Opik MCP Server for Claude
Debug, evaluate, and monitor LLM applications from Claude — read traces and spans, score outputs, save prompts, run evaluation experiments, and query project metrics — with the official Opik MCP server by Comet.
Open the source and read safety notes before installing.
Safety notes
- The `write` tool can create traces, spans, scores, and prompts in your Opik workspace — review any write operations before confirming.
- The `run_experiment` tool executes evaluation experiments end-to-end, which may incur LLM API costs depending on your experiment configuration.
Privacy notes
- LLM traces (including prompts and completions), evaluation scores, experiment results, and prompt library contents from your Opik workspace are surfaced in Claude's context.
- Your `OPIK_API_KEY` is passed as an environment variable — store it in a secrets manager rather than plaintext shell config.
Prerequisites
- An Opik account (cloud at comet.com or self-hosted) and an API key from your workspace settings.
- Python with `uv` installed: `pip install uv` or `brew install uv`.
- An MCP client such as Claude Code or Claude Desktop.
Schema details
- Install type
- cli
- Troubleshooting
- No
- Scope
- Source repo
- Estimated setup
- 5 minutes
- Difficulty
- beginner
Full copyable content
{
"mcpServers": {
"opik": {
"command": "uvx",
"args": ["opik-mcp"],
"env": {
"OPIK_API_KEY": "your-api-key",
"OPIK_WORKSPACE": "default"
}
}
}
}About this resource
Overview
The Opik MCP Server is the official Model Context Protocol server for Opik, Comet's open-source LLM observability platform. It connects Claude directly to your Opik workspace — traces, spans, evaluation scores, prompts, datasets, and experiments — through six outcome-oriented tools.
Install via uvx (the Python-based runner). The earlier TypeScript npx opik-mcp version is
deprecated and sunsets November 15, 2026 per the Opik changelog.
Key capabilities
- Traces & spans — retrieve and filter LLM execution traces and individual span details.
- Scoring — log evaluation scores and qualitative feedback against trace outputs.
- Prompt management — read and write prompts in the Opik prompt library.
- Experiments — run evaluation experiments end-to-end via
run_experiment. - Projects — browse projects, filter by metrics, and query aggregate stats.
- Ollie (AI assistant) —
ask_olliefor investigative queries and synthesis across your data.
Tools
| Tool | Purpose |
|---|---|
read |
Retrieve entities by ID, name, or URI |
list |
Browse collections with filtering and pagination |
write |
Create traces, spans, scores, comments, prompts, experiments |
schema |
Introspect write-operation shapes before calling them |
run_experiment |
Execute evaluation experiments end-to-end |
ask_ollie |
Investigative synthesis via the in-product AI assistant |
How it compares
| Server | Traces & spans | Evals & experiments | Prompt library | Datasets | Auth |
|---|---|---|---|---|---|
| Opik MCP | Yes | Yes | Yes | Yes | API key |
| LangSmith MCP | Yes | Yes | Yes | Yes | API key |
| Arize Phoenix MCP | Yes | Yes | No | Yes | API key |
| Weights & Biases MCP | Yes | Yes | Limited | Yes | API key |
Opik is fully open-source (Apache-2.0) and supports both cloud (comet.com) and self-hosted deployments, giving teams full control over where trace data is stored.
Installation
Claude Code
claude mcp add opik -e OPIK_API_KEY=your-api-key -- uvx opik-mcp
Get your API key from Opik workspace settings → API Keys.
Claude Desktop
{
"mcpServers": {
"opik": {
"command": "uvx",
"args": ["opik-mcp"],
"env": {
"OPIK_API_KEY": "your-api-key",
"OPIK_WORKSPACE": "default"
}
}
}
}
Self-hosted Opik
Set OPIK_URL_OVERRIDE=http://localhost:5173/api to point to your local instance.
Requirements
- An Opik account (cloud or self-hosted) and an API key.
- Python with
uvinstalled. - An MCP client (Claude Code or Claude Desktop).
Security
- API key authentication — generate a dedicated key from workspace settings.
- Self-hosted deployment keeps all trace data within your infrastructure.
Source Verification Notes
Verified on 2026-06-18:
- GitHub repo
comet-ml/opik-mcp(Apache-2.0) is the official source; the README documents the six tools (read,list,write,schema,run_experiment,ask_ollie), theuvx opik-mcpinstall path, and deprecation of the TypeScript package. - GitHub repo
comet-ml/opik(Apache-2.0, 10k+ stars) is the main Opik platform repo documenting the full LLM observability, evaluation, and prompt engineering workflow. - Opik changelog entry for March 3, 2026 at
comet.com/docs/opik/changelog/2026/3/3documents the MCP server revamp with remote MCP support, improved auth, and expanded native features including prompt and dataset workflows.
