opik-mcp

June 4, 2026 · View on GitHub

Migrating from the old npx opik-mcp? The TypeScript server is deprecated and sunsets on 2026-11-15. Swap npx -y opik-mcp for uvx opik-mcp@latest in your MCP client config. Full guide: legacy/typescript/MIGRATION.md.

Model Context Protocol server for Opik + Ollie. Plug your AI host (Claude Code, Cursor, VS Code Copilot, MCP Inspector) directly into your Opik workspace — read traces, log scores, save prompt versions, and ask Ollie investigative questions, all from the chat.

Built for LLM engineers who already run Opik and want to drive it from the same AI assistant they code with.

You:    "Why did the experiment 'gpt-4o-rerank-v3' regress on factuality?"
Claude: → ask_ollie → reads experiment + traces → "Three traces failed because…"

You:    "Score trace 7f2e… 0.9 on helpfulness with reason 'great recovery'."
Claude: → write(score.create) → done

Install

opik-mcp is a Python package (requires Python 3.13+). The recommended way to run it is uvx, which fetches and runs the latest published version on demand — no global install, no virtualenv juggling.

Install uv once:

curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS / Linux
# or: brew install uv

You'll need two things from your Opik workspace:

  • OPIK_API_KEY — get it from comet.com/api/my/settings/.
  • OPIK_WORKSPACE — your workspace name (lowercase, as it appears in the URL). E.g. https://www.comet.com/acme-ai/...OPIK_WORKSPACE=acme-ai. Optional — defaults to default (the Opik SDK convention), which is correct for local/OSS installs; cloud users with a named workspace should set it. COMET_WORKSPACE is accepted as a deprecated alias.

Pre-release note: opik-mcp (Python) is not yet published to PyPI. Until the first PyPI release lands, replace uvx opik-mcp in any snippet below with: uvx --from git+https://github.com/comet-ml/opik-mcp.git opik-mcp

OPIK_WORKSPACE is optional. Omit the OPIK_WORKSPACE line/key in any snippet below and the server uses the default workspace (correct for local/OSS installs). Set it only if you connect to a named cloud workspace.

Claude Code

Add the server with one command:

claude mcp add --transport stdio opik-mcp \
  --env OPIK_API_KEY=<your-key> \
  --env OPIK_WORKSPACE=<your-workspace> \
  -- uvx opik-mcp

Or edit ~/.claude.json directly:

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

Restart Claude Code. Verify with /mcpopik-mcp should appear as connected. Then, in the chat, ask: "list my Opik projects" — Claude will call the list tool and you'll see your workspace's projects.

Cursor

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (project), or open Cmd+Shift+J → Features → Model Context Protocol:

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

Reload Cursor; the green dot next to opik-mcp in the MCP panel confirms the connection. Ask in chat: "list my Opik projects".

Cursor 60s timeout. Cursor enforces a hard tool-call timeout that doesn't reset on progress notifications. Long ask_ollie turns will fail on Cursor. See Known host limits.

VS Code Copilot

.vscode/mcp.json in your workspace (or User Settings JSON):

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

Reload the window; the Copilot Chat MCP indicator shows opik-mcp once the server is reachable. Ask in chat: "list my Opik projects".

MCP Inspector (manual testing)

OPIK_API_KEY=<your-key> OPIK_WORKSPACE=<your-workspace> \
  npx @modelcontextprotocol/inspector uvx opik-mcp

Self-hosted Opik

Add COMET_URL_OVERRIDE (and OPIK_URL if Opik lives at a non-default path) to the same env block in your host config:

{
  "mcpServers": {
    "opik-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["opik-mcp"],
      "env": {
        "OPIK_API_KEY": "<your-key>",
        "COMET_URL_OVERRIDE": "https://opik.your-company.com",
        "OPIK_MCP_ANALYTICS_SOURCE": ""
      }
    }
  }
}

ask_ollie and run_experiment are available on Comet Cloud only — on self-hosted those calls will fail at dispatch, so use read / list / write directly. Setting OPIK_MCP_ANALYTICS_SOURCE="" opts your install out of the cloud-Comet source label on telemetry events.


