Langfuse MCP Server

Usage: 12,518 PyPI downloads last month (pypistats, 2026-08-19). v0.10.1.

Local MCP server and skill for Langfuse. Debug traces, sessions, and exceptions from Claude Code, Codex, Cursor, or any MCP client.

Why this instead of native Langfuse MCP?

Use this for local debug: first-class traces, sessions, and exceptions; route-decision tools; compact / file-dump output; plus the included langfuse skill.

Use official Langfuse MCP for hosted, zero-install access to the broader API (score writes, comments, models, media).

As of June 2026:

	langfuse-mcp	Native Langfuse MCP
Primary fit	Local debug: traces, sessions, exceptions	Hosted, zero-install API surface
Deployment	Local stdio or HTTP, via the Langfuse Python SDK	Native streamable HTTP at /api/public/mcp
Trace / session / exception tools	First-class	Observation/API-oriented access
Route-decision tools	Yes	No
Token & output control	Compact summaries, truncation, file-dump mode, tool-group gating	Hosted tool response + client
Metrics & dataset runs	Yes	Yes
Prompt, dataset, queue & score reads	Yes	Yes
Score writes, comments, models, media	No	Yes

langfuse-mcp for local debug. Native MCP for hosted breadth.

See a failing trace in 2 minutes

Install is uvx langfuse-mcp plus Langfuse API keys — Quick Start for Claude Code, Codex, Cursor, or Docker.

After the client restarts, ask:

find exceptions in the last day

That maps to existing tools:

find_exceptions(age=1440, group_by="file")
find_exceptions_in_file(filepath="<file from the grouping>", age=1440)
get_exception_details(trace_id="<trace_id from the file results>")

find_exceptions returns {group, count, observation_id, trace_id} (top 50 groups) — each group carries a representative observation/trace ID, and find_exceptions_in_file also yields full error records. Then optionally:

fetch_trace(trace_id="<trace_id>", include_observations=true)

The exception tools detect errors by observation level == "ERROR" (strict; a non-empty status_message alone does not count). Counts describe error-level observations, not individual exception events — get_error_count returns exception_count: null for that reason. Exception details come from recorded metadata (exception.type / exception.message / exception.stacktrace, top-level or under metadata.attributes); absent values are null. Results are not a point-in-time snapshot.

Project Links

• Quick Start
• Agent Skill
• Tools
• Selective Tool Loading
• Read-Only Mode
• Other Clients

Quick Start

Requires uv (for uvx) and Python 3.10 or newer. CI verifies Python 3.10 through 3.14.

Get credentials from Langfuse Cloud → Settings → API Keys. If self-hosted, use your instance URL for LANGFUSE_HOST.

# Claude Code (project-scoped, shared via .mcp.json)
claude mcp add \
  -e LANGFUSE_PUBLIC_KEY=pk-... \
  -e LANGFUSE_SECRET_KEY=sk-... \
  -e LANGFUSE_HOST=https://cloud.langfuse.com \
  --scope project \
  langfuse -- uvx langfuse-mcp

# Codex CLI (user-scoped, stored in ~/.codex/config.toml)
codex mcp add langfuse \
  --env LANGFUSE_PUBLIC_KEY=pk-... \
  --env LANGFUSE_SECRET_KEY=sk-... \
  --env LANGFUSE_HOST=https://cloud.langfuse.com \
  -- uvx langfuse-mcp

To pin a CI-verified interpreter explicitly, add --python 3.14 before langfuse-mcp.

Restart your CLI, then verify with /mcp (Claude Code) or codex mcp list (Codex).

Agent Skill

This repo ships a first-party langfuse skill for Claude Code and Codex. The skill gives agents concrete playbooks for trace debugging, exception triage, latency analysis, prompt management, and dataset work.

Install it when you want the agent to know when to reach for Langfuse and which MCP tools to call first.

Via skills (recommended):

npx skills add avivsinai/langfuse-mcp -g -y

Via skild:

npx skild install @avivsinai/langfuse -t claude -y

Manual install:

cp -r skills/langfuse ~/.claude/skills/   # Claude Code
cp -r skills/langfuse ~/.codex/skills/    # Codex CLI

After installing the skill, try:

help me debug langfuse traces
find exceptions in the last day
why was this user's session slow?

The MCP server provides the tools; the skill provides the agent-facing workflow. See skills/langfuse/SKILL.md, skills/langfuse/references/setup.md, and skills/langfuse/references/tool-reference.md.

Tools (48 total)

Category	Tools
Traces	fetch_traces, fetch_trace
Observations	fetch_observations, fetch_observation
Routing	find_route_decisions, get_route_decision, summarize_route_decisions, find_low_confidence_route_decisions
Sessions	fetch_sessions, get_session_details, get_user_sessions
Exceptions	find_exceptions, find_exceptions_in_file, get_exception_details, get_error_count
Prompts	list_prompts, get_prompt, get_prompt_unresolved, create_text_prompt, create_chat_prompt, update_prompt_labels
Datasets	list_datasets, get_dataset, list_dataset_items, get_dataset_item, create_dataset, create_dataset_item, delete_dataset_item, list_dataset_runs, get_dataset_run, list_dataset_run_items, create_dataset_run_item, delete_dataset_run
Annotation Queues	list_annotation_queues, create_annotation_queue, get_annotation_queue, list_annotation_queue_items, get_annotation_queue_item, create_annotation_queue_item, update_annotation_queue_item, delete_annotation_queue_item, create_annotation_queue_assignment, delete_annotation_queue_assignment
Scores	list_scores_v2, get_score_v2
Metrics	query_metrics, get_metrics_schema
Schema	get_data_schema

Dataset Item Updates (Upsert)

Langfuse uses upsert for dataset items. To edit an existing item, call create_dataset_item with item_id. If the ID exists, it updates; otherwise it creates a new item.

