Langfuse vs LangSmith: Open Source vs the LangChain Platform

Langfuse vs LangSmith comes down to one question: do you want to own the observability stack, or rent the platform that also runs your agents? Both tools trace LLM calls, run evaluations, and version prompts. The difference is in what surrounds that core.

• Langfuse is open source. Its product features are MIT-licensed, you can self-host it with no licence fee, and it only observes your app. It doesn't host or run your agent.
• LangSmith is LangChain's commercial platform. It covers tracing, evaluation and prompts, and it also adds a managed runtime for agents (LangSmith Deployment) and an issue-fixing agent (LangSmith Engine). Self-hosting it is an Enterprise add-on.
• Both work without LangChain. Both accept OpenTelemetry traces. "LangSmith only works with LangChain" is an old myth.
• One ownership fact to know: Langfuse was acquired by ClickHouse in January 2026. Langfuse says the licence and self-hosting aren't changing.

If you're searching for a LangSmith alternative that's open source, Langfuse is the obvious one. That doesn't make it the right pick for everyone. Here's where each one wins.

Quick comparison

Langfuse vs LangSmith on licence and ownership

This is the section most comparisons get half right.

Langfuse's LICENSE file is MIT for everything outside the ee/, web/src/ee/ and worker/src/ee/ directories. Code inside those directories falls under a separate enterprise licence. Langfuse's open-source handbook spells out the split: tracing, evaluations, prompt management, experiments, annotation and the playground are MIT with no usage limits. A commercial licence is needed only for enterprise modules such as SCIM, audit logging and data retention policies when you self-host.

So "Langfuse is MIT" is mostly true. "Every line of Langfuse is MIT" isn't.

Then there's ownership. On January 16, 2026, Langfuse announced it was joining ClickHouse. The post says Langfuse stays open source and self-hostable, with "no planned changes to licensing," and that Langfuse Cloud keeps running as-is. The repo's copyright line now reads ClickHouse, Inc. Nothing has changed for users so far. But if your team picked Langfuse partly because it was independent, that's no longer true.

LangSmith is proprietary and built by LangChain, the company behind the LangChain and LangGraph frameworks. There's no source to fork. You're buying a product.

Where Langfuse is stronger

Self-hosting without a sales call. The self-hosting docs offer Docker Compose for local use and testing, Kubernetes via Helm for production, and templates for AWS, Azure and GCP. You'll run four storage services: Postgres, ClickHouse, Redis or Valkey, and S3-compatible blob storage. That's real ops work. But you can start today, and your traces never leave your infrastructure.

One product, three ways to run it. Langfuse says the OSS self-host, the Enterprise self-host and Langfuse Cloud share the same codebase, and you can switch between them. Start on Cloud, move in-house when compliance asks. No rewrite.

Built on OpenTelemetry. The tracing docs say Langfuse uses OpenTelemetry to reduce vendor lock-in. If your app already emits OTel spans, you point them at Langfuse. Native integrations cover LangChain, the OpenAI SDK, the Vercel AI SDK, LlamaIndex, CrewAI, LiteLLM and more.

Prompt management that non-engineers can use. Langfuse prompts carry versions and labels like production and staging. Your app fetches them at runtime with client-side caching, so a prompt change ships without a code deploy. Each prompt version links to its traces, so you can see which version performs better.

Seat count doesn't drive the bill. Langfuse Cloud bills by units: every trace, observation and score you send. Its paid tiers allow unlimited users. For a big team that mostly reads dashboards, that model matters.

Where LangSmith is stronger

It goes past observability. LangSmith is one platform with four jobs: observability, evaluation, prompt engineering, and deployment. LangSmith Deployment is a runtime for agents, with durable execution, streaming and horizontal scaling. You can deploy to LangChain's cloud, BYOC, or your own infrastructure. The docs call it framework-agnostic: LangGraph and LangChain apps, Google ADK agents, and other frameworks, including the Claude Agent SDK and CrewAI, through wrappers. Langfuse has nothing like it. It watches your agent run somewhere else.

Evals that run on live traffic. LangSmith evaluation splits into offline evals on curated datasets and online evals on real user interactions. Evaluators include human review, code rules, LLM-as-judge and pairwise comparison. You can add real production traces to a dataset, so a failure you saw once becomes a test you run every time.

