Langfuse vs. LangSmith

This guide outlines the key differences between Langfuse and LangSmith to help engineering teams choose the right AI engineering platform. All LangSmith facts were checked against public LangChain sources in August 2026.

TL;DR

Choose Langfuse if you want:

1. Control without rebuilding governance. MIT-licensed self-hosting on every tier, including free, with SSO and org RBAC in the OSS distribution, and a framework-agnostic approach built on OpenTelemetry.

2. Data sovereignty. An open-source ClickHouse data plane you can query, join, and hire for. Cloud and self-hosted share the same engine.

3. One usage meter and no seat fees. Traces, observations, and scores bill as units. Langfuse Core and higher plans include unlimited users.

Choose LangSmith if you are an all-in LangChain/LangGraph shop that wants a managed SaaS covering observability, evals, and agent deployment infrastructure in one closed platform.

---

Open source and distribution

Langfuse is open source (MIT) and self-hosting is a first-class deployment mode. Tracing, evaluations, prompt management, experiments, annotation, and playground ship in the MIT distribution with no usage limits. LangSmith is proprietary: its client SDKs are open source, but the backend, UI, and storage layer are not, and self-hosting is an Enterprise-exclusive add-on.

Self-hosting runs the same product codebase and architecture as Langfuse Cloud, so trace handling, tenant boundaries, and access controls do not require rebuilding in-house.

On Langfuse Cloud, Enterprise SSO and project-level RBAC sit behind the Teams add-on or Enterprise. On self-hosted OSS, Enterprise SSO and organization-level RBAC ship in the free distribution.

---

Storage: ClickHouse vs SmithDB

Both platforms were built for high-volume trace data, and both re-architected their storage layers as agent traces grew into thousands of spans with large payloads. They made different choices:

LangSmith's storage engine was ClickHouse. In May 2026, LangChain replaced it on their US Cloud with SmithDB, a proprietary Rust engine on Apache DataFusion and Vortex. Self-hosted LangSmith still runs on ClickHouse and Postgres; there, SmithDB is a public beta: LangChain does not recommend operating it yourself and routes access through a waitlist.

Langfuse runs on open-source ClickHouse and works directly with the database's core team as part of ClickHouse. Every LLM call, tool execution, and agent step is written once to a wide, immutable observations table. Details are in Simplifying Langfuse for Scale.

A purpose-built store is a reasonable choice for closed SaaS. If you operate the platform yourself, ClickHouse is a database you can run, inspect, and query with SQL. High-volume reads go through the Observations and Metrics APIs.

---

Pricing

Langfuse meters traces, observations, and scores as units, with no seat fees. LangSmith meters seats, traces, and a retention upgrade, plus separate lines for Deployment, Engine, Fleet, Sandboxes, and LLM Gateway on their pricing page. The worked example covers observability seats and traces only.

Worked example

500,000 traces/month · 10 spans per trace · 1 score per trace · 12-month retention · 5 users.

Public list prices. Edit the inputs yourself: Langfuse vs LangSmith pricing model

1. LangSmith bills seats on invitations. A ten-person team is $390/month before a single trace.

2. LangSmith's included window is 14 days. Keeping traces longer is a per-trace upgrade, and online evaluators and automation rules promote traces into the paid tier by default, though each evaluator or rule can opt out. Hitting a self-set extended-trace limit blocks further retention-extending evaluators and rules; UI feedback and annotation-queue adds keep working.

3. We bill every span; they bill the trace. We are likely cheaper on shallow traces. The gap narrows as traces deepen. The sheet takes spans-per-trace as an input so you can find the crossover for your workload.

See pricing · Start Free · Talk to us

---

Integrations and framework coverage

Both platforms now accept OpenTelemetry traces and instrument non-LangChain applications. LangSmith's advantage is integration with LangChain and LangGraph through one environment variable, and the same vendor can host the runtime, gateway, and agent builder.

Langfuse has no framework to protect and no runtime to sell. It integrates with 100+ frameworks and model providers (OpenAI, Anthropic, Vercel AI SDK, LangChain, LlamaIndex, LiteLLM, and others) or your own harness. If you change vendors later, traces, evaluators, datasets, and prompts stay in a database you can operate.

The suite-adjacent pieces exist in the open ecosystem too: the ClickHouse-built Agentic Data Stack pairs LibreChat's chat UI with MCP tools and Langfuse tracing in one Docker Compose deployment, and a Langfuse Gateway is on the public roadmap.

