comet-ml/opik-mcp

curl -LsSf https://astral.sh/uv/install.sh | sh  # install uv once
claude mcp add --transport stdio opik-mcp --env OPIK_API_KEY=<your-key> --env OPIK_WORKSPACE=<workspace> -- uvx opik-mcp
Set OPIK_API_KEY (required) and OPIK_WORKSPACE (optional, defaults to 'default')

Usage

• •Ask 'list my Opik projects' — Claude calls the list tool and shows your workspace's projects
• •Use read/list to browse traces, spans, prompts, test suites and experiments; write to log spans, score, comment and save prompt versions
• •Use ask_ollie for investigative questions and run_experiment to run evaluation experiments end-to-end

Key features

• ✓Small outcome-oriented surface: read, list, ask_ollie, write, schema and run_experiment
• ✓Ollie in-product AI assistant with direct workspace access and mid-stream writes (YOLO mode default, audit-logged)
• ✓Universal write dispatcher with schema introspection for valid payload construction
• ✓stdio or streamable-http transport, self-hosted support via COMET_URL_OVERRIDE

Best for

LLM engineers already running Opik who want to drive observability and evaluation from their coding assistant.

Caveats

• ⚠Requires Python 3.13+; currently run via uvx until the first PyPI release
• ⚠ask_ollie and run_experiment are Comet Cloud only — they fail on self-hosted
• ⚠Cursor enforces a hard 60s tool-call timeout that does not reset on progress; long Ollie turns fail there