@limetry/mcp

MCP server for Limetry agent action governance: evaluate ActionIntents, record outcomes, list privacy-safe audit events, and upsert action policies.

Quickstart

1. Run npx -y @limetry/mcp from your MCP host configuration.
2. Point the MCP host at the published package with LIMETRY_API_KEY and LIMETRY_BASE_URL.
3. Ask the agent to call limetry_evaluate before irreversible tool calls; enforce deny in the loop.
4. Call limetry_record_action after execute/skip/block.

The MCP package remains a local stdio process. It makes short authenticated HTTP requests to LIMETRY_BASE_URL; hosted users should set that value to https://api.limetry.com (development: https://api.dev.limetry.com). There is no hosted SSE endpoint.

[!NOTE] Evaluate uses the full intent for policy matching. Audit stores a privacy-safe record (audit_mode=minimal by default). Do not put API keys, passwords, or chat transcripts in metadata / details. limetry_record_action runs client-side redactDetails before POST; the server also minimizes. See Data minimization.

Tools

| Tool | Purpose | | --- | --- | | limetry_evaluate | Allow / deny / wait + reasons + signed receipt | | limetry_record_action | Scrubbed action.recorded event | | limetry_list_audit | List recent privacy-safe audit rows | | limetry_upsert_policy | Create/update ActionPolicy (auditMode supported) |

Run from source

For local development or unreleased changes, build from a checkout:

yarn build:sdk && yarn workspace @limetry/mcp build
node packages/mcp/dist/index.js

Cursor / Claude config example:

{
  "mcpServers": {
    "limetry": {
      "command": "node",
      "args": ["/absolute/path/to/limetry/packages/mcp/dist/index.js"],
      "env": {
        "LIMETRY_BEARER_TOKEN": "your-bearer-token",
        "LIMETRY_BASE_URL": "http://localhost:3810"
      }
    }
  }
}

For normal use, prefer the published npx -y @limetry/mcp command above.