Permit0

Permit0 is an open source action authorization layer for AI agents. It sits between what an agent decides to do and what actually executes, applying deterministic policy to every tool call before it runs. A decorator wraps a function, such as a payment refund, and Permit0 scores the call, then allows it, blocks it or escalates it to a human reviewer. The product ships a canonical taxonomy of agent actions across 21 domains, a risk model built from nine risk flags and ten amplifiers, and compliance mappings for more than ten AI frameworks. The decision path is rules and cryptography rather than a model: every allow, block and escalation is signed and replayable. An optional language model reviewer can run as a second stage, but it may only deny or escalate and can never approve. It works with LangChain, Model Context Protocol, CrewAI, AutoGen, LangGraph, the OpenAI Agents SDK and Claude Code, and it offers three deployment modes. It is aimed at engineering teams shipping agents into production, with a design partner program aimed at fintech, insurtech and healthtech companies. The product is labeled beta, and the architecture specification on the site is still a draft in progress. The vendor does not publish pricing. The code is available on GitHub, and teams can apply to the design partner program or contact the founders.