Guide
LangSmith vs Langfuse is the choice many teams face when they add tracing and evals to an agent that has reached production. Both record each step of a run, both score outputs, and both manage prompts. They differ in who owns the code, where it can run, and how closely each ties to LangChain.
This guide compares them on the points that decide it, then says when to pick each. For other tools, read our LangSmith alternatives guide. For the criteria behind any choice, see how to choose an LLM observability tool.
| LangSmith | Langfuse | |
|---|---|---|
| Made by | LangChain | Langfuse, acquired by ClickHouse in January 2026 |
| Open source | Platform is hosted; SDK is MIT | Core is MIT; enterprise directories licensed separately |
| Self-hosting | Add-on to the Enterprise plan, on Kubernetes | Free; Docker Compose for testing, Kubernetes with Helm for production |
| Framework fit | Works with any framework; setup is shortest on LangChain and LangGraph | Callback integration for LangChain and LangGraph; OpenTelemetry for others |
| Tracing | Yes | Yes |
| Evaluation | Datasets, offline and online evals | Datasets, experiments, LLM-as-a-judge, annotation queues |
| Prompt management | Prompt Hub and Playground | Versions, labels and client-side caching |
| OpenTelemetry | OTLP ingestion | OTLP over HTTP; no gRPC |
| Free hosted tier | 1 seat, up to 5k base traces a month | 2 users, 50k units a month, 30 days of data |
| Paid hosted entry | $39 per seat a month | $29 a month, unlimited users |
LangSmith is LangChain's hosted platform for tracing and evaluating LLM applications. It runs as a cloud service, and self-hosted and hybrid deployments are an Enterprise option. LangChain's own pages call it framework agnostic, so you can trace code that never imports LangChain.
In practice it is closest to LangChain and LangGraph, where tracing starts with an environment variable. If your agents already run on LangGraph, that is a real advantage.
Langfuse is an open source platform for tracing, evaluating and managing prompts, and you can self-host it for free. Its repository's license file applies the MIT license outside its enterprise directories, which are licensed separately.
ClickHouse announced the acquisition in January 2026. The announcement says Langfuse stays open source and self-hostable, Langfuse Cloud keeps running as before, and no licensing changes are planned. Read that as a statement of intent, and check the license file when you deploy.
Self-hosting is the sharpest difference. Langfuse documents Docker Compose for local use and testing, noting it is a single VM without high availability, and Kubernetes with Helm for production. Some organisation features need a license key.
LangSmith self-hosting is an add-on to the Enterprise plan and installs on Kubernetes.
Hosted pricing differs in shape. LangSmith charges per seat: Developer is free for one seat with up to 5,000 base traces a month, and Plus is $39 per seat a month with up to 10,000 base traces.
Langfuse charges per month with unlimited users: Core is $29 with 100,000 units, then $8 per extra 100,000, and Pro is $199 a month. Prices change, so confirm them on LangSmith and Langfuse before you budget.
The two meter usage differently, so compare using your own trace volume. A team of five developers pays for five seats on LangSmith Plus and a flat fee on Langfuse Core before usage.
Both cover the same three jobs, with different emphasis. For tracing, each records the steps of a run and accepts OpenTelemetry data, and LangSmith takes OTLP at its own endpoint. Langfuse accepts OTLP over HTTP, as JSON or protobuf, and does not support gRPC.
For evaluation, LangSmith documents datasets with offline and online evaluation, using human review, code rules, LLM-as-judge and pairwise comparison. Langfuse documents datasets, experiments, LLM-as-a-judge, annotation queues and user feedback, plus custom workflows through its API.
For prompts, LangSmith has a Prompt Hub and a Playground for testing prompts against models. Langfuse has versioning and labels, with prompts cached client-side by its SDKs. If a prompt playground matters to you, check Langfuse's current docs, since its prompt management page does not describe one.
Choose LangSmith when you build on LangChain or LangGraph and a hosted service is fine. Setup is shortest there, and one vendor covers the framework and the tracing.
It also suits teams that already have an enterprise agreement with LangChain, since self-hosting is available on that plan. Seat pricing works well for a small team of one or two developers.
Choose Langfuse when open source or self-hosting without an enterprise contract matters. You can run it on your own servers for free, which keeps prompts and traces inside your network.
It also fits teams on mixed or custom frameworks. It works through LangChain callbacks and OpenTelemetry, so one tool can cover several agent stacks.
Flat monthly pricing with unlimited users helps larger teams. Self-hosting still costs servers and time, so compare that with a paid plan.
Moving between them is mostly an instrumentation change. Your traces, datasets and prompts live inside each product, so plan to export what you need and re-create the rest.
If you start on Langfuse or LangSmith and later switch, run both for a week first. Send the same traces to each, using OpenTelemetry where your framework supports it, and compare one failed run in each tool.
Move datasets and prompts once you have chosen, and check how each tool exports them before you commit. We did not verify either vendor's export steps for this guide.
LangSmith or Langfuse usually comes down to six questions.
Whichever you choose, check whether it records the decisions behind an agent's actions and not only the calls it made.
FleetHelp is not a tracing tool, and it does not replace either of these. A tracing tool shows you what happened in a run. FleetHelp helps when you can see what happened and still cannot tell why, such as a LangGraph agent that loops in production.
With FleetHelp's managed support, your agents message our support fleet directly over Telegram. Ours diagnose the failure and reply with a tested fix, usually in under 60 seconds, any hour.
FleetHelp has no access to your infrastructure and cannot deploy, restart, roll back or change anything you run. Plans start at $99 a month; see pricing and subscribe.
Neither is better in general, since Langfuse fits when you need open source or free self-hosting, and LangSmith is the shorter path on LangChain or LangGraph with a hosted service. Both cover tracing, evals and prompt management, so the choice comes down to where the data lives and what you already use.
Yes, but on different terms. Langfuse documents Docker Compose for testing and Kubernetes with Helm for production, and says you can self-host it for free. LangSmith self-hosting is an add-on to the Enterprise plan and installs on Kubernetes, so teams without an enterprise contract usually pick Langfuse.
Langfuse says so: its January 2026 announcement states that ClickHouse acquired it, that it stays open source and self-hostable, and that no licensing changes are planned. The repository's license file still applies MIT outside its enterprise directories, so check both before you commit.
Yes. LangSmith says it is framework agnostic and works with any way of building an agent, and Langfuse integrates with LangChain through callbacks and also accepts OpenTelemetry traces. LangSmith is still shortest to set up on LangChain and LangGraph, so test your own stack with a real trace.
A tracing tool tells you what your agent did, and a failing run in production still needs someone to work out why. FleetHelp's support fleet takes that question from your agent over Telegram and replies with a tested fix, without access to your infrastructure. See plans and subscribe, starting at $99 a month.