Opik MCP Server

The official Model Context Protocol (MCP) server for Opik, the open-source LLM observability and evaluation platform, built by Comet. Plug your AI host (Claude Code, Cursor, VS Code Copilot, Codex, opencode, or any MCP client) directly into your Opik workspace: read traces, log scores, and save prompt versions, 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:    "Which traces in project 'demo' failed today?"
Claude: → list(entity_type="trace", project_name="demo") → "Three traces failed…"

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

Quick start

One command sets Opik up in the AI clients on your machine, Claude Code included. It finds your Opik (the one saved in ~/.opik.config, one running on localhost, or it asks which), registers the MCP server, installs the Opik skill pack, and checks the connection. On Opik Cloud there is no API key to handle: your client signs in in the browser. It needs uv and no Opik SDK:

uvx opik@latest mcp configure

Without a terminal, as from a coding agent or a script, name the client:

uvx opik@latest mcp configure --ai-client claude-code --skills

--ai-client also takes codex, cursor, vscode and opencode. Run that way it uses the Opik saved in ~/.opik.config, API key included, or one answering on localhost, and otherwise stops and asks for --deployment.

A coding agent asked to "install opik mcp" gets the full steps, including what to ask the user, from uvx opik-mcp@latest --help. They are the same as Install with a coding agent below.

Registering the server by hand, with the client commands under Set it up by hand, skips the check of which Opik you use and the skill pack, so it is for clients the installer doesn't support.

Clients load MCP servers when a session starts, so start a new session afterwards.

Setup guide, troubleshooting and FAQ: comet.com/docs/opik/mcp-server.

Opik Cloud: the hosted server

Comet runs the server at https://www.comet.com/opik/api/v1/mcp. Your client connects over HTTP and opens a browser sign-in the first time. There is nothing to install, no API key, and no workspace to set: the server works in the workspace you pick when you sign in. After adding it, start a new session and ask: "list my Opik projects".

The Quick start command sets this up for Claude Code, Codex, Cursor, VS Code Copilot and opencode, with the skill pack. In Claude Code, the plugin does the same in one command. The buttons below add it to Cursor or VS Code in one click; other ways to register it by hand are under Set it up by hand.

Claude Code plugin

The opik plugin installs the Opik skills and registers the MCP server, by default the hosted one:

/plugin install opik --marketplace comet-ml/opik-mcp

On Claude Code before 2.1.275, run /plugin marketplace add comet-ml/opik-mcp, then /plugin install opik@opik. The skills load as opik:<skill>, such as opik:opik-instrument, and /mcp shows the server as plugin:opik:opik-mcp. The install asks for the Opik MCP server URL: keep the default for Opik Cloud, or enter https://<host>/opik/api/v1/mcp for a self-hosted Comet that runs the MCP OAuth server. /plugin configure opik@opik changes it later. Start a new session, then sign in from /mcp → Authenticate. For any other self-hosted or open-source Opik, use the Quick start command. claude.ai chat ignores a URL built from a setting, so there the plugin adds the skills only. Claude Code does not update a plugin from this marketplace on its own: run /plugin marketplace update opik, then /plugin update opik@opik, or turn on auto-update for the marketplace in /plugin. This marketplace serves main, which can be ahead of the last release. An update arrives only when the plugin's version changes, at each release.

If you added opik-mcp before at the same URL, Claude Code keeps your copy and skips the plugin's. If yours is another server, such as the local uvx opik-mcp, both load: run claude mcp remove opik-mcp to keep only the plugin's. Skills from the skill pack also load next to the plugin's copies; remove them with npx skills remove --global --agent claude-code, which asks which ones.

Self-hosted and open-source Opik: the local server

opik-mcp runs on your machine: the client starts it with uvx opik-mcp and talks to it over stdio. The Quick start command sets it up from your ~/.opik.config or a local Opik; to register it by hand, see Set it up by hand. Install uv once; it fetches the package, and Python 3.13 if needed, on first use:

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

uvx opik-mcp reuses the copy uv has cached, so it starts quickly; uv cache clean opik-mcp makes the next start fetch the newest release. uvx opik-mcp@latest asks PyPI on every start, which adds about a second and a half. Codex waits mcp_optional_startup_grace_ms (1 s by default) for servers before it builds the first tool list, so a slower start can leave the tools out of the first turn. The one-off commands in this README (opik mcp configure, opik mcp status, --help, --check) carry @latest, so they never run a copy uv cached months ago.

