Langfuse

Langfuse

Langfuse is the team's default LLM observability and prompt-management platform. This skill is the single front door for every Langfuse interaction — instrumentation, debugging traces, managing prompts, running evals, capturing user feedback.

Three mechanisms, one front door

Langfuse is reachable three ways. They are complementary, not redundant — pick the right one per intent:

Default routing:

• Reading or mutating a prompt → MCP first (cheaper). Fall back to CLI only if MCP isn't registered or the operation isn't covered.
• Reading or mutating anything else (traces, scores, datasets, sessions, evals, metrics) → CLI.
• Anything that's not data access (instrumentation, migration design, debugging methodology) → load the relevant references/*.md file.

Core principles

Apply these to every Langfuse task:

1. Documentation first — Langfuse evolves frequently. Never implement from memory. Fetch current docs before writing code (see "Documentation access" below).
2. CLI for arbitrary data access; MCP for prompts — see routing table above.
3. Use latest SDK versions — unless the user pins a version or there's a stated reason. See references/sdk-upgrade.md.
4. Best practices by use case — load the matching reference before implementing.

Preflight

Before any Langfuse operation, verify the surface you intend to use:

Credentials (required for CLI and MCP)

echo "${LANGFUSE_PUBLIC_KEY:?missing}"   # pk-lf-...
echo "${LANGFUSE_SECRET_KEY:?missing}"   # sk-lf-...
echo "${LANGFUSE_HOST:?missing}"         # https://cloud.langfuse.com (EU), https://us.cloud.langfuse.com (US), or self-hosted URL

If unset, ask the user for their project keys (Langfuse UI → Settings → API Keys) and the host they're using. Never hardcode keys in committed files — use ${VAR} interpolation.

In repos that use OpenBao (this one does — see docs/openbao-secret-management.md), one line populates all four LANGFUSE_* env vars (including the computed LANGFUSE_BASIC_AUTH):

eval "$("<skill-base-dir>/../bao-vault/scripts/langfuse_env.sh")"

Safe to run unconditionally — falls back silently when OpenBao is unavailable.

CLI reachable

npx langfuse-cli --version    # should print a version string

npx will fetch the package on first run; no global install required. If the user prefers a global install: npm i -g langfuse-cli.

MCP server registered (optional but recommended for prompt work)

Verify by attempting a read-only call:

mcp__langfuse__listPrompts({})

If it errors with "tool not found" or "MCP server not configured", the server isn't registered. See references/mcp-setup.md for full coverage, or run bash "<skill-base-dir>/scripts/install-mcp.sh" for an idempotent setup that targets Claude Code (.mcp.json), Codex (~/.codex/config.toml), and Gemini (~/.gemini/settings.json) in one shot.

Resource model

Shared vocabulary used across CLI args, MCP tools, and the UI:

• Project — top-level container; all keys are project-scoped.
• Session — a logical conversation grouping multiple traces (e.g. one Claude Code session).
• Trace — a single request/response cycle. Has nested observations.
• Observation — a unit of work inside a trace. Subtypes: span (generic), generation (LLM call, has model/usage), event (point-in-time).
• Score — a quality signal attached to a trace or observation. Numeric, categorical, or boolean. Can be model-graded, human-graded, or programmatic.
• Prompt — versioned text or chat template, addressed by name + version (or name + label, e.g. production).
• Dataset — a labeled set of inputs + expected outputs, used for evals.
• Dataset Run — one execution of an experiment against a dataset, producing scored items.

Quick operations

Common enough to handle inline without loading a reference.

Prompts (prefer MCP)

mcp__langfuse__listPrompts({})
mcp__langfuse__getPrompt({"name": "chat-assistant", "version": 3})
mcp__langfuse__getPrompt({"name": "chat-assistant", "label": "production"})
mcp__langfuse__createTextPrompt({"name": "chat-assistant", "prompt": "...", "labels": ["production"]})
mcp__langfuse__updatePromptLabels({"name": "chat-assistant", "version": 4, "newLabels": ["production"]})

CLI fallback when MCP isn't available:

npx langfuse-cli api prompts list --json
npx langfuse-cli api prompts get --name chat-assistant --version 3 --json

Traces, observations, scores (CLI)

npx langfuse-cli api traces list --limit 20 --json
npx langfuse-cli api traces get --trace-id <id> --json
npx langfuse-cli api observations-v2s list --trace-id <id> --json   # prefer v2
npx langfuse-cli api score-v2s list --trace-id <id> --json          # prefer v2
npx langfuse-cli api sessions list --limit 20 --json
npx langfuse-cli api metrics-v2s list --json                        # prefer v2

Discovery

npx langfuse-cli api __schema                          # all resources
npx langfuse-cli api <resource> --help                 # actions on a resource
npx langfuse-cli api <resource> <action> --help        # args for an action
npx langfuse-cli api <resource> <action> --curl        # preview HTTP without sending

Routing — load the right reference

For anything beyond quick operations, load at most one or two of these:

Composition patterns

Multi-step Langfuse workflows chain naturally:

• Ship a new prompt → MCP createTextPrompt → reference it in code → CLI traces list to verify it ran → CLI scores create to grade it.
• Migrate a hardcoded prompt → CLI to inventory traces using the literal string → MCP createTextPrompt for the externalized version → refactor code to call get_prompt() → CLI traces list to confirm the new prompt name appears.
• Debug a regression → CLI traces list --limit 50 --json filtered by name → CLI traces get for the suspect trace → CLI observations-v2s list for the failing span → MCP getPrompt to inspect the prompt version that ran.
• Onboard a new repo to Langfuse → eval "$("<skill-base-dir>/../bao-vault/scripts/langfuse_env.sh")" (or export LANGFUSE_* directly) → bash "<skill-base-dir>/scripts/install-mcp.sh" (registers MCP with all three agents) → python3 "<skill-base-dir>/scripts/install_stop_hook.py" (Claude Code session traces) → instrument app code (see references/instrumentation.md).

When composing, return one unified response covering all steps. Don't ask the user to invoke each step separately.

Documentation access

Three methods, in preference order. Always prefer your application's native fetch tool (WebFetch, mcp_fetch, etc.) over curl when available — the URL patterns below work with any tool.

1. Documentation index (llms.txt)

https://langfuse.com/llms.txt

Structured list of every doc page with titles and URLs. Scan for relevant titles, then fetch the page directly.

2. Fetch a page as markdown

Append .md to any doc path, or send Accept: text/markdown:

https://langfuse.com/docs/observability/overview.md

3. Search across docs + GitHub Issues/Discussions

https://langfuse.com/api/search-docs?query=<url-encoded-query>

Returns JSON with matching pages and excerpts. Useful when debugging — issues and discussions are indexed alongside docs. Responses can be large; extract only relevant portions.

Workflow: start with llms.txt to orient → fetch specific pages once you know which ones → fall back to search when the topic is unclear.

Skill feedback

When the user says this skill gave wrong/outdated guidance, missed a workflow, or could be improved, offer to submit feedback to the upstream Langfuse skill maintainers. Process: references/skill-feedback.md.

Do not trigger this for issues with Langfuse the product — only for issues with this skill's instructions and behavior.