LangSmith vs Langfuse: which LLM observability tool fits?

Before you compare a single feature, answer this: where do you want your agent traces to live, and who do you want running the box they sit on? If the answer is your own servers, under your own control, one of these tools makes that easy and public. If you'd rather hand the whole agent stack to one company and get on with building, the other is designed for exactly that.

That's the decision underneath this pair. Both trace calls, score outputs and show dashboards, so the overlap is wide. Where they part ways is licensing, how much of the stack each one wants to own, and who's behind them. I'll refer to them as the suite and the open tool for much of this page, because that's the clearest way to keep them apart.

Where do you want your traces to live?

The open tool describes itself as open source and publishes self-hosting guides for Docker Compose, Kubernetes with Helm, and Terraform on AWS, GCP and Azure. Its docs say a self-hosted install runs the same infrastructure as its own cloud service. That's a strong promise for anyone with data that can't leave their network. One limit worth knowing: some add-on features need a license key, so open source doesn't mean every feature is unlocked by default.

The suite offers managed cloud, bring-your-own-cloud and self-hosted options too. The difference is the route in. Its FAQ points you to the vendor to talk through the right option for data residency, so self-hosting is a conversation rather than a guide you follow alone. If you need a private deployment next week without a call, that's a real difference. If you'd welcome a vendor walking you through it, it isn't.

Seven gaps at a glance

Theme	LangSmith	Langfuse
Licensing and hosting	Managed cloud, bring-your-own-cloud or self-hosted; data residency goes through the vendor	Open source with public self-hosting guides; some add-ons need a license key
Platform scope	Part of a wider agent platform with deployment, sandboxes, an LLM gateway and no-code agents	Four areas: observability, prompt and skill management, evaluation, and an AI gateway in beta
Framework ties	Made by LangChain; its FAQ says it works with any LLM framework	Neutral pitch, 100+ integrations, OpenTelemetry to reduce lock-in
Storage	SmithDB, a purpose-built trace database, on object storage and Postgres	Open components: Postgres, ClickHouse, Redis or Valkey, and S3 or blob storage
Production monitoring	Latency and error dashboards, online LLM-as-judge evals, webhook and PagerDuty alerts, automatic Insights	Tracing, dashboards, alerts, an in-app agent called Halo, and access through MCP, CLI, API and SDKs
SDK languages	Python, TypeScript, Go and Java	Python and JS, plus any OpenTelemetry exporter
Ownership	Operated by LangChain	Homepage story says it has joined ClickHouse

Who's behind each one, and why it changes the choice

This is the part most comparisons skip. The open tool's homepage carries a story, as of Nov 2026, about joining ClickHouse, with the stated goal of continuing to build the best AI engineering platform. Its self-hosted stack already runs on that database, so the open-source project and its new owner are now tied together. That isn't a warning sign on its own, but it's worth knowing who you're betting on before you put years of traces into a system.

The suite has its own version of this. Its observability page reads as one piece of a bigger agent platform, with deployment, sandboxes and a gateway sitting beside it in the navigation. That's convenient if you want all of it. It's less appealing if you only want tracing and don't fancy a vendor nudging you toward the rest. Both are also adding a gateway: the open tool labels its AI gateway beta, while the suite lists an LLM gateway in its menu. If a gateway matters to you, don't assume either one is mature from the label alone.

Frameworks, languages and lock-in

The suite comes from the company behind the langchain, langgraph and deepagents frameworks, and those sit in its navigation. Its FAQ is clear that it works with any LLM framework, naming the OpenAI, Anthropic and Vercel AI SDKs. So you don't have to use the house frameworks, but if you already do, you'll feel at home faster.

The open tool pitches itself as framework-neutral, with native Python and JS SDKs, more than 100 library and framework integrations, and OpenTelemetry tracing to reduce lock-in. On languages, the suite lists more first-party SDKs: Python, TypeScript, Go and Java. If your back end is written in Go or Java, that's a point in its favour, though the open tool's OpenTelemetry support means any compatible exporter can send it data.

What running each one in production looks like

Once an agent's live, you want to know when it breaks and why. The suite lists dashboards for latency and error rates, online LLM-as-judge evals, alerts through webhooks and PagerDuty, and an automatic Insights feature that clusters traces into usage patterns. That last one is useful if you've got too many traces to read by hand.

The open tool offers tracing, dashboards and alerts, an in-app agent called Halo, and says every feature can be reached through MCP, a CLI, the API and its SDKs. That suits teams who want to script everything or wire observability into their own tools. Under the hood, the suite stores traces in SmithDB, which the vendor says gives sub-second queries across millions of traces. The open tool uses familiar open components, which your engineers may already know how to run and back up.

The verdict: decide on hosting first, scope second

Go with Langfuse if open source, a self-hosting path you can follow without a sales call, and framework-neutral tracing are what you care about most. Go with LangSmith if your team already builds with LangChain tools, writes back-end code in Go or Java, or wants observability, evals, deployment and a gateway from one vendor. Neither site gives grounds to call one better overall. It comes down to where your data has to sit and how much of the agent stack you're happy to hand to one company.

If observability is only one part of the toolkit you're putting together, my guide to the AI tools developers rely on covers coding, testing and the other layers around it.

Alternatives if neither one feels right

Helicone

Best for: A gateway-style proxy with logging that sits in front of your model providers, with request, session and user views plus alerts.

Arize Phoenix

Best for: Open-source tracing and evaluation you can start locally with one command, with a managed upgrade route if you outgrow it.

Braintrust

Best for: Built for teams that put evals at the centre of releases: trace prompts and tool calls, score outputs and block bad releases.

Opik

Best for: Open-source observability and evals from Comet, with LLM-as-a-judge metrics, prompt management and a debugging assistant called Ollie.

Do you make an LLM observability or evaluation tool? Tool makers can ask to be considered for a featured place on this page.

Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

Frequently asked questions

Is Langfuse open source?

Yes, it's described as open source, and it publishes self-hosting guides for Docker Compose, Kubernetes and Terraform. Some add-on features need a license key.

Can I self-host LangSmith?

Its site lists a self-hosted option alongside managed cloud and bring-your-own-cloud, but the FAQ points you to the vendor to choose the right setup for data residency.

Do I need to use LangChain to use LangSmith?

No. Its FAQ says it works with any LLM framework, including the OpenAI, Anthropic and Vercel AI SDKs, though it's made by the same company.

Which one supports more SDK languages?

The suite lists Python, TypeScript, Go and Java SDKs. The open tool lists Python and JS SDKs, and accepts data from any OpenTelemetry exporter.

Does the ClickHouse news change anything for Langfuse users?

The homepage says the goal is to keep building the platform. The self-hosted stack already used ClickHouse, so the tie is close, but the site doesn't describe changes for existing users.