Env vars point the server at your Opik:

• Open source serves its API at /api and has one workspace, default, so it needs no OPIK_WORKSPACE. COMET_URL_OVERRIDE would point the server at /opik/api, which open source does not serve.
• A self-hosted Comet has named workspaces. Set OPIK_WORKSPACE to the segment after /opik/ in your Opik URL; left out, reads come from your account's default workspace, with no error to say so.
• A self-hosted Comet that runs the MCP OAuth server (off by default) can use the hosted flow instead, at https://<host>/opik/api/v1/mcp.
• Substitute every value. The server refuses a placeholder workspace such as <your-workspace> or ${input:OPIK_WORKSPACE}; a placeholder URL is sent as-is and fails to connect.
• OPIK_MCP_ANALYTICS_SOURCE="" in the env block opts a self-hosted install out of the cloud-Comet source label on telemetry events.

Before adding the server to a client, check the settings: run it with the same env and --check. It makes one call to Opik and prints the workspace and how many projects it sees, or what to fix with a link to the docs, and exits 0 or 1. A client can't tell you this before a new session: Claude Code shows Connected even with a wrong key, and Codex starts nothing until then.

OPIK_URL=http://localhost:5173/api uvx opik-mcp@latest --check

Then register it with your client: the local server, by hand.

Install with a coding agent

For an AI agent asked to install the server, as in "install opik mcp". The aim is the server registered for the user's Opik, the Opik skill pack installed, and the user signed in. Work out which Opik from this machine, and ask the user only what it can't tell you: one question per message, each answer before the next. The installer in step 3 does the whole setup in one run, so settle steps 1 and 2 before you run it. uvx opik-mcp@latest --help prints these steps.

1. See what is there.

   uvx opik@latest mcp status

   It shows the Opik saved in ~/.opik.config (URL and workspace, never the key) and the clients that already have the server. In Claude Code, also run claude mcp list: it shows claude.ai connectors and entries under other names. It also prints each entry's command, which for an old npx entry can hold an API key: don't repeat it.

   • A server for https://www.comet.com/opik/api/v1/mcp under any name, such as a claude.ai connector, means it is registered: don't add a second one. Install the skill pack on its own with npx skills add comet-ml/opik-skills -g --all. Go to step 4 if the server shows ! Needs authentication.
   • The installer replaces an existing opik-mcp entry without asking, such as one for the old npx server. Ask the user before you run it.

2. Find the user's Opik.

   • Saved in ~/.opik.config: use it, and tell the user which one you are setting up. The installer reads it, API key included, so don't ask for a key.
   • No ~/.opik.config: curl -s http://localhost:5173/api/is-alive/ping. An answer is an open-source Opik on this machine. In a sandboxed shell, such as Codex's, no answer can mean the shell has no network rather than no Opik.
   • Neither: ask, and don't pick one yourself. An Opik URL in the project, such as OPIK_URL_OVERRIDE in .env, makes it a yes-or-no question. Otherwise: "Which Opik do you use: Opik Cloud (www.comet.com), an open-source Opik (its URL), or a self-hosted Comet platform (its URL and workspace)?"

3. Run the installer for the client you run in. It installs the Opik skill pack with the server, with or without --skills; keep the flag in a command you give the user to run later. On Opik Cloud it opens the browser sign-in and waits a minute for it, so tell the user first.

   uvx opik@latest mcp configure --ai-client claude-code --skills

   --ai-client also takes codex, cursor, vscode and opencode. For an Opik saved in ~/.opik.config or running on localhost, run it as it is: --deployment sets the saved config aside, API key included. With neither, add the flag for the Opik the user named in step 2:

   The user's Opik	Add
   Opik Cloud, www.comet.com	--deployment cloud
   Open-source Opik	--deployment local, and --url <url> when it is not on localhost
   A self-hosted Comet platform, /opik/ in its URL	--deployment self-hosted --url <url> --workspace <workspace>, with OPIK_API_KEY in the shell

   If it fails, as when the Opik doesn't answer, tell the user what it printed and what to check, such as a VPN or the URL. Don't switch to another Opik.

