Opik MCP Server

Official

Connect your AI host directly to your Opik workspace: read traces, log scores, and manage prompts.

How to install ↓ Compare GitHub source repository ↗

Category: Dev Tools #26 of 438
Stars: ★ 220 Very popular
Transport: stdio (local process) · Streamable HTTP
Runtime: Python 3.13+
Credentials: API key / credential required
License: Apache-2.0
Last commit: 3 days ago
Tools: 6

67FMRS · C

Opik MCP Server is the official MCP server for Opik, designed for LLM observability and evaluation. It offers a comprehensive tool set and good client support, but some advanced features are limited, and security configuration requires attention.

Strongest · Maintenance 18/20 Weakest · Reliability 8/20

Reliability

8/20

Security and permissions

14/20

Maintenance

18/20

Documentation

15/20

Setup experience

12/20

Why each score

Reliability 8/20: Source contains a Python test suite (tests/ directory mentioned in README, but no concrete content available). CI workflows are not directly visible. The MANIFEST claims an npm package [email protected] offering stdio transport, but the README indicates the TypeScript server is deprecated and the primary server is a Python package that is not yet published to PyPI (requires git install). This inconsistency makes the startup path unclear. Server tools are declared as read, list, ask_ollie, write, schema, and run_experiment, matching the README; however, the actual tool implementations for the Python package are not verified in the source files. No execution tests available, so initialization flow could not be confirmed. Lack of verifiable CI and test artifacts caps reliability at 12; here scored 8.
Security and permissions 14/20: Required env var OPIK_API_KEY is marked as secret, documented to be passed out-of-process, and query content is not logged. Telemetry section discloses use of SHA-256 digest of API key for analytics, and states raw key never leaves process. Self-hosted deployments support disabling telemetry (OPIK_MCP_ANALYTICS_ENABLED=false) and overriding telemetry URL. Documentation recommends using stdio over HTTP for safer local deployments, and notes HTTP transport does no local credential validation, but authorization is enforced by backend. No obvious malicious patterns or default irreversible operations found. No default enablement of dangerous features (e.g., auto-deploy) observed, but network calls and telemetry require explicit user consent. Documented confirmation mechanism (OPIK_MCP_AUTO_APPROVE) exists, but default is auto-approve, which could lead to overprivileged operations. Security score 14 due to auto-approve default and potential data leakage of telemetry external network traffic.
Maintenance 18/20: Repository is active, with 216 stars, 32 open issues, not archived. License Apache-2.0. Clear migration plan and deprecation timeline (TypeScript server sunset 2026-11-15; Python version under development). Evidence of regular commits and releases exists but not directly visible. Multiple supported clients (Claude, Cursor, VS Code) and MCP Inspector. Issues are public, with support and community channels (Slack, GitHub Issues). Version 2.0.1 is released with MANIFEST. Strong maintenance; dependency updates and release cadence cannot be directly confirmed, but activity is evident. Score 18 due to limited verifiable evidence on release cadence and dependency updates.
Documentation 15/20: README provides thorough docs: installation (uvx, Claude Code, Cursor, VS Code, self-hosted), tool descriptions (6 tools) overview, environment variable configuration (including self-hosted, telemetry), troubleshooting (common errors like OPIK_API_KEY not set, Cursor timeout), and migration guide from v2. Example CLI commands and JSON configs. Known limitations (Cursor 60s timeout) are documented. Gaps: no full reference for tool parameters (relies on schema tool and Ollie knowledge), permissions and telemetry details are not exhaustive, and current status of Python package is pre-release (noted in README, but not a documentation section). Documentation overall strong, score 15.
Setup experience 12/20: Installation instructions are clear, using uvx or git. Step-by-step configuration examples for Claude Code, Cursor, and VS Code. Python 3.13+ requirement is noted. MCP Inspector setup provided. However, server is not yet published as a Python package (needs git install), so installation is limited. Additionally, standard flow requires an existing Opik workspace, API key, and possibly endpoint configuration. The server includes many environment variable settings, which add configuration complexity. No verifiable CI pipeline or committed end-to-end tests proving installation works out-of-the-box, so setup score capped at 15, here scored 12.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessUses the networkChanges third-party account data