An agent that works your traces. As of September 2026, LangSmith Engine scans your tracing projects, clusters recurring issues, diagnoses root causes, and proposes fixes. It can open a pull request for agents built with Deep Agents, LangChain or LangGraph. On LangSmith Cloud it runs on LangChain-managed inference and bills in LangChain Standard Units (LSUs). Its red-teaming feature is marked beta. Nothing in Langfuse's docs matches this today.

Zero-friction if you're already on LangChain. Tracing a LangChain or LangGraph app takes a couple of environment variables. The Prompt & Context Hub versions prompts and whole bundles of agent instructions and tools, and connects to the public LangChain prompt hub.

It isn't locked to LangChain either. LangSmith supports OpenTelemetry-based tracing "from any OpenTelemetry-compatible application." It also integrates with OpenAI, Anthropic, CrewAI, the Vercel AI SDK and Pydantic AI.

Self-hosting and billing, side by side

Most teams end up choosing on these two points. Neither vendor hides them, but they're easy to misread.

Self-hosted LangSmith is "an add-on to the Enterprise plan." You get a licence key from LangChain's sales team and install on Kubernetes. A hybrid setup exists too: you run the Agent Servers in your own infrastructure and send traces to LangSmith Cloud or a self-hosted instance.

Langfuse self-hosting needs no licence key unless you want the enterprise modules.

On billing, per each vendor's pricing page:

Check LangSmith's pricing page and Langfuse's before you budget. Both change.

Deciding: a short checklist

1. Compliance says traces can't leave your network, and there's no enterprise budget? Langfuse.
2. You want one vendor to trace, evaluate and host your agents? LangSmith.
3. Your stack is LangGraph end to end? LangSmith is the path of least resistance. Langfuse still works through its LangChain integration.
4. You're mixing frameworks, or might switch next year? Either works over OpenTelemetry. Langfuse's open code makes leaving cheaper.
5. Many people need read access? Langfuse's unit-based billing won't grow with your headcount.
6. You want automated issue triage from traces? LangSmith Engine. Nothing equivalent exists in Langfuse today.

Still picking a framework? Read LangChain alternatives and our overview of agentic AI frameworks first. Observability follows the framework, not the other way round.

Where Taku fits

Langfuse and LangSmith are for teams who build LLM apps and need to see inside them. If that's you, pick one of them. Taku doesn't trace, evaluate or deploy anything.

Many people reading about agent tooling never get that far, though. They see an AI workflow someone else built, want to use it, and stall at the setup. Taku is an AI-native desktop workspace for that situation. You mirror an AI app, skill or workflow someone already got working, run it, remix it for your own files, and keep it. Taku is in Beta, and the Mac app is available now. Browse the free app library to see what that looks like.

FAQ

Is Langfuse fully open source?

Mostly. Tracing, evaluations, prompt management, experiments, annotation and the playground are MIT-licensed with no usage limits. Code in the ee directories is under a separate enterprise licence. When you self-host, that licence covers features like SCIM, audit logging and data retention policies.

Is LangSmith only for LangChain apps?

No. LangSmith accepts OpenTelemetry traces from any compatible app and integrates with OpenAI, Anthropic, CrewAI, the Vercel AI SDK and others. LangSmith Deployment also runs agents from frameworks beyond LangGraph.

Can I self-host LangSmith?

Yes, but only as an add-on to the Enterprise plan, with a licence key from LangChain's sales team. It installs on Kubernetes.

Who owns Langfuse now?

ClickHouse. The acquisition was announced on January 16, 2026. Langfuse said there were no planned licensing changes and that Cloud and self-hosting continue as before.

Which is better for LLM observability?

For tracing and evals alone, both are mature enough. Langfuse wins on openness, self-hosting cost and seat-free billing. LangSmith wins if you also want managed deployment, online evaluation, and automated issue triage in one place.

Can I use Langfuse with LangGraph?

Yes. Langfuse's LangChain integration traces LangChain and LangGraph apps through a callback handler, and you can also send OpenTelemetry spans directly.