe and cheap but don't prove it; Compass hands you measurable numbers. The guardrail policy scores 100/100 in CI (against a 61-case corpus), cost routing achieves ~61% cheaper performance than all-Opus at ~98% quality, and the spending ceiling actually halts the agent's next tool call once the cap is hit—not just a warning. All of this runs locally with no telemetry and no external service calls.

What to watch

Installation is reversible and requires only Claude Code (or Codex/Gemini) plus git; the simplest path is a git clone and ./quickstart.sh, or via the Claude Code plugin marketplace. The autonomous PR loop (which runs in GitHub Actions) needs model auth and a fine-grained PAT token; without it, the loop still runs but won't auto-re-fire after a fix. Releases carry keyless SLSA provenance, verifiable with compass verify v0.17.2.