create_dataset_item(dataset_name="qa-test-cases", item_id="item_123", input={"question": "What is 2+2?"}, expected_output={"answer": "4"})

Metrics Queries

query_metrics aggregates telemetry server-side (cost, latency, tokens, counts, score values) so agents can answer "what did inference cost?" or "what's p95 latency by model?" without pulling raw traces. Call get_metrics_schema for the full view/dimension/measure catalog.

query_metrics(
    view="observations",
    metrics=[{"measure": "totalCost", "aggregation": "sum"}, {"measure": "latency", "aggregation": "p95"}],
    dimensions=["providedModelName"],
    age=1440,  # last 24h; or pass from_timestamp / to_timestamp
)

High-cardinality fields (id, traceId, userId, sessionId) must be used in filters, not dimensions. The v2 metrics endpoint is Langfuse Cloud-only; self-hosted instances may return 404.

Selective Tool Loading

Load only the tool groups you need to reduce token overhead:

langfuse-mcp --tools traces,prompts

Available groups: traces, observations, routing, sessions, exceptions, prompts, datasets, annotation_queues, scores, metrics, schema

The routing group is router-neutral. It reads Langfuse span observations with metadata.schema_version: "mcp.route_decision.v1" and filters on route-decision fields stored in observation metadata, such as decision_id, router_name, provider, and capability_id.

Read-Only Mode

Disable all write operations for safer read-only access:

langfuse-mcp --read-only
# Or via environment variable
LANGFUSE_MCP_READ_ONLY=true langfuse-mcp

This disables: create_text_prompt, create_chat_prompt, update_prompt_labels, create_dataset, create_dataset_item, delete_dataset_item, create_dataset_run_item, delete_dataset_run, create_annotation_queue, create_annotation_queue_item, update_annotation_queue_item, delete_annotation_queue_item, create_annotation_queue_assignment, delete_annotation_queue_assignment

Default Output Mode

Set the MCP-exposed default output_mode so clients that omit the parameter automatically use your preferred mode:

langfuse-mcp --default-output-mode full_json_file
# Or via environment variable
LANGFUSE_MCP_DEFAULT_OUTPUT_MODE=full_json_file langfuse-mcp

Supported values: compact, full_json_string, full_json_file

This updates the default shown in MCP tool schemas. Clients can still override it per call by passing output_mode explicitly.

Other Clients

Cursor

Create .cursor/mcp.json in your project (or ~/.cursor/mcp.json for global):

{
  "mcpServers": {
    "langfuse": {
      "command": "uvx",
      "args": ["langfuse-mcp"],
      "env": {
        "LANGFUSE_PUBLIC_KEY": "pk-...",
        "LANGFUSE_SECRET_KEY": "sk-...",
        "LANGFUSE_HOST": "https://cloud.langfuse.com",
        "LANGFUSE_MCP_DEFAULT_OUTPUT_MODE": "full_json_file"
      }
    }
  }
}

Docker (single project)

docker run --rm -i \
  -e LANGFUSE_PUBLIC_KEY=pk-... \
  -e LANGFUSE_SECRET_KEY=sk-... \
  -e LANGFUSE_HOST=https://cloud.langfuse.com \
  ghcr.io/avivsinai/langfuse-mcp:latest

HTTP transport — shared server for multiple projects

Run one persistent server instance and route each MCP client to its own Langfuse project by passing credentials in the Authorization header.

# Start a shared server (binds to localhost by default)
docker run -d -p 127.0.0.1:8000:8000 \
  -e LANGFUSE_HOST=https://cloud.langfuse.com \
  ghcr.io/avivsinai/langfuse-mcp:latest \
  --transport streamable-http --bind-host 0.0.0.0

Security note: --bind-host 0.0.0.0 exposes the port on all interfaces. In production, place the server behind a TLS-terminating reverse proxy (nginx, Caddy, Cloudflare Tunnel) that enforces HTTPS. The Authorization header containing your keys is transmitted in plaintext over plain HTTP. If startup credentials are set, the proxy must enforce authentication; otherwise unauthenticated callers without an Authorization header can use the default project. For shared public HTTP deployments, omit default LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY credentials unless the fronting proxy authenticates every request.

Register each project separately in your MCP client, passing its credentials as a Basic auth header (base64(public_key:secret_key)):

# Generate the header value for each project:
echo -n "pk-lf-YOURKEY:sk-lf-YOURSECRET" | base64
# cGstbGYtWU9VUktFWTpzay1sZi1ZT1VSU0VDUkVU

# Register in Claude Code (one entry per project):
claude mcp add langfuse-audit \
  --transport http http://localhost:8000/mcp \
  -H "Authorization: Basic cGstbGYtWU9VUktFWTpzay1sZi1ZT1VSU0VDUkVU"

claude mcp add langfuse-staging \
  --transport http http://localhost:8000/mcp \
  -H "Authorization: Basic <base64 for staging project>"

Auth semantics: Basic here carries Langfuse API keys, not user passwords. An absent header falls back to startup env credentials (LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY). Any malformed header is rejected outright — there is no silent fallback to a different project.

Optional environment variables

Variable	Default	Description
LANGFUSE_MAX_AGE_DAYS	7	Caps the lookback window for time-based tools (fetch_traces, fetch_observations, etc.). Set to match your Langfuse instance's data retention — e.g. 30 if your retention is 30 days.
LANGFUSE_MCP_TRACE_TIMEOUT_SECONDS	120	Per-request read timeout (seconds) for single-trace fetches (fetch_trace). Raise it if large traces with include_observations=True time out. Must be a positive integer.

Development

uv venv --python 3.14 .venv && source .venv/bin/activate
uv pip install -e ".[dev]"
pytest

License

MIT