Langfuse, the LLM Observability Stack I Actually Run

Langfuse is the only LLM observability tool I let through to empire production. I tried PostHog with the LLM addon, I tried Helicone, I tried OpenLLMetry, and I came back to Langfuse because it does the one thing I want, which is trace every Claude call across every empire app, in one searchable place, without metering me on traces.

What I run today is Langfuse Cloud-Free for the public empire endpoints (PramaanAI API, taxwallaai webhooks, autokaam social composer). The free tier gives me 50K units a month and that has been enough so far. For the experimental stuff and anything that touches private data, I run Langfuse self-hosted on the same Coolify box that hosts kaam-tracker. The setup is one compose file in their repo, except I am Docker-banned in my empire, so I deploy it as separate Coolify services: Postgres, ClickHouse, Redis, and S3-compatible storage alongside the Langfuse containers.

The reason traces matter more in 2026 than they did in 2025 is that empire workloads are now agentic. A single user request fans out to four or five Claude calls, each with tool use, each with retries. When something silently produces a bad output, the only way to find out which sub-step lied is to look at the trace tree. I caught a bug last month where the autokaam social composer was passing the wrong article slug into the headline-rewriter call, and the only reason I caught it was a Langfuse trace that showed input slug and output slug mismatched. Without Langfuse I would have been grepping logs for hours.

The Indian-operator angle here is cost. Datadog's pricing for the equivalent stack is usage-based across several products. PostHog with the LLM hooks is reasonable but its UX for prompt comparison is weak. Langfuse self-hosted costs me the marginal Postgres, ClickHouse, Redis, and object storage on the Coolify box, and the cloud free tier covers most empire production.

What I would change is the eval module. The Langfuse evals UI is good for spot-checking but bad for building a regression bench. I run a separate evals harness in Python and only push the headline numbers back into Langfuse as datasets. Two systems, fine.

If you build anything LLM-shaped this year and you are not tracing, you are flying blind. Langfuse is the cheapest way to see clearly.

Agentic workflows fan out a single request into several LLM calls. Without traces you cannot debug. Per-call observability vendors charge brutally; Langfuse is open source, can be self-hosted, and its cloud Hobby plan is free. You cannot optimize what you cannot measure.