The README's comparison table is the argument: traces, sessions, exceptions, datasets, scores, annotation queues, comments and model definitions are all absent from the official Langfuse MCP, and prompt access there is read-only. That makes this the option when the loop you want is find the failing traces, add them to a review queue, then ship a new prompt version — all without a browser. Pagination is one-based, limits default to 20, and several tools delete permanently, so read the annotations before auto-approving anything.
A Spring Boot MCP server that puts the Langfuse public REST API in front of an agent over Streamable HTTP. It goes well past reading traces: prompts can be created and relabelled, datasets and experiment runs managed, score configs defined, annotation queues filled for human review, and custom cost models registered.
- Trace work, including the error path: `fetch_traces` with filters on user, session, tags and time; `find_exceptions` for ERROR-level traces; `find_exceptions_in_file` for errors whose metadata names a given file; `get_error_count` over a window
- Session and user views — `fetch_sessions`, `get_session_details` with all its traces, `get_user_sessions`, `get_user_traces`
- Prompt management in both directions: `list_prompts`, `get_prompt` pinned to a version or a label such as production, `create_prompt`, `delete_prompt`, `update_prompt_labels`
- Evaluation plumbing — datasets and items, dataset runs and run items, plus scores and score configs including numeric ranges and categorical label sets
- `create_annotation_queue` and its item tools, so a trace, observation or session can be pushed to a human reviewer from the same conversation
- `get_cost_metrics`, which takes a JSON query against the Metrics API and aggregates cost, tokens, latency and usage server-side
- `get_data_schema`, which returns the full Langfuse data model and is the one to call first
- Destructive counterparts exist and are marked as irreversible: `delete_trace`, `delete_traces`, `delete_dataset_item`, `delete_dataset_run`, `delete_model`
A Langfuse account and an API key pair: `LANGFUSE_PUBLIC_KEY`, `LANGFUSE_SECRET_KEY` and `LANGFUSE_HOST` are all required, and `LANGFUSE_TIMEOUT` defaults to 30s. Java 21 or later plus Maven 3.9+, or the multi-stage Dockerfile which compiles inside Docker. Build produces `langfuse-mcp-1.0.0.jar`; the server listens on port 8080 and the MCP endpoint is `/mcp` — the legacy SSE path is not used. Self-hosted Langfuse works by pointing the host at your instance. MIT licensed.