MCP Server

Start Runs and read verdicts from Claude Desktop, Cursor, or Claude Code.

The @peerlm/mcp server exposes the same API to Claude Desktop, Cursor, Claude Code, and other MCP-compatible clients. It is available on Pro, Team, and Enterprise.

Setup

{
  "mcpServers": {
    "peerlm": {
      "command": "npx",
      "args": ["-y", "@peerlm/mcp"],
      "env": { "PEERLM_API_KEY": "<YOUR_PEERLM_API_KEY>" }
    }
  }
}

Use a read-write key when the client must create a Monitor, ingest traffic, or start a Run. Keep the key in the client's environment, not in a repository.

Monitor tools

• list_workloads — list Monitors
• create_workload — create a Monitor under the organization capacity lock
• ingest_logs — ingest normalized records by workload name and auto-create a Monitor
• list_audits — list comparison Runs
• get_audit — retrieve a Run receipt and recommendation
• start_audit — start a comparison Run

The MCP names retain the API's compatibility vocabulary: workloads are Monitors and audits are Runs.

Catalog and capacity tools

• list_models — find exact catalog model IDs
• get_usage — read plan, Monitor capacity, and pooled Runs

Compatibility tools

The package still exposes list_suites, get_suite, create_suite, run_eval, get_results, create_system_prompt, and create_test_prompt for existing authored-evaluation integrations. They operate the internal Suite engine and have no customer dashboard counterpart.

Recommended workflow

1. Use list_models to resolve the exact production and candidate IDs.
2. Use create_workload with the exact production model.
3. Connect OpenTelemetry in the dashboard, or call the REST ingest endpoint with the returned Monitor ID.
4. Confirm enough replayable records exist, then use start_audit.
5. Use get_audit to read the switch, route, hold, or inconclusive result.