Opik MCP Server

uvx opik-mcp@latest

{
  "mcpServers": {
    "opik-mcp": {
      "command": "uvx",
      "args": [
        "opik-mcp@latest"
      ]
    }
  }
}

Comet's LLM observability platform stores traces, spans, prompts, datasets and experiment results, and the obvious way to expose that is one tool per entity type. This server does the opposite: read, list, write and schema each take an entity name — project, trace, span, test_suite, experiment and so on — so the tool list stays at six while the entity catalogue can grow. schema is the piece that makes it work, because the model asks what fields an entity has before it writes a filter, which is the same discipline that keeps hand-written queries valid. Two tools break the pattern on purpose: ask_ollie hands an investigation to Opik's in-product assistant, and run_experiment runs an evaluation end to end. The older TypeScript build published as npx opik-mcp is deprecated in favour of the Python one.