Skip to content

What Model MCP deliberately does not do

A short list of limits, so that you know what to expect before you install.

  • It does not run an AI model. There is no model, no API key and no chat window. You bring an MCP client and its model provider.
  • It does not send data to the maintainer. No telemetry, analytics or crash reports. The only automatic requests are update and package checks that you can turn off.
  • It does not host anything. There is no cloud service, account or license server.
  • It does not sandbox C# execution. revit_execute_code runs with your Windows rights. A confirmation token makes you look first. If that is too much risk, turn the tool off by policy.
  • It does not stop an MCP client that confirms for itself. The client receives each confirmation token. Use read-only mode or policy for unattended runs.
  • It does not provide TLS or single sign-on. The HTTP listener is optional, loopback by default and uses one bearer token. Use a tunnel or TLS proxy for remote access.
  • It does not authenticate individual callers. All HTTP callers share the token.
  • It does not encrypt its files. Use disk encryption.
  • It does not write to the Windows Event Log or forward logs. Logs are local files.
  • It does not offer full family authoring or manage ACC or Forma models. Family work is limited to the tools in the tool reference. Revit Server paths are for multi-model runs.
  • It does not run in Revit 2021 or older. Supported Revit versions are 2022 to 2027.
  • It does not change tools or descriptions after install. The tool list is fixed per release.
  • It does not install silently over a signed channel yet. Release binaries are unsigned; see code signing.
  • It does not replace Revit's own checks. Results report what the Revit API returns. A non-empty skipped list means a result is incomplete.

See the security whitepaper for the full list of controls and residual risks.