4. Sign-in, for Opik Cloud. If the installer reports the server as not signed in yet, give the user claude mcp login opik-mcp to run, or /mcp in a Claude Code session, rather than running it yourself: it waits on the browser. A claude.ai connector signs in from /mcp. The sign-in asks for a workspace; name the one saved in ~/.opik.config, if any.

5. Finish. Say what was installed and what is left. Name the skill pack: the Opik skills, installed for every project, which teach the agent to instrument, evaluate and debug with Opik. Clients load MCP servers and skills when a session starts: ask the user to start a new session, then try "list my Opik projects".

On every path:

• Don't ask the user to paste an API key into the chat. The hosted server needs none; otherwise take it from the shell environment ("$OPIK_API_KEY"), or let the user run the command.
• Don't guess the workspace. It is the segment after comet.com/opik/ (or /opik/ on a self-hosted Comet) in the user's Opik URL.
• Don't print ~/.claude.json or ~/.opik.config: both can hold API keys. uvx opik@latest mcp status and claude mcp list show what you need, and the installer reads the config itself.
• If a permission check blocks a command, don't work around it. Give the user the command to run (in Claude Code, ! before it runs it from the prompt).

For a client the installer doesn't know, register by hand with the commands in Set it up by hand: the hosted server for Opik Cloud, the local server for the rest.

• Before registering the local server, run it with the env you will pass and --check, with OPIK_API_KEY exported in the shell rather than written on the command: OPIK_URL=http://localhost:5173/api uvx opik-mcp@latest --check. It exits 0 and names the workspace it sees, or prints what to fix. Once registered, Claude Code reports Connected even with a wrong key, and Codex has no check.
• Claude Code refuses to add over an existing opik-mcp entry, so remove it first with claude mcp remove opik-mcp --scope user; Codex's add replaces it.
• After registering, in Claude Code, read only the status line: claude mcp get opik-mcp | grep Status. The full output prints the env block, API key included. The hosted server shows ! Needs authentication until claude mcp login opik-mcp has run. Codex has nothing that starts the server before a session; codex mcp get opik-mcp shows what was stored, with env values masked.

Set it up by hand (no Opik check, no skill pack)

The Quick start command does all of this for Claude Code, Codex, Cursor, VS Code Copilot and opencode, and also checks which Opik you use and installs the skill pack. Register by hand only for another client, or to see what the installer writes. A coding agent installing for a user runs the installer: see Install with a coding agent.

Hosted server (Opik Cloud)

Claude Code, by hand

claude mcp add --scope user --transport http opik-mcp https://www.comet.com/opik/api/v1/mcp

Then sign in from /mcp → Authenticate in a Claude Code session. It opens a browser, so the person at the keyboard does it; from their own terminal, claude mcp login opik-mcp does the same. --scope user makes the server available in every project; without it, Claude Code registers it for the current directory only. Over SSH, claude mcp login opik-mcp --no-browser prints the sign-in URL to open on your own machine; the last step needs an interactive terminal (ssh -t).

Codex, by hand

codex mcp add opik-mcp --url https://www.comet.com/opik/api/v1/mcp

add opens the sign-in; codex mcp login opik-mcp opens it again. In ~/.codex/config.toml the same server is:

[mcp_servers.opik-mcp]
url = "https://www.comet.com/opik/api/v1/mcp"

Cursor, by hand

Use the button under the hosted server, or add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "opik-mcp": {
      "url": "https://www.comet.com/opik/api/v1/mcp"
    }
  }
}

VS Code Copilot, by hand

Use the button under the hosted server, or add to .vscode/mcp.json in your workspace or to your user mcp.json (MCP: Open User Configuration):

{
  "servers": {
    "opik-mcp": {
      "type": "http",
      "url": "https://www.comet.com/opik/api/v1/mcp"
    }
  }
}

Other clients

Any client that takes a URL can use the hosted server:

npx add-mcp https://www.comet.com/opik/api/v1/mcp --name opik-mcp

A client that can only start local commands can reach it through npx -y mcp-remote https://www.comet.com/opik/api/v1/mcp.

