Back to directory

cMCP: Confidential MCP Runtime

An open-source gateway that enforces Model Context Protocol tool-call policy inside a hardware trusted execution environment, so a governed agent process cannot alter or bypass its own policy engine.

cMCP: Confidential MCP Runtime

cMCP intercepts every MCP tool call an agent makes, evaluates it against a Cedar policy bundle, and enforces the decision from a control plane the governed process cannot reach, running inside a Trusted Execution Environment when available or in a signed-only software mode otherwise. Each session produces a signed TRACE Claim, a tamper-evident record that a separate verifier can check without having to trust the operator running the gateway.

The project targets a specific weakness in software-only policy enforcement: an administrator or a compromised evaluator process can swap a policy bundle or flip an allow or deny decision in memory without leaving reliable evidence. cMCP is described as a developer preview launched at the Confidential Computing Summit, installable as a Python package and runnable in software mode without special hardware.

More in MCP servers

See all 734
  • Light Cloud

    light-cloud.com

    A hosting platform that deploys a static site, a container, or a managed Postgres database straight from a GitHub, GitLab, or Bitbucket push, with a preview URL for every branch and an MCP server for deploying from Claude Code.

  • ForensicDbg

    forensicdbg.com

    A post-mortem debugger for native Windows x86 and x64 crashes, with an MCP interface so AI tools can consume crash-dump analysis directly.

  • Karada.ai

    karada.ai

    A CI/CD platform that auto-compiles existing APIs into Go MCP servers, attaches composable middleware, and exposes them through one gateway usable from Cursor, Claude Code, and Windsurf.