@pooml/mcp

MCP server for pooml: query your logs and metrics with plain SQL from Claude, or any other MCP client.

Runs on your machine (stdio), talks to your pooml instance over its read-only query API. Your credentials stay local; the pooml server needs nothing MCP-specific.

Tools

• query_logs - SQL over the logs (levels, services, full-text search via FTS5)
• query_metrics - SQL over the metrics (counters and gauges)
• list_metrics - what metrics exist, so the model doesn't guess names

Strictly read-only: every query goes through pooml's layered SQL validation (SELECT-only, allow-listed tables, read-only connection, timeouts, row caps).

Setup

1. On the pooml server, enable the query API: POOML_QUERY_API_ENABLED=true and POOML_QUERY_API_AUTH_SECRET=<min 32 chars>.
2. Add to your MCP client. For Claude Code:

claude mcp add pooml \
  -e POOML_URL=https://your-pooml-host:8080 \
  -e POOML_QUERY_API_AUTH_SECRET=your-query-secret \
  -- npx -y @pooml/mcp

Or in .mcp.json:

{
  "mcpServers": {
    "pooml": {
      "command": "npx",
      "args": ["-y", "@pooml/mcp"],
      "env": {
        "POOML_URL": "https://your-pooml-host:8080",
        "POOML_QUERY_API_AUTH_SECRET": "your-query-secret"
      }
    }
  }
}

If pooml sits behind a Cloudflare Tunnel with Access, add the service token: POOML_CF_ACCESS_CLIENT_ID and POOML_CF_ACCESS_CLIENT_SECRET.

Why SQL

pooml's whole pitch is that your observability data is SQLite queried with SQL - and LLMs already speak SQL fluently. No custom query language for the model to hallucinate around: it writes SELECT service, COUNT(*) FROM logs WHERE level >= 4 ... and gets answers.

License

Apache-2.0