Tools

opik-mcp exposes a small, outcome-oriented surface — six tools that cover the full lifecycle (read → annotate → curate → author → iterate).

ToolPurpose
readUniversal read by id / name / opik:// URI
listUniversal list with optional name filter + pagination
ask_ollieInvestigate / synthesize via the Opik in-product assistant
writeUniversal write — log traces/spans, score, comment, save prompts, manage test suites & experiments
schemaIntrospect write-operation schemas (used by the LLM to construct valid payloads)
run_experimentRun an evaluation experiment end-to-end via Ollie

read

One tool for any "show me X" question. Takes an entity_type plus an id (UUID or, for nameable types, a name) or a full opik:// URI. Composite reads (trace, prompt) inline their children so a single call returns the full picture.

Supported entities: project, trace, span, test_suite, experiment, prompt. Name-based lookup is available for project, experiment, prompt, test_suite (slower — two API calls — and may return multiple matches).

read(entity_type="trace", id="7f2e3c8a-…")
read(entity_type="project", id="demo")          # name lookup
read(entity_type="trace", id="opik://traces/7f2e3c8a-…")

list

Browse a collection with optional name filter and pagination. Project-scoped types (trace, test_suite_item, prompt_version) require their parent UUID.

list(entity_type="experiment", page=1, size=25)
list(entity_type="experiment", name="rerank")          # name substring filter
list(entity_type="trace", project_id="<project-uuid>") # traces of one project

ask_ollie

For investigative questions, cross-entity synthesis, or anything that needs Opik domain expertise. Ollie has direct read access to your workspace and can execute writes (scores, comments, test-suite items, prompt versions) mid-stream when asked.

ask_ollie(query="Why are spans in project 'demo' slower this week than last?")
ask_ollie(query="Compare experiments A and B on factuality. Score the bottom 5 traces of A 0.2 with reason.")

Returns the assistant's final text plus a thread_id. Pass it back on follow-ups to preserve context — Ollie has no memory across threads.

YOLO mode (default). Writes Ollie performs mid-stream execute without a per-action confirmation. Each auto-approval is logged as a JSON audit row on the opik_mcp.audit Python logger. To require confirmation instead, set OPIK_MCP_AUTO_APPROVE=disabled — Ollie's confirm requests then surface as typed errors you can manually re-issue.

Available on Comet Cloud only.

write

Universal write dispatcher. Pass operation + data and the dispatcher validates the payload, applies the right REST verb, and returns the backend response.

Operations:

OperationWhat it does
trace.createLog a single trace (or a batch). Parent for spans / scores / comments.
trace.updateFinalize or amend an existing trace.
span.createLog a span on an existing trace (or a batch).
score.createAttach a numeric feedback score to a trace, span, or thread.
comment.createAttach a free-text comment to a trace, span, or thread.
prompt_version.saveSave a new prompt version (creates the prompt by name if missing).
test_suite.createCreate an evaluation test suite.
test_suite_item.upsertUpsert items into a test suite (always the envelope shape).
experiment.createCreate an experiment scoped to a test suite.
experiment_item.createAttach trace + dataset_item rows to an experiment.
write(operation="score.create", data={
  "target": "trace",
  "target_id": "7f2e3c8a-…",
  "name": "helpfulness",
  "value": 0.9,
  "reason": "great recovery"
})

schema

Inspect the exact JSON shape and required fields of any write operation before you call it — useful when you're not sure what data should look like. Returns the schema, OAuth scope, and one validated example. Pure lookup, no backend call.

schema(operation="score.create")
schema(operation="prompt_version.save")

run_experiment

Run an evaluation experiment end-to-end via Ollie. Takes a single experiment_config dict that mirrors Opik's experiment shape (prompt, test suite, scorers); Ollie executes the run and writes results back as an Opik experiment.

run_experiment(experiment_config={
  "test_suite_name": "qa-eval-v2",
  "prompt_name": "welcome-msg",
  # … see `schema(operation="experiment.create")` for the full shape
})

Available on Comet Cloud only.


