What problem does it solve? Standard logging cannot prove what an AI agent actually did, leaving regulated teams without tamper-evident evidence of agent behavior. This cookbook walks through setting up cryptographically signed, hash-chained audit receipts on every Claude Code tool call so auditors can verify the full chain offline. ## Core Features & Use Cases - Policy-gated tool execution: A Cedar policy evaluated in a PreToolUse hook allows or denies each tool call before it runs, with explicit forbid rules for destructive commands. - Signed Ed25519 receipts: A PostToolUse hook signs JCS-canonical, hash-chained receipts for every tool call, verifiable offline with a single CLI command. - Tamper detection and CI/CD gating: Demonstrates how modifying any receipt field breaks verification, and how to gate merges on receipt chain integrity in GitHub Actions. - Use Case: A fintech team running Claude Code in CI needs to prove to auditors that a policy gate held for every automated build step; this recipe produces a verifiable receipt chain archived as a build artifact and composable with SLSA provenance. ## Quick Start Ask the assistant to walk you through setting up signed audit receipts for Claude Code tool calls using the protect-mcp hooks and a Cedar policy.