Opik Cloud with an API key

Where nobody can complete the browser sign-in, as when an agent runs unattended from a script or a client has no MCP OAuth support, use an API key from comet.com/api/my/settings/ instead, with the local server pointed at Opik Cloud:

claude mcp add --scope user opik-mcp \
  --env OPIK_API_KEY="$OPIK_API_KEY" \
  --env OPIK_WORKSPACE=<workspace> \
  -- uvx opik-mcp

The other clients take the same two variables in their env block. Set OPIK_WORKSPACE to the segment after comet.com/opik/ in your Opik URL (https://www.comet.com/opik/acme-ai/projects → acme-ai). Left out, the server sends default, which Comet resolves to your account's default workspace, so reads can come from the wrong workspace without an error.

Local server (self-hosted and open-source Opik)

The examples use a local open-source Opik. For a self-hosted Comet, swap in the three variables from the table. After adding the server, start a new session and ask: "list my Opik projects".

Claude Code, by hand

claude mcp add --scope user opik-mcp --env OPIK_URL=http://localhost:5173/api -- uvx opik-mcp

On a self-hosted Comet, with the key in your shell's OPIK_API_KEY:

claude mcp add --scope user opik-mcp \
  --env COMET_URL_OVERRIDE=https://<host> \
  --env OPIK_WORKSPACE=<workspace> \
  --env OPIK_API_KEY="$OPIK_API_KEY" \
  -- uvx opik-mcp

Or edit ~/.claude.json directly:

{
  "mcpServers": {
    "opik-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["opik-mcp"],
      "env": {
        "OPIK_URL": "http://localhost:5173/api"
      }
    }
  }
}

claude mcp get opik-mcp shows ✔ Connected once the server starts. That does not prove the URL or key are right, because only a tool call reaches Opik; --check (the local server) does. It also prints the env block, API key included.

Codex, by hand

codex mcp add opik-mcp --env OPIK_URL=http://localhost:5173/api -- uvx opik-mcp

Or edit ~/.codex/config.toml:

[mcp_servers.opik-mcp]
command = "uvx"
args = ["opik-mcp"]
env = { OPIK_URL = "http://localhost:5173/api" }
# On a self-hosted Comet, forward the key from the environment Codex starts in,
# so it stays out of this file:
# env_vars = ["OPIK_API_KEY"]
# Time allowed for the server to start (default 10 s); the first start
# downloads the package.
startup_timeout_sec = 30

Codex gives a local server only the variables in env and the names in env_vars, so a key exported in your shell does not reach it otherwise.

Cursor, by hand

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

{
  "mcpServers": {
    "opik-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["opik-mcp"],
      "env": {
        "OPIK_URL": "http://localhost:5173/api"
      }
    }
  }
}

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

VS Code Copilot, by hand

Add to .vscode/mcp.json in your workspace, or to your user mcp.json (MCP: Open User Configuration):

{
  "servers": {
    "opik-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["opik-mcp"],
      "env": {
        "OPIK_URL": "http://localhost:5173/api"
      }
    }
  }
}

MCP Inspector (manual testing)

OPIK_URL=http://localhost:5173/api npx @modelcontextprotocol/inspector uvx opik-mcp

Coming from npx opik-mcp?

The TypeScript server (npm opik-mcp@2) is deprecated and stops serving requests on 2026-11-15. For a new install, run the Quick start command; uvx opik-mcp is the command a client runs to start the server, not an installer. To move an existing config on Opik Cloud, switch to the hosted server and drop the API key. Otherwise, in your MCP client config, replace npx -y opik-mcp with uvx opik-mcp. Some env vars were renamed and command-line flags are no longer read: see the migration guide. Support policy: DEPRECATED.md. The TypeScript source is at the git tag legacy-typescript-final.

Tools

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

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, thread, agent_insights_issue) inline their children so a single call returns the full picture.