Configuration

Every setting is an environment variable. Required ones in bold.

Identity / endpoint

VariableDefaultNotes
OPIK_API_KEYRequired for ask_ollie and any authenticated read/write.
OPIK_WORKSPACEdefaultWorkspace name. Optional — falls back to default (Opik SDK convention). Cloud users with a named workspace should set it.
COMET_WORKSPACEDeprecated alias for OPIK_WORKSPACE (backward compat). OPIK_WORKSPACE wins if both are set.
COMET_WORKSPACE_IDOptional workspace UUID. Stamped into analytics events when set so BI can join on a stable id rather than the (mutable) workspace name.
COMET_URL_OVERRIDEhttps://www.comet.comSet to your self-hosted Comet host, or https://dev.comet.com for staging.
OPIK_URLderived from COMET_URL_OVERRIDE + /opik/apiOverride only if Opik lives on a different host/path than the Comet UI.
OPIK_DEFAULT_PROJECT_NAMEunsetWhen set, the per-session instructions blob tells the LLM to pass this as project_name on every tool call unless the user names a different project.

Server / transport

VariableDefaultNotes
OPIK_MCP_TRANSPORTstdiostdio for host-launched, streamable-http to listen on a port.
OPIK_MCP_HOST127.0.0.1uvicorn bind host (streamable-http only).
OPIK_MCP_PORT8080uvicorn bind port (streamable-http only).
OPIK_MCP_RELOADfalsetrue to enable uvicorn --reload (dev only).
OPIK_MCP_AS_URLunsetOAuth Authorization Server URL, advertised in /.well-known/oauth-protected-resource (RFC 9728) and used as the proxy target for AS-discovery probes. Required for MCP hosts to bootstrap the OAuth dance over HTTP.
OPIK_MCP_RESOURCE_URIunsetCanonical public URI of this server, advertised as resource in the protected-resource metadata and used to derive the WWW-Authenticate hint.
OPIK_MCP_LOG_LEVELINFOstderr logger threshold.

Choosing a transport

opik-mcp performs no local credential validation on HTTP transport: any well-formed Authorization: Bearer … (an Opik API key or an opik_at_… OAuth access token) is forwarded verbatim to opik-backend, which is the single point of auth enforcement. Pick the transport by deployment shape:

ScenarioTransport
MCP client and Opik on the same machine (local OSS install)stdio (recommended — simplest, no port, no OAuth setup)
Local MCP client → remote Opik (Comet cloud / self-hosted)stdio with OPIK_API_KEY, or HTTP with OAuth (OPIK_MCP_AS_URL pointing at the backend)
Hosted opik-mcp behind the same edge as opik-backendHTTP — bearers are validated by the backend per request

Note for local OSS installs: the OSS backend does not authenticate requests, so an HTTP opik-mcp in front of it is as open as the OSS REST API itself. Keep the default 127.0.0.1 bind (and prefer stdio) on shared networks.

Ollie / long calls

VariableDefaultNotes
OPIK_MCP_AUTO_APPROVEenableddisabled to require a per-action approval before Ollie's mid-stream writes proceed. On hosts that advertise the MCP elicitation capability the user sees a yes/no prompt; on dumber hosts the request surfaces as a typed error you can manually re-issue.
OPIK_MCP_ELICIT_TIMEOUT_SECONDS60How long Ollie's mid-stream confirmation prompt may wait for the user before being treated as a cancel. 0 disables the bound (debug only).
OPIK_MCP_POD_READY_TIMEOUT_S120Ollie pod cold-start poll cap.
OPIK_MCP_POD_READY_INTERVAL_S2Cold-start poll interval.
OPIK_MCP_HEARTBEAT_INTERVAL_S15.0Watchdog cadence — emits a notifications/progress tick when the pod is silent, keeping host timeouts at bay.
OPIK_MCP_STREAM_IDLE_TIMEOUT_S300.0Hard ceiling on pod silence before ask_ollie aborts. 0 disables (debug only).

