What is Langfuse MCP Server?
Langfuse, the open-source LLM observability platform, has its MCP server built into the product. You do not install anything: each Langfuse Cloud region (US at https://us.cloud.langfuse.com/api/public/mcp, plus EU, Japan and HIPAA US) and every self-hosted instance serves it at /api/public/mcp over Streamable HTTP. It authenticates with a project-scoped public and secret key pair sent as a Basic Authorization header. It replaces the older langfuse/mcp-server-langfuse repo, a prompts-only server with no commits since February 2025.
The MCP Reference lists about 95 tools across prompts, skills, observations, annotation queues, comments, datasets and experiments, scores, metrics, models, evaluators and evaluation rules, dashboards and alerts. Read and write tools are both on by default. A separate unauthenticated server at https://langfuse.com/api/mcp searches the documentation. The platform itself is reviewed under repos/langfuse.
How to install Langfuse MCP Server
https://us.cloud.langfuse.com/api/public/mcp (also EU, Japan, HIPAA US and self-hosted /api/public/mcp)Is Langfuse MCP Server worth it?
This is a vendor MCP server done the simple way: it lives inside Langfuse, so it is always on the same version as your project and works identically on Cloud and self-hosted. Setup is one claude mcp add command with a base64-encoded key pair. The scope goes well beyond the old prompts-only server. An agent can list and filter observations, read scores, run batch evaluations, create dataset items from production traces, update prompt labels and build dashboard widgets.
Rate limits are what readers need to plan around. MCP calls count against Langfuse's General API bucket, which is shared by every project and key in the organisation. It allows 30 requests a minute on Hobby, 100 on Core and 1,000 on Pro and above. An agent that pages through observations one getObservation at a time can use up a Hobby minute on one question and then wait out a 429.
Writes need care. Roughly half the tools change state, including deleteDatasetRun, deleteEvaluationRule, deleteDashboard and updatePromptLabels, which can move the production label to a different prompt version. Langfuse's advice is to allowlist read-only tools in the client if that is all you want. The key in the header is a full project secret key, stored in plain text in the client's config file.
Traces contain whatever your users typed, so an agent reading them is reading untrusted text with prompt-injection risk. Langfuse itself recommends its Agent Skill and CLI over MCP wherever the agent can run shell commands, and says tool schemas may change without notice. That is a fair warning for anyone pinning workflows to tool names.
Where it earns its place
An AI engineer on a support-bot team using Langfuse Cloud US
Find the conversations from last night where the bot scored badly and see why
listObservations with filters and listScores pull the failing traces into Claude Code, which summarises the common cause. On Core's 100 requests a minute this is comfortable; on Hobby, large pulls hit 429s.
A prompt engineer maintaining a production system prompt
Promote a tested prompt version to production without opening the web UI
getPrompt fetches the candidate, and updatePromptLabels moves the production label after review. Prompt GET calls have no rate limit, but the label change takes effect for live traffic at once, so keep it behind an approval prompt.
An evaluation lead at a 40-person AI startup on Langfuse Pro
Turn real failure cases into a regression dataset and run an evaluator over it
The agent uses batchUpsertDatasetItems from selected observations, then runBatchEvaluation. Platform cost is the Pro plan at $199 a month plus $8 per 100k units beyond the included 100k; evaluator LLM calls are billed by your model provider.
Before you commit
- Setup
- Create a project API key, base64-encode pk-lf-...:sk-lf-..., and register the endpoint with an Authorization: Basic header. Langfuse documents Claude Code, Codex, Cursor and Pi; it is a few minutes to a first listPrompts call.
- Integrations
- About 95 tools in the MCP Reference covering prompts, skills, observations, annotation queues, comments, datasets, experiments, scores, metrics, models, LLM connections, evaluators, evaluation rules, dashboards and alerts. Roughly half write or delete.
- Data and privacy
- Hosted by Langfuse on Cloud, so it reads your trace data where it already lives; self-hosting keeps it on your infrastructure. One project per key. Trace contents are untrusted input for prompt injection.
- Cost at scale
- No separate fee. MCP shares the General API bucket at 30, 100 or 1,000 requests a minute on Hobby, Core and Pro; Enterprise ($2,499/month) can request custom limits. Usage beyond plan allowances is $8 per 100k units.
- Limits
- Rate limits are per organisation and fixed-window; 429 responses carry Retry-After. Schemas are self-describing and may change. Self-hosted instances behind a reverse proxy need the Host header preserved or LANGFUSE_MCP_ALLOWED_HOSTS set, or they return 403.
What’s good
- Hosted in every Langfuse region and self-hosted instance; nothing to run
- About 95 tools across prompts, traces, datasets, evals and dashboards
- Project-scoped keys limit reach to one project
- Docs MCP server for documentation search needs no key
What’s not
- 30 requests a minute on Hobby, shared with all other API use
- Write and delete tools enabled by default
- Full secret key sits in client config as Basic auth
- Tool schemas may change without notice
Best for
Teams already on Langfuse who want a coding agent to inspect traces, curate datasets and manage prompts from the editor.
Langfuse MCP Server alternatives
Langfuse MCP Server: common questions
- Is Langfuse MCP Server worth it?
- Nothing to install and close to the whole Langfuse API, but on the free plan's 30 requests a minute an agent digging through traces runs out of road quickly. We rate it 4.2 out of 5.
- How much does Langfuse MCP Server cost?
- Included with Langfuse. Cloud: Hobby free (50k units/month); Core $29/month; Pro $199/month; Enterprise $2,499/month; extra usage $8 per 100k units. Self-hosted is free.
- What are the downsides of Langfuse MCP Server?
- 30 requests a minute on Hobby, shared with all other API use; write and delete tools enabled by default; full secret key sits in client config as Basic auth; tool schemas may change without notice.
- Who is Langfuse MCP Server best for?
- Teams already on Langfuse who want a coding agent to inspect traces, curate datasets and manage prompts from the editor.
- How much does Langfuse MCP Server cost for a team or in production?
- No separate fee. MCP shares the General API bucket at 30, 100 or 1,000 requests a minute on Hobby, Core and Pro; Enterprise ($2,499/month) can request custom limits. Usage beyond plan allowances is $8 per 100k units.
- Is Langfuse MCP Server safe for company data?
- Hosted by Langfuse on Cloud, so it reads your trace data where it already lives; self-hosting keeps it on your infrastructure. One project per key. Trace contents are untrusted input for prompt injection.
- What are the best Langfuse MCP Server alternatives?
- The closest alternatives we have reviewed are Sentry MCP and Grafana MCP Server.
Answers checked Oct 31, 2026.
Make Langfuse MCP Server? Show this rating on your site