Best for

LLM engineers already using Opik
Developers who want to manage LLM observability directly from their coding environment
Workflows needing deep integration with the Opik platform

Not for

Users not using Opik
Those not interested in LLM observability
Users needing offline or non-cloud features (ask_ollie and run_experiment are Comet Cloud only)

Required permissions

Read access to Opik projects, traces, spans, test suites, experiments, and prompts
Write access to traces, spans, scores, comments, prompt versions, and experiments
Access to Opik API key (required)

Risks and side effects

Requires API key, which poses security risks if exposed
ask_ollie and run_experiment are limited to Comet Cloud; not available for self-hosted
Cursor enforces a 60-second tool-call timeout; long operations may fail
Default YOLO mode auto-approves Ollie's writes, potentially leading to unintended modifications

Setup

Before you start

Runtime：Python 3.13+

OPIK_API_KEY requiredsecret Required Opik API key for authenticating all reads/writes; get it from comet.com/api/my/settings/.

Other optional settings (25)

OPIK_WORKSPACE optional Opik workspace name (optional, defaults to 'default'); cloud users with a named workspace should set it to the lowercase name from the URL.

COMET_WORKSPACE optional Deprecated alias for OPIK_WORKSPACE; OPIK_WORKSPACE wins if both are set.

COMET_WORKSPACE_ID optional Optional workspace UUID stamped into analytics events for stable BI joins.

COMET_URL_OVERRIDE optional Self-hosted Comet host, defaults to https://www.comet.com.

OPIK_URL optional Opik API base URL (derived from COMET_URL_OVERRIDE by default); set only if Opik lives elsewhere.

OPIK_DEFAULT_PROJECT_NAME optional When set, instructs the LLM to pass this project_name on every tool call unless the user names another.

OPIK_MCP_TRANSPORT optional Transport mode: stdio (default) or streamable-http.

OPIK_MCP_HOST optional uvicorn bind host for HTTP mode, default 127.0.0.1.

OPIK_MCP_PORT optional uvicorn bind port for HTTP mode, default 8080.

OPIK_MCP_RELOAD optional Enables uvicorn --reload in dev, default false.

OPIK_MCP_AS_URL optional OAuth Authorization Server URL, required for MCP hosts to bootstrap OAuth over HTTP transport.

OPIK_MCP_RESOURCE_URI optional Canonical public URI of this server, advertised in protected-resource metadata.

OPIK_MCP_LOG_LEVEL optional stderr log level, default INFO; set DEBUG to troubleshoot.

OPIK_MCP_AUTO_APPROVE optional Whether Ollie's mid-stream writes auto-approve (default enabled); set disabled to require per-action confirmation.

OPIK_MCP_ELICIT_TIMEOUT_SECONDS optional How long Ollie's confirmation prompt may wait, default 60 seconds.

OPIK_MCP_POD_READY_TIMEOUT_S optional Ollie pod cold-start poll cap, default 120 seconds.

OPIK_MCP_POD_READY_INTERVAL_S optional Cold-start poll interval, default 2 seconds.

OPIK_MCP_HEARTBEAT_INTERVAL_S optional Heartbeat cadence, default 15 seconds, emitting progress notifications to keep host timeouts at bay.

OPIK_MCP_STREAM_IDLE_TIMEOUT_S optional How long ask_ollie waits before aborting on pod silence, default 300 seconds.

OPIK_MCP_ANALYTICS_ENABLED optional Anonymous telemetry switch, default true; set false to opt out.

OPIK_MCP_ANALYTICS_URL optional Telemetry endpoint, default https://stats.comet.com/notify/event/.

OPIK_MCP_ANALYTICS_ENVIRONMENT optional Environment tag (prod/staging/dev), default prod.

OPIK_MCP_ANALYTICS_SOURCE optional Event source label, default comet.com; on-prem installs should override to empty string.

OPIK_MCP_ANALYTICS_CONNECT_TIMEOUT_S optional Telemetry HTTP connect timeout, default 5 seconds.

