I built this after getting paged one time too many for an agent that leaked a Slack token to Claude. It's an OpenAI-compatible proxy plus MCP-tool hooks that bl

I built this after getting paged one time too many for an agent that leaked a Slack token to Claude. It's an OpenAI-compatible proxy plus MCP-tool hooks that block credentials, prompt injection, and out-of-policy tool calls before they hit the provider. Every call is hash-chained so you can prove what the agent actually did. Curious what breaks first for people — happy to debug live.