Brex open-sourced CrabTrap, an HTTP proxy that uses an LLM-as-a-judge to inspect and block unsafe agent actions in production.
Brex engineering released CrabTrap, an HTTP proxy layer designed to sit between AI agents and external APIs in production environments. It uses an LLM-as-a-judge pattern to evaluate agent HTTP requests in real time, blocking or flagging actions that violate defined policies. The tool targets teams running autonomous agents where unchecked tool calls represent a live security and compliance risk. No pricing was mentioned — it appears to be an open-source release from Brex's engineering team.
CrabTrap inserts an LLM-as-a-judge at the HTTP proxy layer — every outbound agent request is evaluated against a policy before it hits an external API. This means you can enforce rules like 'never DELETE records' or 'only call payment endpoints with user-approved amounts' without baking logic into the agent itself. It's a clean separation of concerns: your agent stays dumb, your proxy stays paranoid.
Deploy CrabTrap in front of your agent's tool-call layer this week and write a policy that blocks any HTTP DELETE or POST to /payments without an explicit user-confirmed flag — measure how many calls it intercepts in your staging environment.
Clone the CrabTrap repo: git clone https://github.com/brexhq/crabtrap && cd crabtrap
Tags
Sources