OPIK_MCP_ANALYTICS_TOTAL_TIMEOUT_S optional Telemetry HTTP total timeout, default 10 seconds.

Install uv (curl -LsSf https://astral.sh/uv/install.sh | sh)
Obtain OPIK_API_KEY from your Opik workspace; optionally set OPIK_WORKSPACE (defaults to default)
Add the server to your MCP client config using uvx opik-mcp and set environment variables as shown
Restart the client, verify connection, and start using tools

.mcp.json

{
  "mcpServers": {
    "opik-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["opik-mcp"],
      "env": {
        "OPIK_API_KEY": "<your-key>",
        "OPIK_WORKSPACE": "<your-workspace>"
      }
    }
  }
}

Shown for Claude Code. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json

{
  "servers": {
    "opik-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "opik-mcp"
      ],
      "env": {
        "OPIK_API_KEY": "<your-key>",
        "OPIK_WORKSPACE": "<your-workspace>"
      }
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal

claude mcp add opik-mcp -e 'OPIK_API_KEY=<your-key>' -e 'OPIK_WORKSPACE=<your-workspace>' -- uvx opik-mcp

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After restarting the host, run /mcp in Claude Code — opik-mcp should appear as connected. Then ask "list my Opik projects" in chat; Claude should call the list tool and return your workspace's projects.

Troubleshooting

If ask_ollie complains about missing API key, verify the environment variables are in the MCP client's env block
If ask_ollie hangs, it may be a cold-start issue; retry
Self-hosted deployments do not support ask_ollie or run_experiment; use read/list/write instead
Cursor timeouts can be mitigated by shortening queries or using Claude Code

Things to try

Once connected, you can ask your AI assistant things like:

List my Opik projects.
Why did the experiment 'gpt-4o-rerank-v3' regress on factuality?
Score trace 7f2e… 0.9 on helpfulness with reason 'great recovery'.
Show me the traces in project 'demo' from this week.

Tools 6

read read-only

Read Opik entities (project, trace, span, test_suite, experiment, prompt) by id, name, or opik:// URI.

list read-only

List Opik entities with optional name filter and pagination.

ask_ollie writes

Investigate and synthesize via Opik's in-product AI assistant, able to answer complex queries and execute writes.

write writes

Universal write operation for logging traces/spans, scoring, commenting, saving prompts, and managing test suites and experiments.

schema read-only

Introspect write-operation schemas, required fields, and examples to construct valid payloads.

run_experiment writes

Run an evaluation experiment end-to-end via Ollie and write results back to Opik.

Use cases

Quickly fetch details about projects, traces, and spans

Add scores and comments to traces or spans for evaluation

Save and version prompts

Analyze experiment regressions with Ollie

Run evaluation experiments and inspect results

Supported clients

Claude Code

CursorPartial support

VS Code

MCP InspectorPartial support

Listed from the project's documentation, not tested by this site.

Overview

Opik MCP Server is the official Model Context Protocol (MCP) server for Opik, the open-source LLM observability and evaluation platform built by Comet. It enables AI hosts like Claude Code, Cursor, and VS Code Copilot to interact with your Opik workspace, allowing you to read traces, log scores, save prompt versions, and ask Ollie, Opik's in-product AI assistant, investigative questions. The server provides six tools—read, list, write, ask_ollie, schema, and run_experiment—covering the full lifecycle from reading and annotating to curating, authoring, and iterating.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 62.9k · Tools 2 Compare with this →

NPM Sentinel MCP Server 79 · B

AI-powered NPM package analysis MCP server

★ 18 · Tools 19 Compare with this →

MartinLoop MCP Server 79 · B

An independent control layer that gives AI coding agents budgets, brakes, and receipts

★ 410 Compare with this →

GitHub MCP Server 76 · B

GitHub's official MCP server for managing repos, issues, PRs, and workflows via natural language

★ 33.5k · Tools 15 Compare with this →

Compare with other MCP servers

Source revision e73758164ca2 Data synced 2026-10-12 Read the FMRS scoring method