PermitRail

An authorization, proof, and audit layer for what AI agents actually do. A policy decides which tool calls run freely, which are blocked, and which need a person to sign off. On approval, PermitRail issues a short-lived Ed25519-signed proof bound to the exact action, the tool runs once, and a signed receipt lands in the audit log. Zero runtime dependencies, runs on Node, browsers, and edge runtimes, and ships a runnable MCP server.