The record you name comes back whole. Inlined children do not: their bodies are fetched with the backend's truncate=true, so a field over ~10 KB is cut in ClickHouse and base64 images are replaced with "[image]" — one attachment echoed across 200 spans would otherwise cost more than everything else in the read. The answer says so in spanBodies / messageBodies, and any child is whole again through its own read("span", id) or read("trace", trace_id), which hit endpoints that have no truncate parameter at all.

An inlined collection is also bounded in length: 200 spans, 200 turns, 100 prompt versions. Past that, spansTruncated / messagesTruncated / versionsTruncated is true and a moreSpans / moreMessages / moreVersions line beside it carries the count and the exact list(...) call that continues from where the inlined part stopped.

Supported entities: project, trace, span, dataset, dataset_item, experiment, prompt, thread, agent_insights_issue. Name-based lookup is available for project, experiment, prompt, dataset (slower — two API calls — and may return multiple matches). thread and agent_insights_issue are project-scoped: pass project_id or project_name, or a link/URI that carries the project. dataset and dataset_item were called test_suite and test_suite_item before; the old names still resolve, but they are not advertised and new code should use the new ones.

read(entity_type="trace", id="7f2e3c8a-…")
read(entity_type="project", id="demo")  # name lookup
read(entity_type="trace", id="opik://traces/7f2e3c8a-…")
read(entity_type="agent_insights_issue", id="<issue-uuid>", project_id="<project-uuid>")
read(
    entity_type="agent_insights_issue",
    id="https://www.comet.com/opik/<ws>/projects/<pid>/diagnostics?issue=<id>",
)

A link copied from the Opik UI works as the id: a thread link or a Diagnostics page link carries the project, so no project_id is needed and the entity type is taken from the link.

A project read answers "how is my project doing" in one call. It returns {project, summary, vocabulary, contains, url}: the record, then the four figures the Logs page shows as cards (trace count, error rate, average duration, total cost) for the last 7 days against the 7 before, SDK traffic only, as on screen. since / until move that window; since="30d" is what the UI opens on. A rate or an average over a period with no traces comes back as null, because 0% errors on a week with no traffic reads as a healthy week.

vocabulary is the map you need before you can ask anything else: the project's feedback score names, its token usage keys, and the automation rules scoring its traces. These are the names that go into a filter or into series= below, and guessing them returns an empty page that reads like good news. Score names and rules are capped, always report the true total, and name the call that returns the rest; usage keys are listed in full, since nothing else enumerates them. contains names the freshest experiment, dataset, prompt version and optimization run, so "what has been happening here" does not need four more calls. A part that failed to load says so instead of looking empty, and an empty one is omitted.

An agent_insights_issue read returns {issue, example_trace_ids, details}: the Diagnostics issue record (name, description, cause, suggested fix, severity, status), the deduplicated ids of the traces that exhibit it (the same sample the Diagnostics page shows — open one with read("trace", id)), and the per-day breakdown. Trace bodies are not inlined, so the read stays one backend call. since / until narrow the per-day rows; the default is all-time. When the server knows the Opik URL and the session's workspace, the read also carries url (the issue's Diagnostics page) and trace_url_template (a deep link for any of the example traces), so the assistant can hand you something clickable; under an OAuth session whose workspace could not be resolved the links are omitted rather than guessed.

Traces themselves carry no URL — a link for one is not derivable from the fields a read or list returns, and a guessed shape 404s. The session instructions name a template for it instead, .../v1/session/redirect/projects/?trace_id={trace_id}&path=..., so the assistant fills in an id and hands you a link. It goes through opik-backend's redirect, which resolves the project and the workspace from the trace, so it works where a direct project URL cannot, an OAuth session with an unresolved workspace included. It is the same link the Python SDK prints for a trace.

list

Browse or search a collection with pagination. Project-scoped types (trace, span, thread, agent_insights_issue, dataset_item, prompt_version) need their parent: a project UUID or name, a dataset UUID, or a prompt UUID.

