Connect Langfuse to Cursor via MCP: Traces and Prompts

💡 No install? Use cloud.anythingmcp.com directly. Sign in, click Connectors → Langfuse, paste your credentials, mint an MCP API key and you're done. No Docker, no git clone, no local server.

Langfuse + Cursor

Read a Langfuse project with your AI: LLM traces, observations and sessions with cost, tokens and latency, metrics over time, scores and evaluations, prompts with their versions and labels, datasets, dataset items and experiments. It can also publish prompt versions, move labels, add scores and build datasets.

Prerequisites

See the full setup instructions baked into the connector (visible in the in-app store when you select the connector). The required environment variables for this connector are:

LANGFUSE_HOST, LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY

Step 1 — Get credentials

Setting up takes about 3 minutes. You need a Langfuse project on Langfuse Cloud (EU, US, Japan or HIPAA) or on a self-hosted Langfuse v4.

1. In Langfuse open the project, then Settings, API Keys, Create new API keys. Use a project key, not an organization key.
2. Copy the public key (pk-lf-...) into LANGFUSE_PUBLIC_KEY and the secret key (sk-lf-..., shown only once) into LANGFUSE_SECRET_KEY. The connector sends them as HTTP Basic auth.
3. Set LANGFUSE_HOST to the region of the project, without a trailing slash: https://cloud.langfuse.com (EU), https://us.cloud.langfuse.com (US), https://jp.cloud.langfuse.com (Japan), https://hipaa.cloud.langfuse.com (HIPAA) or your self-hosted address. Keys of one region do not work on another.
4. Install the connector and run langfuse_get_project to check the keys; langfuse_get_health shows the server version.

Costs: no tool runs a model. The connector only reads and writes Langfuse data; the costs it shows are the ones Langfuse recorded for your applications' LLM calls.

Tracing data: in Langfuse v4 a trace is the set of observations (spans, generations, tool calls, events) that share a traceId, and a session is the set of traces that share a sessionId. The connector reads them from Observations API v2 and Metrics API v2, the real-time read path; the old /traces, /sessions and /metrics/daily endpoints are deprecated (Langfuse Cloud serves them until 16 November 2026). Other endpoints can lag about 10 minutes behind ingestion.

• langfuse_list_observations lists rows newest first, always bounded by from_start_time and to_start_time. Filters: type (GENERATION for LLM calls), name, user_id, session_id, trace_id, level (ERROR for failures), environment, is_root_observation (true gives one row per trace) or a structured filter. fields picks the field groups core, basic, time, io (input and output), metadata, model, usage (tokens and cost), prompt, metrics (latency) and trace_context; the default is core,basic,usage.
• langfuse_get_trace returns every observation of one trace with input, output, model, usage and latency; the root has no parentObservationId and carries the trace's input and output.
• langfuse_get_session returns one row per trace of a session by default; root_only: false gives every observation.
• langfuse_query_metrics aggregates. query is a JSON string with view (observations, scores-numeric, scores-boolean, scores-categorical), metrics (measures such as count, latency, totalTokens, totalCost or timeToFirstToken, with the aggregations sum, avg, count, max, min, p50, p75, p90, p95, p99), dimensions (such as providedModelName, traceName, promptName, promptVersion, environment, release), timeDimension (per minute, hour, day, week or month), filters, fromTimestamp and toTimestamp. High-cardinality fields such as userId and sessionId can be filtered on but not grouped by.

Scores: langfuse_list_scores filters by name, source (API, ANNOTATION, EVAL), data_type, trace, session, observation, experiment, value range and time range; langfuse_list_score_configs shows the score definitions. langfuse_create_score adds one score (numeric, boolean as 0 or 1, categorical or text) to a trace, observation, session or experiment run. It is processed asynchronously (HTTP 202), and sending the same score_id again overwrites it.

Prompts: langfuse_list_prompts lists names with versions, labels and tags; langfuse_get_prompt returns the version labelled production unless version or label is given. langfuse_create_prompt_version publishes a new version of a text prompt (with {{variables}}) or a chat prompt, and a new name creates the prompt; langfuse_update_prompt_labels moves labels such as production or staging to a version. Moving production changes what your applications fetch, so the AI asks you first. Folder names such as support/greeting are passed as they are.