Telemetry

Anonymous usage events (event type + timing only — no query content). A SHA-256 digest of your API key is included so support can find your account; the raw key never leaves the process. Opt out: OPIK_MCP_ANALYTICS_ENABLED=false.

VariableDefaultNotes
OPIK_MCP_ANALYTICS_ENABLEDtrueSet to false to disable all telemetry.
OPIK_MCP_ANALYTICS_URLhttps://stats.comet.com/notify/event/Override for staging.
OPIK_MCP_ANALYTICS_ENVIRONMENTprodTag on every event (prod / staging / dev).
OPIK_MCP_ANALYTICS_SOURCEcomet.comReceiver uses this to mark on_prem=False. On-prem installs should override to "" or their own domain.
OPIK_MCP_ANALYTICS_CONNECT_TIMEOUT_S5.0HTTP connect timeout.
OPIK_MCP_ANALYTICS_TOTAL_TIMEOUT_S10.0HTTP total request timeout.

Known host limits

The MCP spec lets hosts reset their tool-call timeout on notifications/progressopik-mcp emits one per Ollie SSE event plus a 15-second watchdog heartbeat. Reality is uneven:

  • Claude Code — no documented tool-call timeout; heartbeat keeps the call alive until message_end. Recommended.
  • Cursor — hard 60s timeout that does not reset on progress (upstream bug). Long Ollie turns will fail. Keep ask_ollie queries focused.
  • MCP InspectorMAX_TOTAL_TIMEOUT bounds total duration (default 60s). Raise it in the Inspector UI for long operations.

If a call gets stuck, set OPIK_MCP_LOG_LEVEL=DEBUG — heartbeat failures (usually host disconnects) are logged on opik_mcp.ask_ollie at debug level.


Troubleshooting

OPIK_API_KEY is required to use ask_ollie — the var isn't reaching the server process. In Claude Code / Cursor / VS Code, env vars only apply when inside the env block of the MCP server config, not your shell. Restart the host after editing.

ask_ollie returns "pod not ready" after 2 minutes — the Ollie pod cold-start exceeded OPIK_MCP_POD_READY_TIMEOUT_S. Retry — the second call usually hits a warm pod.

ask_ollie / run_experiment fails with a dispatch error on self-hosted Opik — those tools are available on Comet Cloud only. Use read / list / write directly on self-hosted.

Cursor call times out at 60s — Cursor's known bug, not opik-mcp. Either shorten the Ollie query, or run the same operation on Claude Code which has no hard cap.


Development

git clone git@github.com:comet-ml/opik-mcp.git
cd opik-mcp
make install        # uv sync --extra dev
make check          # lint + typecheck + test
make run-dev        # uvicorn with --reload + DEBUG logs
make inspect        # MCP Inspector against the running server

Common targets:

TargetWhat it does
make installuv sync --extra dev
make runRun the MCP server (stdio by default).
make run-devRun with DEBUG logging + uvicorn --reload.
make devRun via mcp dev (Inspector dev-mode wrapper).
make inspectLaunch MCP Inspector against a running server.
make testuv run pytest -q.
make test-liveLive end-to-end against dev.comet.com (set OPIK_API_KEY + OPIK_WORKSPACE).
make lintruff check + format check.
make formatruff format + ruff check --fix.
make typecheckmypy.
make checklint + typecheck + test.

Repo layout:

opik-mcp/
├── src/opik_mcp/        ← server, tools, ask_ollie, analytics
├── tests/               ← pytest suites
├── scripts/             ← live-BE smoke + MCP-session smoke
├── legacy/typescript/   ← deprecated v2 TS server
├── pyproject.toml
└── Makefile

Get help


Upgrading from v2? The legacy TypeScript server still ships on npm as opik-mcp@^2 (npx -y opik-mcp); source is preserved under legacy/typescript/. See legacy/typescript/DEPRECATED.md for the support policy.


License

Apache-2.0.