LangSmith vs Langfuse
Both trace agent runs, score outputs and catch regressions before a user does. The choice usually comes down to who is allowed to hold the data, not which dashboard looks better.
The short answer
Choose LangSmith when your team already builds on LangChain or LangGraph and wants tracing managed for you. Choose Langfuse when you need to self-host, run a framework other than LangChain, or keep trace data inside your own infrastructure.
LangSmith is built by the LangChain team and integrates with one line of setup if you already use LangChain or LangGraph. It is a hosted product first, with self-hosting available on higher tiers.
Langfuse is open source under an MIT-family license for its core. It ships a self-hosted Docker deployment as a first-class path, and instruments any framework through OpenTelemetry or its SDK. The two tools log traces, scores and prompt versions in similar shapes. The decision is mostly about deployment, not features.
Side by side
Positioning taken from each project's own documentation, checked in August 2026.
| Dimension | LangSmith | Langfuse |
|---|---|---|
| Built by | The LangChain team | Langfuse GmbH, an independent company |
| Source | Closed source | Open source core, MIT-family license |
| Self-hosting | Available on enterprise plans | First-class, documented Docker deployment |
| Framework fit | Deepest with LangChain and LangGraph | Framework neutral, via SDK or OpenTelemetry |
| Setup outside LangChain | Works, but loses the one-line integration | Same setup regardless of framework |
| Evaluation | Built-in datasets, LLM-as-judge and human review | Built-in datasets, LLM-as-judge and human review |
| Prompt management | Included, tied to the LangSmith workspace | Included, versioned independently of any framework |
| Best-fit team | Already committed to the LangChain ecosystem | Wants data residency control or runs a mixed stack |
- Agent runLangGraph, a plain API call, or anything else
- InstrumentationSDK call or OpenTelemetry export
- LangSmith cloudManaged by the LangChain team, unless you pay for self-hosting
- Langfuse instanceYour own Docker deployment, or their cloud
Both tools trace the same run. Where the data lands afterward is the actual decision.
LangSmith
Where it wins
- One-line setup if your agent already runs on LangChain or LangGraph.
- Built by the same team, so new LangChain and LangGraph features tend to show up in tracing quickly.
- The hosted product needs no infrastructure of your own to run or patch.
- Evaluation datasets and human review sit in the same workspace as the traces.
Where it hurts
- Self-hosting is gated behind enterprise pricing, so most teams run it as someone else's cloud by default.
- Outside LangChain, the integration is another SDK call rather than the one-line hook.
- Trace data leaving your infrastructure is a real conversation with a security team in regulated industries.
- The product roadmap follows LangChain's roadmap, which is fine until your stack diverges from it.
Langfuse
Where it wins
- Self-hosting is documented and treated as a normal deployment path, not an upsell.
- Framework neutral from the start, so a mixed stack or a later framework switch does not mean re-instrumenting.
- OpenTelemetry support means a lot of existing observability tooling already speaks its language.
- An open core means you can read the code that touches your trace data before trusting it with production.
Where it hurts
- Self-hosting is your infrastructure to run, patch and back up, which is real ongoing work.
- Being framework neutral means less of the LangChain-specific convenience some teams expect.
- Smaller company than the LangChain team, so support channels and release cadence are different.
- The managed cloud option exists, but it gives up the self-hosting advantage that draws most teams to it.
How to choose
- Choose LangSmith if your team is fully on LangChain or LangGraph and wants tracing set up in minutes with no server to run.
- Choose LangSmith if you are fine with trace data sitting in a managed cloud and have no residency constraint against it.
- Choose Langfuse if a contract, regulation or internal policy says trace data stays inside your own infrastructure.
- Choose Langfuse if your stack mixes frameworks, or you expect to change frameworks later without re-instrumenting.
- Choose neither yet if nobody has written down what a trace should look like when something goes wrong. Both tools show you data; deciding what counts as a failure is a separate job.
Questions engineers ask before committing
01Can I switch from LangSmith to Langfuse later?
Yes, and it is a moderate rewrite rather than a rebuild. Both tools wrap the model calls in your code. Migrating means swapping the instrumentation calls and re-pointing dashboards, not changing the agent logic itself. Budget time to rebuild evaluation datasets, since those rarely export cleanly between vendors.
02Does Langfuse work with LangChain and LangGraph?
Yes. Langfuse ships a LangChain callback handler alongside its own SDK and OpenTelemetry support. LangChain and LangGraph runs trace into it without losing the framework-neutral setup that also covers other frameworks.
03Is Langfuse's free self-hosted version missing features LangSmith has?
The self-hosted core covers tracing, evaluation, prompt management and datasets. Some enterprise features, such as advanced access controls, are reserved for Langfuse's paid tiers. Compare the current feature matrix on each vendor's pricing page before deciding, since both change it periodically.
04Do we need an observability tool if we already log requests?
Plain logs record what happened. An observability tool built for LLM calls adds structure: nested traces across multi-step agent runs, cost per call, and scoring against a known-good answer set. Once an agent makes more than one model call per task, plain logs stop being enough to debug a bad run.
05How much engineering time does self-hosting Langfuse take?
A single Docker Compose deployment is a short setup for a team that already runs its own infrastructure. Running it reliably in production, with backups and updates, is the ongoing cost worth weighing against LangSmith's managed option before committing.