list(entity_type="experiment", page=1, size=25)
list(entity_type="experiment", name="rerank")  # name substring filter
list(entity_type="agent_insights_issue", project_name="demo")  # open Diagnostics issues
list(entity_type="agent_insights_issue", project_id="<uuid>", status="resolved")
list(entity_type="trace", project_name="demo")  # latest traces of one project
list(
    entity_type="trace", project_name="demo", filters="error_info is_not_empty AND duration > 5000"
)
list(
    entity_type="span",
    project_name="demo",  # spans across the whole project
    filters='type = "llm" AND usage.total_tokens > 10000',
)
list(
    entity_type="thread",
    project_name="demo",
    filters="number_of_messages > 20 AND feedback_scores.helpfulness < 0.5",
)
list(entity_type="experiment", filters='dataset_id = "<dataset-uuid>" AND tags contains "baseline"')

Filters. trace, span, thread, experiment and dataset_item take an OQL string, the same grammar as the SDK's search_traces(filter_string=…):

<field>[.<key>] <op> <value> [AND ...]
ops: = != > >= < <= contains not_contains starts_with ends_with is_empty is_not_empty in not_in

Strings go in double quotes, numbers are bare, duration is in milliseconds, dates are ISO-8601 instants with a timezone ("2026-09-08T10:00:00Z"). Scores and dictionaries take a key: feedback_scores.accuracy < 0.5, metadata.environment = "prod". AND is the only connector.

Like the UI's Logs page, trace, span and thread lists add source = "sdk" so evaluator, playground and experiment traces stay out of the way; name source yourself to see them. The first output line echoes the filter that was applied.

A bad filter fails before reaching the backend with what is needed to fix it: the position of a syntax error, the closest field name, the valid operators for the field's type, or the expected value format. Fields with a closed set of values (source, span type, thread status, visibility_mode) are checked against it too, every element of an in list included. source is the one the backend validates itself, and it answers an unknown value with a 500 rather than a 400, so source = "SDK" would otherwise be an opaque server error for a capital letter. The rest are compared as strings and answer with an empty page, which reads as "no matches" when it means "no such value". Ask schema("list.trace") (or list.span, list.thread, list.experiment) for the full field reference, accepted values included.

Finding one case in a dataset. list(entity_type="dataset_item", dataset_id=…) filters on the case itself: data.<key> for the keys the dataset was built with, full_data for a substring of the whole payload (a full scan — name a key when you can), plus id, tags, source, trace_id, span_id and the timestamps. data.<key> takes the six string operators only (=, !=, contains, not_contains, starts_with, ends_with); the backend answers a comparison with a 400, so this one is refused before the call. The endpoint has no sorting and no free-text search — sort is refused rather than dropped. read(entity_type="dataset_item", id=…) returns one case whole, which is how a value the table cut is read back.

list(entity_type="dataset_item", dataset_id="<uuid>", filters='data.question contains "install"')
list(
    entity_type="dataset_item", dataset_id="<uuid>", filters='trace_id = "<trace-uuid>"'
)  # the case made from that trace
read(entity_type="dataset_item", id="<item-uuid>")  # the case, uncut

With experiment_ids the same list is the comparison instead — the cases with each run attached — and it filters on the runs (feedback_scores.<name>, output, duration). The two are different field sets on two backend endpoints: schema("list.dataset_item_case") is the dataset's own cases, schema("list.dataset_item") the comparison.

Sort. trace, span, thread and experiment take sort="<field> [asc|desc]", desc by default and one field only: sort="duration desc", sort="total_estimated_cost", sort="feedback_scores.accuracy asc", sort="usage.total_tokens". The field is checked against the entity's sortable list before the call, because the backend silently ignores fields it cannot sort by. On very large workspaces the backend drops sorting altogether; the header says so when that happens.

dataset_item sorts only as a comparison (with experiment_ids): the items endpoint takes no sorting parameter, so a sort on a plain listing is refused rather than dropped.

Time window and search. trace, span and thread take since and until, each a relative span ("30m", "1h", "7d") or an ISO-8601 instant with a timezone, so "the last hour" needs no clock arithmetic. The window is by record creation time, which is cheap for the backend and agrees with start_time within seconds for live traffic. For an exact bound, put start_time in filters. The same three types take search, free text matched anywhere in id, name, input, output, metadata, tags and thread id. Search scans the whole project on the backend, so the first call on a large project can take tens of seconds. Those calls get a 60-second timeout. Adding since makes them fast again.

