Langfuse, traces you can keep inside the client's network

It is open source and self-hostable, which in a forward deployed context is not a preference, it is the whole argument. It is built on OpenTelemetry, so your instrumentation is a standard rather than a proprietary SDK you will have to tear out in eighteen months. Traces cover non-model spans as well as model calls, so the retriever, the reranker and the tool calls appear in the same timeline as the completion, which is almost always where the actual bug is hiding. Sessions group multi-turn conversations, and agent runs render as a graph rather than a wall of logs. Prompt management gives versioning and deployment separate from a code release, so a subject-matter expert on the client side can change wording without waiting for your deploy pipeline. Datasets, model-as-judge scoring, code evaluators, user feedback and manual labelling live in the same place as the production traces, which is how you stop exporting CSVs between two vendors that disagree about what a run is.

Self-hosting is not one container and a good afternoon. A production deployment wants Postgres, ClickHouse, Redis and blob storage, so four stateful services become your responsibility inside a client environment where you may not control the cluster or the on-call rota. Teams that begin self-hosted often drift to the cloud offering, which is a perfectly reasonable outcome unless data residency was the reason you picked it, in which case you have quietly undone your own decision. The evaluation half is broad rather than deep: good enough to notice a regression, thinner than a dedicated eval tool the moment you want statistical comparison across runs. And trace volume is real money in storage. Nothing prevents an agentic workload making forty tool calls per request from becoming the largest table in the client's database. Sampling exists, and you will need it considerably sooner than you expect to.

Spend an hour on it if you are about to debug a multi-step pipeline by adding print statements, which is where most first deployments genuinely are. Skip it if you have one prompt, one call and no retrieval, because a log line does that job and the infrastructure is not free. If the client's data is not allowed to leave their network, this is the shortest honest path to seeing what your system actually did.

Helicone if a one-line proxy change is all you can get approved. Phoenix if you want OpenTelemetry-native tracing with no vendor at all.