Developer-grade Claude Code + Codex configuration: cost-tiered subagents, workflow commands, guardrail hooks, MCP parity, and an installable plugin/marketplace.
README
Anyone can say "safe" and "cheap." compass hands you the number — and lets you reproduce it in 30 seconds: guardrails 100/100 on a 61-case bypass corpus, a router measured ~61% cheaper than all-Opus at ~98% quality, signed releases you verify in one command. One config you own for every agent, in every repo — not a service. No curl | sh, no telemetry. You always merge.
In plain terms: a tireless senior teammate for your AI coding agents — it reviews and fixes its own work, spends your money wisely, refuses the dangerous stuff, and still can't ship anything without your yes.
</div> <p align="center"> <a href="docs/01-architecture.md" title="placeholder link — repoint to a demo/video/landing"><img src="assets/explainer.svg" alt="compass in three beats: ONE CONFIG (install once) → EVERY AGENT (Claude Code · Codex · Gemini · Cursor, one AGENTS.md, no drift) → AUTONOMOUS PRs (reviews · fixes itself · you merge). All opt-in: guardrails · red-team hardening · cost-tiered router · subagents/commands/MCP · scheduled fleet · human merge gate." width="900"></a> </p> <p align="center"> <a href="#-the-part-people-screenshot-it-fixes-its-own-prs">▶ Watch it fix its own PR</a> · <a href="#why-its-differe