---

Evaluation and experiments

Both platforms cover offline and online evaluation. Langfuse runs LLM-as-a-judge and deterministic code evaluators inside the platform and gates CI/CD pipelines on experiment results. Online scoring does not auto-promote traces into a paid retention tier the way LangSmith online evaluators do by default. LangSmith pairs its evals with dataset tooling, Tuned Evaluators (public beta on Plus and Cloud Enterprise, US region), and LangSmith Engine, which clusters production failures on a 6-hour schedule and is metered in LCUs.

---

Observability and monitoring

Both platforms cover trace inspection in the UI and threshold-based alerting on production metrics.

Langfuse records cost on generations, groups conversations into sessions, and renders agent graphs from the trace. Custom dashboards and Pulse sit on the same observations table. Alerts watch metrics over observations and scores (for example observation counts, average cost, p95 latency, or an evaluation score), support separate warning and alert thresholds plus no-data handling, and route notifications to Slack, webhooks, or GitHub Actions. Alerts are available on Langfuse Cloud and self-hosted v4+.

LangSmith Alerts fire on run count, error rate, run latency, feedback score, or cost over 5- or 15-minute aggregation windows and route to Slack (native integration, Cloud only), PagerDuty, Dynatrace, or any webhook.

---

Open platform, extensibility, and agentic access

Langfuse is API-first, so teams can treat observability data as their own — and so can their agents. Both platforms ship an agent-facing surface: Langfuse has a CLI built for coding agents, a native MCP server, and an agent skill; LangSmith has an agent-first CLI and a hosted Remote MCP server.

The differences are depth and availability. The Langfuse skill follows the open Agent Skills standard and packages best-practice workflows, so a coding agent can instrument your app, analyze production traces, build datasets, and run evaluations without leaving the editor. Because the CLI and MCP server are part of the OSS product, agentic access works on any self-hosted instance, including the free tier. On Cloud, the in-product Langfuse Assistant answers questions about project data in plain language.

---

Enterprise security

Both products publish SOC 2 Type II, ISO 27001, GDPR, and HIPAA. Langfuse signs a Business Associate Agreement (BAA) from the Pro plan in the dedicated HIPAA Cloud region. Langfuse is used by 21 of the Fortune 50 and 129 of the Fortune 500. Air-gapped operation and self-hosted SSO/RBAC are in Open source and distribution.

---

Which should you choose

Choose Langfuse if you want an independent, open-source platform that adapts to your stack: OpenTelemetry-based instrumentation for any framework, MIT self-hosting with SSO and org RBAC included, observability data in an open database you can query directly, and unit pricing with no seat fees.

Choose LangSmith if you are committed to LangChain and LangGraph and want one vendor to host the agent runtime and observe it: zero-config instrumentation and managed deployment for their frameworks, sandboxes, a no-code agent builder (Fleet), and an LLM gateway in public beta.

---

Why teams move away from LangSmith

The same reasons recur in migration conversations. Cost leads, and it is several distinct problems:

• Trace-volume pricing. LangSmith's per-trace usage grows linearly with traffic, so a bill that works in a pilot becomes the blocker at production scale. At 500k traces with 5 users, the worked example lands at $3,895/month on LangSmith Plus vs $621 on Langfuse Pro.
• Seat fees tax adoption. LangSmith bills $39/seat on invitation, so giving product managers, reviewers, and domain experts view access has a per-head price. Langfuse meters usage, not people.
• Retention is its own meter. LangSmith's included window is 14 days, keeping traces longer is a per-trace upgrade, and online evaluators and automation rules promote traces into the paid tier by default — $1,250 of the worked example's monthly LangSmith total is retention.
• Compliance routes through sales. Self-hosting, custom SSO, and role-based access control are Enterprise-only on LangSmith, so teams that must run in their own infrastructure cannot evaluate on the free or Plus tiers. Langfuse self-hosts MIT-licensed on every tier and signs a BAA from the Pro plan.
• Framework coupling. Platform teams serving heterogeneous stacks don't want the observability layer tied to one agent framework. Off LangChain and LangGraph, LangSmith's zero-setup instrumentation is gone, and Deployment, Engine, Fleet, and the Gateway are priced separately on top.
• Turning trace data into answers. LangSmith captures the traces, but slicing them into business views or moving them into your own analytics stack hits walls: bulk export is Enterprise-gated and there is no direct database access on LangSmith Cloud. Langfuse pairs custom dashboards and Pulse with the Observations and Metrics APIs for high-volume reads on Cloud, and self-hosters can additionally query the ClickHouse table directly in SQL.

---

Why teams choose Langfuse

A pattern repeats across production adopters: start where governance requires, often self-hosted, then grow into the platform.