Source citations
Add this badge to your README
How it compares
Opik MCP Server for Claude side by side with 3 alternatives on trust, install, platform support, and disclosed safety notes — all from reviewed registry metadata.
| Field | Opik MCP Server for Claude Debug, evaluate, and monitor LLM applications from Claude — read traces and spans, score outputs, save prompts, run evaluation experiments, and query project metrics — with the official Opik MCP server by Comet. Open dossier | Arize Phoenix MCP Server for Claude Inspect LLM traces and spans, manage prompts, explore datasets, and review evaluation experiments from Claude — with the official Arize Phoenix MCP server, built into the open-source Phoenix AI observability platform. Open dossier | LangSmith MCP Server for Claude Connect Claude to LangSmith — retrieve conversation threads and traces, fetch and push prompts, browse evaluation datasets and experiments, and access billing usage — with the official LangSmith Model Context Protocol server from LangChain. Open dossier | OpenTelemetry MCP Server for Claude Analyze distributed traces and LLM observability data from Claude — search traces and spans, find errors, list services, analyze LLM token usage, identify slow LLM operations, and discover AI model usage patterns — with the OpenTelemetry MCP server supporting Jaeger, Grafana Tempo, and Traceloop backends. Open dossier |
|---|---|---|---|---|
| Trust | ||||
| Install risk | Review first | Review first | Review first | Review first |
| Notes | Safety ✓ Privacy ✓ | Safety ✓ Privacy ✓ | Safety ✓ Privacy ✓ | Safety ✓ Privacy ✓ |
| Category | mcp | mcp | mcp | mcp |
| Source | source-backed | source-backed | source-backed | source-backed |
| Author | Comet | Arize AI | LangChain | Traceloop |
| Added | 2026-06-18 | 2026-06-18 | 2026-06-18 | 2026-06-18 |
| Platforms | Claude CodeCodexCursorClaude Desktop | Claude CodeClaude Desktop | Claude CodeClaude Desktop | Claude CodeClaude Desktop |
| Source repo | — | — | — | — |
| Safety notes | ✓The `write` tool can create traces, spans, scores, and prompts in your Opik workspace — review any write operations before confirming. The `run_experiment` tool executes evaluation experiments end-to-end, which may incur LLM API costs depending on your experiment configuration. | ✓The `write` and upsert tools can create or modify prompts and add dataset examples — review Claude's proposed changes before confirming. For self-hosted deployments, ensure your Phoenix instance is accessible from the machine running the MCP server. | ✓The server runs locally via `uvx` and connects to the LangSmith cloud API using your API key. The server provides read access to your traces, prompts, datasets, and experiments; it can also push prompts (write). Trace data may contain sensitive LLM inputs and outputs from your production systems — treat retrieved traces as confidential. | ✓All tools are read-only — the server queries trace data but does not modify your application or tracing backend. Trace data may contain sensitive information (request parameters, user IDs, SQL queries) — ensure Claude has appropriate access to this data. |
| Privacy notes | ✓LLM traces (including prompts and completions), evaluation scores, experiment results, and prompt library contents from your Opik workspace are surfaced in Claude's context. Your `OPIK_API_KEY` is passed as an environment variable — store it in a secrets manager rather than plaintext shell config. | ✓LLM traces (including prompts and completions), evaluation annotations, session data, datasets, and experiment results from your Phoenix workspace are surfaced in Claude's context. Your `PHOENIX_API_KEY` is passed as an environment variable — store it in a secrets manager rather than plaintext shell config. | ✓Conversation threads, prompt content, run inputs/outputs, and dataset examples are surfaced in Claude's context. LANGSMITH_API_KEY is a secret — store it in your environment or MCP client config, never in repositories. | ✓Distributed trace content including service names, operation names, error messages, HTTP parameters, and LLM prompt/response metadata may be surfaced in Claude's context. For Traceloop backend, `BACKEND_API_KEY` is required and grants access to your Traceloop organization's trace data. |
| Prerequisites |
|
|
|
|
| Install | | | | |
| Config | | | | |
| Citations | ||||
| Claim | Unclaimed | Unclaimed | Unclaimed | Unclaimed |
Signals
Loading live community signals…
A short, calm digest of reviewed Claude resources. Unsubscribe any time.