Protect Mcp

功能

• Cedar policy enforcement before tool execution
• Ed25519 signed receipts for allow/deny decisions
• Hash-chained audit trail for integrity
• Offline verification of receipts
• Command-line tools for verification and auditing

使用场景

• Securing AI agent actions in production environments
• Providing auditable evidence for compliance requirements
• Detecting and preventing unauthorized tool usage
• Establishing tamper-evident logs of AI agent activity

非目标

• Acting as a general-purpose logging system
• Storing sensitive data beyond cryptographic receipts
• Replacing the Claude Code session log entirely

工作流

1. User installs plugin and configures hooks.
2. User creates a Cedar policy file.
3. User starts the signing server or relies on environment variables.
4. Claude Code agent makes a tool call.
5. PreToolUse hook evaluates call against Cedar policy.
6. If denied, tool call is blocked; if allowed, tool executes.
7. PostToolUse hook signs a receipt with decision, input, and output.
8. Receipt is saved locally and linked to the previous receipt.
9. User can verify individual receipts or audit the chain offline.

实践

• Policy Enforcement
• Cryptographic Auditing
• Access Control

先决条件

• Node.js installed
• Claude Code environment
• Policy file (e.g., ./protect.cedar)
• Signing key file (e.g., ./protect-mcp.key)