• SumUp saved 30% of external BPO cost by deflecting 50% of support conversations to AI (Ana Casado). They started self-hosted for the PoC, then moved to Langfuse Cloud for production.
• Canva runs framework-agnostic raw LLM clients: Java via OpenTelemetry, Python via the Langfuse SDK. Non-technical users contribute domain expertise through judges. They self-hosted first, then moved to Cloud.
• Merck runs about 80 GenAI use cases on Langfuse with 200+ people building on it. They self-host for data sovereignty, provision via API, and cited the public roadmap and shipping velocity in the selection.
• Khan Academy built a custom Go client on the open API. 100+ users across 7 product and 4 infrastructure teams use it.

---

Switching from LangSmith

Teams that switch rarely migrate everything. They move live instrumentation and the durable assets, not their trace history. Most see their first traces in Langfuse the same day.

• Instrumentation: point your OpenTelemetry exporter at the OTLP endpoint, or swap the LangSmith environment variables for the Langfuse integration in LangChain and LangGraph apps. Both platforms can trace the same app side by side during the transition.
• Datasets, prompts, and evaluators: import datasets via CSV, SDK, or API and recreate prompts and judges in prompt management and evaluators. A coding agent with the agent skill can do this for you.
• Trace history: stays readable in LangSmith for its retention window while live data builds up in Langfuse from day one.

Talk to us if you want help planning a migration.

---

Start free: Cloud or self-host

Start on Langfuse Cloud (Hobby: 50k units/mo, no credit card) or self-host the MIT-licensed stack. You can explore a populated example project before instrumenting anything. Already on LangSmith? See Switching from LangSmith.

---

FAQ

Is Langfuse the best LangSmith alternative?

Yes, if you want an independent open-source platform to trace, evaluate, and improve agents without adopting LangChain's runtime suite. See Which should you choose.

Does Langfuse cover the same product areas as LangSmith?

For the AI engineering workflow, yes. Tracing, prompt management, evaluation, and experiments are natively integrated in one Langfuse project, from prototype to production scale. What Langfuse deliberately does not do is host or run your agents; it works with any runtime you choose instead.

How does LangSmith pricing compare to Langfuse?

LangSmith Plus is $39/seat/month with 10,000 base traces included, plus per-trace usage and optional retention upgrades (as of August 2026). Langfuse Cloud Hobby includes 50,000 units/month, Core starts at $29/month, and the MIT distribution is free to self-host. On the rates in our editable public model, 500k traces / 10 spans / 1 score / 5 users / 12-month retention is $621 Langfuse Pro vs $3,895 LangSmith Plus. We bill every span; they bill the trace, so span depth changes the crossover. Details in the pricing section.

Does Langfuse have production alerting?

Yes. Langfuse Alerts watch any observation or score metric with filters and aggregations, support separate warning and alert thresholds with no-data handling, and deliver to Slack, webhooks, or GitHub Actions. They are available on Langfuse Cloud and self-hosted v4+. Third-party comparisons claiming Langfuse has no alerting are out of date.

Does Langfuse run deterministic evaluators on production traffic?

Yes. Code evaluators run custom Python or TypeScript logic on live observations and on experiments, and they are generally available. Self-hosted deployments configure a code evaluator dispatcher.

Is LangSmith open source?

No. The LangSmith platform (backend, UI, storage, including SmithDB) is closed source. LangChain, LangGraph, and the LangSmith client SDKs are MIT open source. Langfuse is MIT and self-hostable.

Does LangSmith require LangChain?

No. LangSmith accepts OpenTelemetry and non-LangChain SDKs (docs). The lowest-friction path is still LangChain and LangGraph. Managed Deep Agents, Fleet, and LangSmith Deployment are scoped to their suite.

Can I self-host LangSmith?

On the Enterprise plan. As of LangSmith Self-Hosted v0.16, SmithDB, Engine, and Sandboxes are available to self-hosted customers, with SmithDB in public beta behind a waitlist. The LLM Gateway is not in the v0.16 stable release; self-hosted access is by request.

How do I migrate from LangSmith to Langfuse?

Point your OpenTelemetry exporter at the Langfuse OTLP endpoint, or swap the LangSmith environment variables for the Langfuse integration in LangChain and LangGraph apps. Import datasets via CSV, SDK, or API, and recreate prompts and evaluators. Most teams see first traces the same day; trace history stays in LangSmith. See Switching from LangSmith.

What is SmithDB?

SmithDB is LangSmith's proprietary trace database (May 2026), written in Rust on Apache DataFusion and Vortex, with trace data on object storage. LangChain reports up to 15× faster core experiences. Langfuse runs on open-source ClickHouse with an observations-first data model, so Cloud and self-hosted share the same architecture. See Storage: ClickHouse vs SmithDB.

---

---