LangSmith → Langfuse
What it saves, what actually moves, what you rebuild — and the thing that catches people.
What it costs, and what it saves
LangSmith bills per trace and teams with real volume reach hundreds monthly. Langfuse self-hosted is Postgres and a container.
Moves cleanly
The concept of a trace. Langfuse has an OpenTelemetry-compatible ingest and SDKs for the same languages.
You rebuild
Instrumentation calls, and any evaluation datasets you have built up.
What LangSmith costs you today
There is a free developer tier with a monthly trace allowance. Paid plans are seat-based with usage pricing on traces beyond the included volume, and enterprise plans (including self-hosted deployment) are quote-only. The cost that surprises people is trace volume rather than seats: a chatty agent generates many spans per user action, so the bill tracks how verbose your application is. Verified against LangChain's published pricing on 2026-07-30 — confirm current figures, as this pricing has changed repeatedly.
What actually holds you in
Moderate and mostly in the accumulated history. Instrumentation is the movable part — if you instrumented with OpenTelemetry conventions, switching backends is configuration. What does not move is your trace history, evaluation datasets and the scoring you built on top of them, and those are the things that took months to become useful.
What you are moving to
Langfuse is the most widely adopted open LLM observability platform: tracing, evaluation, prompt management and cost tracking, deployable with Docker on your own infrastructure so traces containing customer data never leave your network. It is deliberately framework-agnostic — LangChain, LlamaIndex, the provider SDKs directly, or OpenTelemetry — which makes it the natural exit from a LangChain-shaped tool. At 32k stars with commits landing daily it is the healthiest project in this category. Licensing needs precision: the core is MIT Expat, while the ee/ directories carry a separate enterprise licence.
Free to self-host — Docker Compose or Kubernetes, costs are your own infrastructure. A managed cloud with a free tier and paid plans exists if you would rather not run it.
Langfuse strengths
- Self-hosted: prompts and completions stay inside your network
- Framework-agnostic — not tied to any one orchestration library
- Tracing, evals, prompt management and cost tracking in one tool
- 32k stars and daily commits — the most active project in this category
What you give up
- Self-hosting means running Postgres, ClickHouse and the collector yourself
- The MIT core stops at the ee/ directories; SSO and some enterprise features sit behind a separate licence
- Evaluation tooling is less polished than LangSmith's if you live inside LangChain
- Trace storage grows fast — retention policy is your problem to plan
The migration, step by step
- 1Deploy Langfuse with docker compose and Postgres
- 2Swap the SDK — for LangChain applications this is a callback handler change
- 3Verify traces appear with the nesting you expect before removing LangSmith
- 4Recreate evaluation datasets; these do not transfer and are often the real work
- 5Run both briefly, then cancel
The gotcha
Langfuse's core is MIT but some enterprise features are separately licensed. Check which side of that line SSO and fine-grained permissions fall on before you plan around them.
When to stay on LangSmith
You are deep in the LangChain ecosystem and value the first-party integration.
More Layer 4 migrations
Figures verified 2026-07-30 against vendor pricing pages. Prices change and migrations differ by estate — treat the cost delta as a starting model, not a quote. Rankings and recommendations here are merit-only; affiliate income never changes a verdict. See our methodology.