ui
without luthien
Claude ignores your CLAUDE.md rule and you correct it manually.
with luthien
Luthien catches the violation and auto-corrects. No human intervention needed.
Open-source proxy that sits between Claude Code and the Anthropic API.
Logs every request. Enforces your rules.
Claude ignores your CLAUDE.md rule and you correct it manually.
Luthien catches the violation and auto-corrects. No human intervention needed.
🧑💻 You ⇔ 💻 Claude Code ⇔ Luthien ⇔ ☁️ Anthropic API | logs every request and response. ~5-15ms you can configure rules/policies to modify or block certain responses or requests: | |-- did it do what I asked? |-- did it follow CLAUDE.md? +-- did it do something suspicious?
Luthien can call a separate model to check whether each response follows your rules. Runs alongside normal requests, adding almost no delay.
When Claude Code compacts or starts a new session, Luthien still remembers. Your rules stay enforced from first prompt to last.
How to set up a custom policy
ML engineer with over a decade of experience at Amazon, Meta, and startups. A MATS alum who co-authored RE-Bench, which METR now uses in its AI Agent time-horizon tracking (The Chart). Founded Luthien in January 2025.
"I saw the website and immediately thought, 'I want that.' I have to correct the pip install thing multiple times per day. You see pip and you're like, I already gave you permissions for uv. But it blocks on that permission and you lose hours of work. Awesome job, guys, it's like CLAUDE.md that actually works."
Try it + give feedback
Run Luthien on your codebase. Tell us what works, what breaks and what's missing.
Start a trial →Open source. Growing fast.
Apache 2.0 · contact@luthienresearch.org
