Langfuse
Langfuse ingests OpenTelemetry: point the OTLP driver at it. Install e-track[otel].
ETRACK_OTLP_TARGETS='{"langfuse": {
"endpoint": "https://cloud.langfuse.com/api/public/otel",
"headers": {"Authorization": "Basic <base64 of public_key:secret_key>"}}}'
Self-hosted: http://<host>:3000/api/public/otel. Build the header with
printf '%s' "pk-lf-…:sk-lf-…" | base64 -w0.
| In Langfuse | Comes from |
|---|---|
| sessions | the thread (langfuse.session.id) |
| trace | the run — its id is the run id |
| observation types | the span kind: generation (llm), tool (tool, mcp), agent (agent, subagent), retriever, embedding, evaluator, chain (run, node), span (the rest) |
| usage and cost | exclusive buckets: input, output, input_cache_read, input_cache_creation, output_reasoning — cached input is priced as cached |
| metadata | the root span's app.* attributes (→ LangGraph) |
Measured against Langfuse's own LangChain handler on the same runs (deeptrack's tracing A/B): the same nesting, generations and tokens. Before 0.1.2 cached input was priced at the full input rate — up to twice the cost; it is priced apart since. Everything else about the driver → OpenTelemetry.