Langfuse LLM Observability
AdoptFactorial rates Langfuse as Adopt (2026). We run our own Langfuse instance, so prompts and client data in traces stay on our infrastructure.
Langfuse records what an LLM application actually did: every trace with its prompts, model responses, tool calls, latencies, token counts and costs, grouped by session and user. On top of the traces it offers prompt management, datasets and scoring for evals.
It can be used as a cloud service or self-hosted. Since version 4 it ingests data through OpenTelemetry (OTLP), so any instrumented application can send traces to it.
Updates
We run our own Langfuse instance, so prompts and client data in traces stay on our infrastructure. Our AI products report to it: Subtool traces every agent run with one span per model iteration, the Rust engine of FlowDrop maps workflows to traces and nodes to spans, and our controlling app, built on Drupal, traces the generation of its AI summaries. It is where we look first when an agent gives a strange answer or costs more than expected.
One lesson from the upgrade to Langfuse 4: the old ingestion API answered with a success code while silently rejecting the data. We are moving our exporters to OTLP, and treat exports that the server rejects as errors that must show up in the logs.