Deterministic governance for AI agent tool calls in Claude Code, MCP, LangGraph, or any agent loop. Pin approved capabilities, detect drift, and refuse unsafe or unapproved actions before execution. No model in the decision path.
https://recusal.io ↗// momentum
// why it's here
- Active: last commit today
- Built almost entirely by one person
- Has issues labelled for new contributors
- Health 93/100 with only 7 stars
- 14 forks for 7 stars: people build on it
// how to help
// readme
Recusal
Recusal checks each tool call an AI agent proposes against your policy, and refuses the unsafe or unapproved ones before the tool runs. No model makes that decision; your policy functions do.
It drops into Claude Code as a hook, MCP tool calls included, and gates a Claude Messages API loop or any Python agent loop. For agent frameworks there are eight adapter packages, each tested against its framework: LangGraph, LangChain, the OpenAI API and Agents SDK, Google ADK, CrewAI and Pydantic AI. The kernel has zero runtime dependencies and runs on Python 3.9+.
Website: recusal.io, with an interactive demo of the gate.
Contents: See it refuse · Works with · Adapters · Install · Policies · MCP · Audit · How it works · CI · Built to be checked · Documentation