Langfuse vs. LangSmith

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

TL;DR:

• Choose Langfuse if you prioritize open source (MIT), data sovereignty (full self-hosting on an open storage stack), a framework-agnostic approach built on OpenTelemetry, and transparent unit-based pricing.
• 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. LangSmith is a proprietary, closed-source platform: its client SDKs are open source, but the backend, UI, and storage layer are not, and self-hosting is an Enterprise-exclusive add-on.

Scale and storage architecture

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: LangChain built a proprietary database, and Langfuse invested in its data model on top of an open-source one.

In May 2026, LangChain announced SmithDB, a purpose-built trace database written in Rust on Apache DataFusion and the Vortex file format, with trace data on object storage and stateless ingestion and query services. LangChain reports that SmithDB serves LangSmith's US Cloud ingestion and tracing queries and makes core experiences up to 15x faster. As of the May 2026 announcement, SmithDB for self-hosted LangSmith was in early access, so self-hosted deployments do not yet run the same storage engine as their cloud.

Langfuse stores traces in ClickHouse, the open-source columnar database (Apache 2.0). Langfuse is part of ClickHouse and works directly with the database's core team. Rather than building a new engine, Langfuse redesigned its data model: since March 2026, every LLM call, tool execution, and agent step is written once to a single wide, immutable observations table, which eliminates joins and read-time deduplication. The engineering details and benchmarks are public in Simplifying Langfuse for Scale:

• 1. Initial table loads for large data volumes dropped from seconds to tens of milliseconds, and dashboard load times for large projects improved by at least 10x on longer time ranges.
• 2. The redesign carried a 19x year-over-year growth in processed data and a migration of petabytes of historic traces, done in collaboration with ClickHouse engineers, including a fix contributed upstream to ClickHouse itself.
• 3. Token-based full-text search across observation inputs and outputs runs directly on ClickHouse.
• 4. Compute and storage are decoupled: Langfuse's web and worker containers are stateless, and ClickHouse supports object-storage-backed deployments that separate compute from storage, the same architectural property SmithDB cites.

A proprietary purpose-built store is a defensible engineering choice for a closed SaaS. The difference shows up when you operate the platform yourself: with Langfuse, the storage layer is a widely deployed database you can run, inspect, and query with SQL, operational expertise for it is hireable, and the same engine backs Langfuse Cloud and self-hosted deployments today.

Integrations and framework coverage

Both platforms now accept OpenTelemetry traces and instrument non-LangChain applications. The difference is degree: Langfuse's own SDKs are built on OpenTelemetry, while LangSmith's deepest, zero-setup integration remains LangChain and LangGraph, with OTel ingestion alongside its native SDK.

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; LangSmith pairs its evals with dataset tooling and, since May 2026, LangSmith Engine for automated failure clustering.

Monitoring and alerting

Both platforms ship threshold-based alerting on production metrics as of July 2026.

Langfuse Monitors watch metrics over observations and scores (for example average cost per trace, 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 workflows. Monitors are available on Langfuse Cloud.

LangSmith Alerts fire on error rate, run latency, feedback score, or cost over 5- or 15-minute aggregation windows and route to PagerDuty, Dynatrace, or any webhook. Alerts are configured per project.

Pricing

The economic models differ. Langfuse charges for the depth of ingested data (units), while LangSmith charges per seat plus per root execution (traces), with retention priced into the trace rate. Numbers below are as of July 2026.

Open platform and extensibility

Langfuse is API-first, so teams can treat observability data as their own. LangSmith exposes APIs for traces and datasets within its managed platform.

Enterprise security

Both platforms are enterprise-ready with major certifications. Langfuse adds ISO 27001 to SOC 2 Type II and offers a fully air-gapped compliance path through open-source self-hosting.

Feature highlights

Langfuse:

• Agent graph views render a trace as a graph of steps, with an aggregated overview mode and an expanded as-it-ran mode for multi-agent debugging.
• Prompt management provides versioning, deployment labels, and a playground without tying you to one framework's syntax.
• Evaluations combine LLM-as-a-judge, in-platform code evaluators, and human annotation, with experiment gates in CI/CD.
• ClickHouse-native analytics give unrestricted raw SQL access to your data when self-hosting.

LangSmith:

• LangSmith Deployment runs LangGraph agents as managed, durable APIs, complemented by Managed Deep Agents since May 2026.
• Zero-setup tracing instruments LangChain applications through environment variables alone.
• Interrupt 2026 additions include LangSmith Engine (automated failure clustering with proposed fixes), Context Hub (versioned agent instructions), and an LLM Gateway.

Which should you choose

Choose Langfuse if you need to self-host with full feature parity, run a mixed or non-LangChain stack on OpenTelemetry, want your observability data in an open database you can query directly, or prefer usage-based pricing without seat fees.

Choose LangSmith if your team builds primarily on LangChain and LangGraph, wants tracing enabled by environment variables alone, and values a single managed vendor for observability plus agent deployment infrastructure, with self-hosting reserved for Enterprise contracts.

FAQ

Is LangSmith free?

LangSmith has a free Developer plan with one seat and 5,000 base traces per month at 14-day retention (as of July 2026). Beyond that, base traces cost $2.50 per 1,000, extended 400-day retention traces cost $5.00 per 1,000, and the Plus plan is $39 per seat per month. It is free to start but not free software: the platform is proprietary and self-hosting requires an Enterprise contract. Langfuse's free tier includes 50,000 units per month, and the MIT-licensed open-source version is free to self-host without feature gates.

Is LangSmith open source?

No. The LangSmith platform (backend, UI, and storage layer, including SmithDB) is closed source. The LangChain and LangGraph frameworks and the LangSmith client SDKs are open source under MIT, which is a common point of confusion. Langfuse is open source end to end: the full platform is MIT-licensed and self-hostable.

Does LangSmith require LangChain?

No. LangSmith traces applications built without LangChain through its SDKs and accepts OpenTelemetry traces via an OTLP endpoint (as of July 2026). Its lowest-friction path remains LangChain/LangGraph, where tracing activates via environment variables. Langfuse is likewise framework-agnostic, with SDKs built on OpenTelemetry and integrations for 100+ frameworks and model providers.

What is SmithDB?

SmithDB is LangSmith's proprietary trace database, announced in May 2026 and written in Rust on Apache DataFusion and the Vortex file format, with trace data on object storage. LangChain reports it makes core LangSmith experiences up to 15x faster; at announcement, self-hosted availability was early access only. Langfuse achieves its scale characteristics on open-source ClickHouse with an observations-first immutable data model, so cloud and self-hosted deployments run the same engine.

---

---