Reading the table. Durations are labelled duration_ms / ttft_ms and shown as whole milliseconds; the field stays duration in filters and sort. Timestamps are shown to the second and costs as plain decimals. Project rows carry last_updated_trace_at so you can see which project has live traffic; thread rows carry the first message. An empty page under a time window says when the project's last trace landed, and an empty page under the default source = "sdk" says how to see the other sources. A misspelled project_name comes back with the closest existing name.

list(
    entity_type="trace",
    project_name="demo",
    since="1h",
    filters="error_info is_not_empty",
    sort="duration desc",
)
list(entity_type="trace", project_name="demo", search="order-42")

Diagnostics issues. agent_insights_issue is the Diagnostics page over the MCP: the recurring failures Opik's Diagnostics job grouped for a project, ranked as the UI ranks them (most recently seen first). Columns are severity, status, total_occurrences (all-time sum), latest_count (the most recent report day, the number the issue's own description refers to) and last_seen. Open issues are listed by default; pass status="resolved" or "closed" for the rest. read and list also answer to issue, which is what the UI calls these; the long name is the one in the entity_type enum, so that one entity does not appear there twice. Counts are all-time so they match the UI; the same since / until as for traces narrow the window, truncated to UTC report days because Diagnostics aggregates per day.

An empty list says why it is empty, because "nothing is broken" and "nobody turned Diagnostics on" read the same otherwise. There are five states: Diagnostics is unavailable on this deployment, not enabled for this project, turned off, enabled but not scanned recently, or enabled and clean with the time of the last scan. The ones you can act on name the call to make, and every state links the project's Diagnostics page.

A non-empty list dates itself. The issues are whatever the last scan grouped, so the reply ends with Report covers data through <time>, and when the window you asked about runs past that, it names the uncovered tail and how to close it: a trigger when a rescan reaches back far enough, otherwise raw traces with the since it gives you. Ask for a week on a project scanned nightly and the last day is missing from the grouped answer; this is what says so.

write("agent_insights_job.enable", {"project_name": "demo"}) turns Diagnostics on. It scans daily from then on, and calling it again is safe. write("agent_insights_job.trigger", …) scans the last 24 hours now, without waiting for the nightly run. Both take the permission that reading issues takes, and both refuse where the deployment has no Diagnostics.

An issue moves through its lifecycle with write("agent_insights_issue.resolve", {"issue_id": "<uuid>", "project_name": "demo"}) — dealt with — or …close for one not worth acting on, and …reopen to put either back on the open list. All three take the same permission and answer with a link to the view the issue moved to, since a resolved issue is no longer on the default page. Whether a failure is fixed is a judgment call, so these are for when you ask: the assistant has no business tidying the list while triaging it.

Metrics over time. project_metric charts one metric for a project as a table of time buckets: trace, span and thread counts, durations, error rates, costs, token usage and feedback scores. It answers the question that follows the overview, which is when something changed.

list(entity_type="project_metric", project_name="demo", metric_type="trace_count")
list(
    entity_type="project_metric",
    project_name="demo",
    metric_type="trace_error_rate",
    since="14d",
    interval="daily",
)
list(
    entity_type="project_metric", project_name="demo", metric_type="span_count", breakdown="model"
)  # one column per model
list(
    entity_type="project_metric",
    project_name="demo",
    metric_type="span_duration",
    breakdown="model",
    series="p99",
)  # the p99 of each model

Rows are time buckets, not records, so page, size and sort are refused rather than ignored. interval is hourly, daily, weekly or total; left out, it follows the window the way the Metrics tab does — hourly up to 3 days, daily up to 30, weekly beyond — so a default chart is a few dozen rows whatever the range, and an hourly month (721 rows) is something you ask for. since / until take the same forms as everywhere else and default to the last 7 days. filters uses the fields of whichever entity the metric is about, so a span metric is filtered by span fields.

breakdown splits each bucket by tags, name, error_info, error_type, model, provider, span_type, guardrail_name or metadata.<key>. Not every metric accepts every one of those, and seven accept none at all; the tool knows which and says so before calling the backend, naming a metric that does answer the same question where one exists. Three families come back as several series at once (a duration as p50/p90/p99, a feedback score per name, token usage per key), and the backend charts one of them at a time when grouping, so series= picks it: a percentile, a score name, or a usage key. Duration defaults to p50 and token usage to total_tokens, and whichever was used is echoed on the first line.

