Luthien Proxy

Luthien

Claude Code builds. You stay in control.

What does it look like? | What can it do? | How does it work? | Quick start

Open-source proxy that sits between Claude Code and the Anthropic API.
Logs every request. Enforces your rules.

What does it look like?

Say your CLAUDE.md has this rule:

Without Luthien
With Luthien

Claude ignores your CLAUDE.md rule and you correct it manually.

Luthien catches the violation and auto-corrects. No human intervention needed.
:rotating_light: Luthien is in active development. Star this repo to follow updates, or Watch > Releases to get notified on new versions.
Found a bug or have a question? Open an issue.
Want to try it and help us out in the process? The feedback guide explains what kind of feedback is most useful.

What can it do?

Enforce arbitrary policies
Use Luthien to define policies and run them on every request and/or response. What it catches depends on the rules you write.

- Rewrite pip into uv: a drop-in preset (PreferUvPolicy) that corrects pip install to uv add in every response from Claude.
- Plain-English rules, checked by an LLM-as-Judge: write your rules in English, and a second LLM (SimpleLLMPolicy) checks every response from Claude against them and rewrites or flags anything that doesn't comply.
- Enforce scope boundaries: only allow changes to files mentioned in the request.
- Catch subtle test cheating: not just deleted tests. Flag tests that accept any response code (assert response_code in [200, 302, 400, 404, 500]), or except: return {"mock": "data"} fallbacks that hide failures so no test ever fails.
- Clean up AI writing tics: remove em dashes, curly quotes, over-bulleting, and excessive comments.

Example: an LLM-as-Judge on every response from Claude (SimpleLLMPolicy). You write rules in plain…