opik-mcp: Connect AI Assistants to Opik

Overview

opik-mcp connects MCP-compatible assistants to Opik, an LLM observability and evaluation platform. It gives an assistant a way to inspect workspace data and perform selected Opik actions without switching between chat and the Opik interface.

It is most useful for teams already using Opik who want to investigate traces, annotate results, or manage prompts from their coding assistant. The server provides a local option for self-hosted and open-source deployments; Opik Cloud also offers a hosted MCP endpoint.

Key Features

• Read projects, traces, spans, datasets, experiments, prompts, threads, and Diagnostics issues.
• Filter, search, sort, and paginate supported project data, including time-window queries.
• Write trace and span data, scores, and comments through a validated operation dispatcher.
• Save prompt versions and manage datasets and experiments.
• Inspect schemas for write operations and supported list filters before making calls.
• Connect through stdio or configure streamable HTTP for suitable deployments.
• Configure supported MCP clients, including Claude Code, Cursor, VS Code Copilot, Codex, and opencode.

Use Cases

• LLM engineers can ask an assistant to find failed or unusually slow traces while debugging an application.
• Evaluation teams can record feedback scores and comments on traces as they review model outputs.
• Prompt developers can save prompt versions and inspect existing prompts without leaving their coding workflow.
• Opik administrators can review Diagnostics issues and manage their status from an MCP client.

Getting Started

The README recommends configuring detected AI clients with the Opik CLI:

uvx opik mcp configure

This requires uv. For manual setup, self-hosted deployments, credentials, and troubleshooting, see the repository README and Opik MCP documentation.

Alternatives

• phoenix: Phoenix is a self-hosted tracing and evaluation platform, while opik-mcp exposes Opik workspace workflows to MCP clients.
• LangWatch: LangWatch provides a platform for LLM testing and production monitoring, while opik-mcp connects assistants to an Opik workspace.
• Langsmith-sdk: LangSmith SDKs connect applications to LangSmith for debugging and evaluation, rather than providing an MCP server for Opik.

Considerations

• The local Python server requires Python 3.13 or newer and uses uvx in the recommended setup.
• Local or self-hosted use requires configuring the appropriate Opik endpoint and authentication. Workspace settings differ for API-key, OAuth, and open-source deployments.
• The README documents a 60-second hard tool-call timeout in Cursor, which can affect long-running searches.
• Anonymous usage telemetry is enabled by default and can be disabled with OPIK_MCP_ANALYTICS_ENABLED=false.
• The README says the older TypeScript server is deprecated and sunsets on 2026-11-15. New setups should use the Python package.