Langfuse vs LangSmith: which one fits your stack?

Langfuse is MIT-licensed open source, with a managed cloud that runs the same code you can self-host. LangSmith is proprietary, owned by the LangChain team, and there is no self-hosted build. If your procurement rules require an OSS option, or you need to inspect the code path that ships your prompts to a vendor, that decision is effectively already made before you read a single benchmark.

The practical consequence: with Langfuse you can switch from cloud to self-hosted in a weekend and keep the same SDK in your application code. With LangSmith you are locked to a single vendor for the lifetime of the project. For a regulated workload — healthcare, finance, public sector — that is often a hard constraint rather than a preference.

Both platforms support nested traces, LLM spans, tool calls, token usage, and evaluation runs against those traces. The difference is opinionation. LangSmith is more opinionated and tightly integrated with LangChain and LangGraph: if you build agents on those frameworks, much of the instrumentation is wired in automatically and the dataset UI is the most polished in the category.

Langfuse is more framework-agnostic. Its SDKs are clean, but you wire them in yourself for custom stacks. For teams running DSPy, Haystack, a hand-rolled agent loop, or a polyglot mix of Python and TypeScript, Langfuse is the easier fit. If you are on LangChain end-to-end, LangSmiths defaults save you real engineering time.

Langfuse ships an official Docker compose and a Helm chart, both maintained against a Postgres dependency. The minimum production deploy is a single node plus a managed Postgres, and you can run the whole stack on a small VM for development. LangSmith is cloud-only — no self-host path, no enterprise appliance, no air-gapped option.

Langfuse Cloud is free below a generous monthly event threshold, then usage-based on events and storage. LangSmith is per-seat with usage tiers on top. For small teams the free tier of Langfuse Cloud is hard to beat. For larger organisations with predictable seat counts, LangSmiths per-seat pricing is easier to forecast in a procurement cycle — even if the unit cost is higher.

The honest answer is: run your projected volume through both pricing calculators before you decide.

Pick Langfuse if you are an OSS-first team, run a multi-framework stack, work in a regulated environment, or are cost-sensitive at low to mid volume. The DX is excellent, the release cadence is fast, and you can move workloads between self-hosted and cloud without rewriting instrumented code.

Pick LangSmith if your stack is LangChain or LangGraph end-to-end, you want a single-vendor story for support and procurement, and you value opinionated defaults over flexibility. The dataset UI in particular is one of the most polished in the category, and if you run experiments against LangChain own evaluators the integration is friction-free. The tradeoff is vendor lock-in and no self-host path.