Empty buckets are left out and counted underneath, so a quiet month is a few rows instead of a column of zeros, and a rate over a bucket with no traces is absent rather than reported as zero.

Ask schema("list.project_metric") for the metric table, the intervals and the per-metric grouping matrix.

A project's names. score_name lists the feedback score names recorded in a project and online_rule the automation rule evaluators configured on it, which is where most of those names come from. Both are the same lists read("project", …) carries, in full and paginated, for when the capped version in the overview is not enough.

list(entity_type="score_name", project_name="demo")
list(entity_type="online_rule", project_name="demo")

write

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

Operations:

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")

The same tool answers list.trace, list.span, list.thread and list.experiment with the list tool's reference for that entity: every filterable field with its type and valid operators, the sortable fields, whether a time window and free-text search apply, and two example filters.

schema(operation="list.trace")

Configuration

These configure the local server; every setting is an environment variable. The hosted server on Opik Cloud takes none of them.

Identity / endpoint

Server / transport

Choosing a transport

Two bearer shapes, two contracts on HTTP transport. An opik_mcp_at_… OAuth access token is validated on every request against opik-backend's token introspection endpoint (cached, see OPIK_MCP_OAUTH_VALIDATION_CACHE_TTL_S); an expired or revoked token gets an HTTP 401 with WWW-Authenticate: Bearer error="invalid_token", which is what MCP hosts key their silent refresh_token grant on. An Opik API key is not validated locally: it is forwarded verbatim to opik-backend, which is its single point of enforcement. Pick the transport by deployment shape:

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.

Telemetry

Known host limits

Hosts differ in how long they let a single tool call run:

• Claude Code — no documented tool-call timeout. Recommended.
• Cursor — hard 60s timeout that does not reset on progress (upstream bug).
• MCP Inspector — MAX_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 for the full request log.

Troubleshooting

Does the setup work? Run the server with the env from its client config and --check: OPIK_URL=http://localhost:5173/api uvx opik-mcp@latest --check. It names the workspace and the projects it sees, or says what to fix. The server's own errors also say what to change, and link the docs page for it.

OPIK_API_KEY isn't picked up — 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; Codex also forwards the names listed in env_vars. Start a new session after editing, since clients read the config when a session starts.

Requests go to /opik/api on an open-source Opik — COMET_URL_OVERRIDE is for a self-hosted Comet platform. Open source serves its API at /api: set OPIK_URL=http://localhost:5173/api (or https://<host>/api) instead. The server warns about this at startup when the host is this machine, and --check answers "No Opik API at …".

Cursor call times out at 60s — Cursor's known bug, not opik-mcp. Either narrow the call (smaller size, a tighter window), or run the same operation on Claude Code which has no hard cap.

Server not showing, sign-in not opening, wrong workspace, uvx not found. These are covered in the troubleshooting section of the docs. uvx opik@latest mcp status lists every client that has the server configured and whether its config has drifted.

Development

git clone [email protected]:comet-ml/opik-mcp.git
cd opik-mcp
make install        # uv sync --locked --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:

Repo layout:

opik-mcp/
├── src/opik_mcp/        ← server, tools, analytics
├── tests/               ← pytest suites
├── scripts/             ← live-BE smoke + MCP-session smoke
├── legacy/typescript/   ← migration guide for the deprecated v2 TS server (source: tag `legacy-typescript-final`)
├── pyproject.toml
└── Makefile

Get help

• Open an issue for bugs and feature requests
• Opik docs for SDK / backend documentation
• Comet community Slack for questions

License

Apache-2.0.