Langfuse Skill

Debug AI agents and LLM applications through Langfuse observability.

This skill is the agent-facing companion to langfuse-mcp. It tells Claude Code and Codex when to use Langfuse, which MCP tool to call first, and how to move from broad trace discovery to a concrete root-cause hypothesis.

Triggers: langfuse, traces, debug AI, find exceptions, set up langfuse, what went wrong, why is it slow, datasets, evaluation sets

What This Skill Provides

• Setup steps for connecting langfuse-mcp to Claude Code or Codex.
• Playbooks for exception triage, trace inspection, latency analysis, sessions, prompts, and datasets.
• A quick reference for the highest-value MCP tools.
• Links to full setup and tool references for deeper troubleshooting.

Use the playbooks before guessing at individual tools. Start broad, identify the relevant trace/session/observation, then drill into the exact failure or slow path.

Setup

Step 1: Get credentials from https://cloud.langfuse.com → Settings → API Keys

If self-hosted, use your instance URL for LANGFUSE_HOST and create keys there.

Step 2: Install MCP (pick one):

Requires Python 3.10 or newer. CI verifies Python 3.10 through 3.14.

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

Add --python 3.14 before langfuse-mcp if you want to pin a CI-verified interpreter explicitly.

Step 3: Restart CLI, verify with /mcp (Claude) or codex mcp list (Codex)

Step 4: Test: fetch_traces(age=60)

Read-Only Mode

For safer observability without risk of modifying prompts or datasets, enable read-only mode:

# CLI flag
langfuse-mcp --read-only

# Or environment variable
LANGFUSE_MCP_READ_ONLY=true

This disables write tools: create_text_prompt, create_chat_prompt, update_prompt_labels, create_dataset, create_dataset_item, delete_dataset_item.

Default Output Mode

If you want MCP clients to default to writing full payloads to files when they omit output_mode, configure:

langfuse-mcp --default-output-mode full_json_file

# Or via environment variable
LANGFUSE_MCP_DEFAULT_OUTPUT_MODE=full_json_file

For manual .mcp.json setup or troubleshooting, see references/setup.md.

Playbooks

"Where are the errors?"

find_exceptions(age=1440, group_by="file")

→ Shows error counts by file. Pick the worst offender.

find_exceptions_in_file(filepath="src/ai/chat.py", age=1440)

→ Lists specific exceptions. Grab a trace_id.

get_exception_details(trace_id="...")

→ Full stacktrace and context.

"What happened in this interaction?"

fetch_traces(age=60, user_id="...")

→ Find the trace. Note the trace_id.

If you don't know the user_id, start with:

fetch_traces(age=60)

fetch_trace(trace_id="...", include_observations=true)

→ See all LLM calls in the trace.

fetch_observation(observation_id="...")

→ Inspect a specific generation's input/output.

"Why is it slow?"

fetch_observations(age=60, type="GENERATION")

→ Find recent LLM calls. Look for high latency.

fetch_observation(observation_id="...")

→ Check token counts, model, timing.

"What's this user experiencing?"

get_user_sessions(user_id="...", age=1440)

→ List their sessions.

get_session_details(session_id="...")

→ See all traces in the session.

"Manage datasets"

list_datasets()

→ See all datasets.

get_dataset(name="evaluation-set-v1")

→ Get dataset details.

list_dataset_items(dataset_name="evaluation-set-v1", page=1, limit=10)

→ Browse items in the dataset.

create_dataset(name="qa-test-cases", description="QA evaluation set")

→ Create a new dataset.

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

→ Add test cases.

create_dataset_item(
  dataset_name="qa-test-cases",
  item_id="item_123",
  input={"question": "What is 3+3?"},
  expected_output={"answer": "6"}
)

→ Upsert: updates existing item by id or creates if missing.

"Manage prompts"

list_prompts()

→ See all prompts with labels.

get_prompt(name="...", label="production")

→ Fetch current production version.

create_text_prompt(name="...", prompt="...", labels=["staging"])

→ Create new version in staging.

update_prompt_labels(name="...", version=N, labels=["production"])

→ Promote to production. (Rollback = re-apply label to older version)

Quick Reference

age = minutes to look back (max 10080 = 7 days)

Troubleshooting

MCP connection fails

• Verify credentials: check LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_HOST
• Restart CLI after adding/updating MCP config
• Test MCP independently: fetch_traces(age=60) — if this fails, the issue is MCP, not the skill
• See references/setup.md for detailed troubleshooting

No traces found

• Increase the age parameter (default lookback may be too short)
• Verify your application is sending traces to the correct Langfuse project
• Check LANGFUSE_HOST points to the right instance (cloud vs self-hosted)

Permission denied

• Regenerate API keys from Langfuse dashboard
• Ensure keys have the required scopes for the operation
• Write operations require read-write keys (not read-only mode)

References

• references/tool-reference.md — Full parameter docs, filter semantics, response schemas
• references/setup.md — Manual setup, troubleshooting, advanced configuration