Datasets and experiments: langfuse_list_datasets, langfuse_get_dataset and langfuse_create_dataset (optionally with JSON schemas for input and expected output); langfuse_list_dataset_items, langfuse_get_dataset_item and langfuse_upsert_dataset_item (creates an item, or updates it when item_id exists; status ARCHIVED hides it from new runs). langfuse_list_experiments lists dataset runs and experiments with their scores; from_start_time is required.

Langfuse's own MCP server: Langfuse also runs a native MCP server at <LANGFUSE_HOST>/api/public/mcp with the same key pair, mainly for prompt management. This connector covers prompts too, plus tracing data, metrics, scores and datasets.

Switched off at install: nothing. All 20 tools are on; the write tools add prompt versions, labels, scores, datasets and dataset items, and none of them deletes anything.

Data formats: times are ISO 8601, UTC. Costs are in USD; in v2 observations inputPrice, outputPrice and totalPrice are strings for decimal precision, and input and output arrive as JSON text. Latency and time to first token in metrics are milliseconds.

Paging: observations, traces, sessions, scores and experiments use cursors: the AI passes meta.cursor back as cursor until it is missing (observations: 50 rows by default, up to 1,000). Prompts, datasets, dataset items and score configs use page (from 1) and limit (default 50, at most 100); meta.totalPages says how many pages exist.

Limits: self-hosted Langfuse must be v4 or later for the observation and metrics tools. Data outside Observations and Metrics API v2 can lag about 10 minutes behind.

Common errors: 400 means a bad parameter (a malformed filter or query JSON, a group-by on a high-cardinality field, a limit over the maximum); 401 a wrong key pair or the wrong region; 403 the key may not use this endpoint (use a project key); 404 an unknown prompt, dataset or item, or a self-hosted version without the v2 endpoints; 429 rate limiting (the AI waits instead of retrying in a loop).

The connector does not delete traces, prompts, datasets or scores and does not send traces (the Langfuse SDKs do that over OpenTelemetry). Annotation queues, LLM connections and organization or SCIM administration are not included.

Step 2 — Install the adapter

curl -fsSL https://raw.githubusercontent.com/HelpCode-ai/anythingmcp/main/docker-compose.quickstart.yml -o docker-compose.yml
printf 'JWT_SECRET=%s\nENCRYPTION_KEY=%s\n' "$(openssl rand -hex 32)" "$(openssl rand -hex 32)" > .env
docker compose up -d

Step 3 — Add the connector in Cursor

Your server URL: in AnythingMCP, open MCP Servers → the server this connector is on, and copy its URL (https://cloud.anythingmcp.com/mcp/…). Use it wherever this guide shows YOUR_SERVER_ID.

Cursor reads MCP servers from ~/.cursor/mcp.json. Add this entry:

{
  "mcpServers": {
    "anythingmcp": {
      "url": "https://cloud.anythingmcp.com/mcp/YOUR_SERVER_ID",
      "headers": { "X-API-Key": "YOUR_MCP_API_KEY" }
    }
  }
}

1. Create an MCP API key in AnythingMCP under MCP Servers → your server → API keys, and paste it in place of YOUR_MCP_API_KEY.
2. Save the file and restart Cursor.
3. Open Cursor → Settings → MCP and check that anythingmcp is listed and connected.
4. Start chatting: the agent can now call the Langfuse tools.

Available tools

Example prompts

• "What did our LLM calls cost per model per day over the last 7 days?"
• "Show me yesterday's generations with level ERROR, with model, input and output."
• "Show me the whole conversation of session chat-8f21 with the input and output of each turn."
• "What is the p95 latency per trace name this week, and which one got slower?"
• "Publish a new version of the prompt support/greeting with a friendlier tone, but do not move the production label yet."

FAQ

Does Cursor support custom MCP servers on the free tier? Yes, MCP is available on Hobby, Pro and Business plans.

Next steps

• Connector landing page
• Connect to Claude
• Connect to ChatGPT
• Connect to Gemini
• Connect to Meta Muse
• Connect to Microsoft Copilot