The cMCP (Confidential MCP Runtime) is a new open-source gateway that secures AI agent tool calls by enforcing Cedar policies within a hardware-based Trusted Execution Environment.

Key Points

• cMCP intercepts agent tool calls and evaluates them against Cedar policy bundles inside a secure Trusted Execution Environment (TEE).
• The runtime supports hardware attestation via TPM 2.0, AMD SEV-SNP, and Intel TDX to ensure policy integrity and prevent unauthorized tampering.
• Each session generates a signed, tamper-evident TRACE Claim that provides verifiable proof of policy enforcement without requiring trust in the operator.
• The system is designed to meet regulatory standards, including the EU AI Act and NIST SP 800-207, by providing immutable audit logs.
• Developers can test the runtime using a software-only mode, which requires no specialized hardware, before deploying to production environments.

Why it Matters

This technology addresses critical security gaps in agentic AI by moving policy enforcement out of the reach of potentially compromised host operating systems. By providing hardware-attested proof of compliance, it enables organizations to safely integrate AI agents with sensitive APIs while meeting strict regulatory and audit requirements.