/langfuse
Investigate AI traces, observations, exceptions, latency, sessions, prompts, datasets, annotation queues, and scores through Langfuse MCP. Use when the request names Langfuse or asks to diagnose recorded AI behavior.
$ npx -y skills add avivsinai/langfuse-mcp --skill langfuse --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
- Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
- You can call itInvoke it directly when you want it.
- Slash command
/langfuse
Context preview
The summary Claude sees to decide when to auto-load this skill.
Investigate AI traces, observations, exceptions, latency, sessions, prompts, datasets, annotation queues, and scores through Langfuse MCP. Use when the request names Langfuse or asks to diagnose recorded AI behavior.
SKILL.md
langfuse.SKILL.mdname: langfuse
version: 0.14.0
description: Investigate AI traces, observations, exceptions, latency, sessions, prompts, datasets, annotation queues, and scores through Langfuse MCP. Use when the request names Langfuse or asks to diagnose recorded AI behavior.
metadata:
short-description: Langfuse observability via MCP
compatibility: claude-code, codex-cli
Langfuse
Use Langfuse evidence to move from a broad symptom to the relevant trace, observation, session, prompt, or dataset. Start with the narrowest read-only query that can locate the event, then inspect its concrete inputs, outputs, timing, and errors.
Route the task
- For trace, observation, exception, latency, or session diagnosis, read
[diagnostic and management workflows](references/workflows.md).
- For installation, credentials, tool groups, output defaults, connection
failures, or empty results, read [setup](references/setup.md).
- For exact parameters, filters, pagination, output modes, and response shapes,
read [the tool reference](references/tool-reference.md).
A common investigation starts with `fetch_traces(age=60)` or `find_exceptions(age=1440, group_by="file")`, then fetches the selected trace or observation by ID. Use `full_json_file` only when the complete payload is needed; exports can contain sensitive user data.
Evidence and mutation boundaries
- Treat model names, prompts, timestamps, and other values in historical traces
as recorded evidence. Do not rewrite or normalize old model identifiers to a current name such as `gpt-6-astra`.
- Trace discovery and inspection are read-only. Prompt creation or relabeling,
dataset changes, annotation-queue changes, and deletions mutate Langfuse; perform them only when the user requested that change.
- Prefer read-only server mode for diagnostic work. It disables prompt and
dataset write tools:
langfuse-mcp --read-only
# or LANGFUSE_MCP_READ_ONLY=true
- Never commit Langfuse credentials. Keep public key, secret key, and host in
the supported environment or MCP configuration described in the setup reference. Rotate any exposed key.
- Preserve exact IDs and filters in findings so another reader can reproduce
the query. State when pagination, lookback limits, or compact output limits the conclusion.
Read more
name: langfuse version: 0.14.0 description: Investigate AI traces, observations, exceptions, latency, sessions, prompts, datasets, annotation queues, and scores through Langfuse MCP. Use when the request names Langfuse or asks to diagnose recorded AI behavior. metadata: short-description: Langfuse observability via MCP compatibility: claude-code, codex-cli
Langfuse
Use Langfuse evidence to move from a broad symptom to the relevant trace, observation, session, prompt, or dataset. Start with the narrowest read-only query that can locate the event, then inspect its concrete inputs, outputs, timing, and errors.
Route the task
- For trace, observation, exception, latency, or session diagnosis, read
[diagnostic and management workflows](references/workflows.md).
- For installation, credentials, tool groups, output defaults, connection
failures, or empty results, read [setup](references/setup.md).
- For exact parameters, filters, pagination, output modes, and response shapes,
read [the tool reference](references/tool-reference.md).
A common investigation starts with `fetch_traces(age=60)` or `find_exceptions(age=1440, group_by="file")`, then fetches the selected trace or observation by ID. Use `full_json_file` only when the complete payload is needed; exports can contain sensitive user data.
Evidence and mutation boundaries
- Treat model names, prompts, timestamps, and other values in historical traces
as recorded evidence. Do not rewrite or normalize old model identifiers to a current name such as `gpt-6-astra`.
- Trace discovery and inspection are read-only. Prompt creation or relabeling,
dataset changes, annotation-queue changes, and deletions mutate Langfuse; perform them only when the user requested that change.
- Prefer read-only server mode for diagnostic work. It disables prompt and
dataset write tools:
langfuse-mcp --read-only # or LANGFUSE_MCP_READ_ONLY=true
- Never commit Langfuse credentials. Keep public key, secret key, and host in
the supported environment or MCP configuration described in the setup reference. Rotate any exposed key.
- Preserve exact IDs and filters in findings so another reader can reproduce
the query. State when pagination, lookback limits, or compact output limits the conclusion.
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.