Configure the Limetry MCP server

@limetry/mcp exposes evaluate, record, audit, and policy tools to MCP hosts such as Claude Desktop, Cursor, and GitHub Copilot. Evaluate and audit sensitive agent actions from the editor without embedding credentials in prompts.

Register the server

Add this block to Claude Desktop or Cursor MCP config:

{
  "mcpServers": {
    "limetry": {
      "command": "npx",
      "args": ["-y", "@limetry/mcp"],
      "env": {
        "LIMETRY_API_KEY": "sk-your-bearer-token",
        "LIMETRY_BASE_URL": "http://localhost:3810"
      }
    }
  }
}

Point LIMETRY_BASE_URL at your self-run evaluation server (see [self-run setup](/docs/server/self-run setup)).

Call the tools

Once registered, the IDE agent can call limetry_evaluate, limetry_record_action, limetry_list_audit, and limetry_upsert_policy. Attempt a blocked action (for example http_post to a production URL) and confirm deny plus a privacy-safe audit record. Do not put secrets in metadata/details.

Until @limetry/mcp is on npm, run the built binary from a source checkout.