**LangSmith** is the tracing and evaluation product built by LangChain Inc. It is closed-source, hosted only on LangChain's servers (with a self-hosted enterprise option), and reads its trace data primarily from LangChain's callback handler. Committing to LangSmith is a commitment to LangChain as both a framework and a SaaS vendor.
**Langfuse** is an open-source tracing and evaluation product. The core is MIT-licensed and runs on your own Postgres and ClickHouse instances; there is also a hosted Langfuse Cloud for teams that want a managed offering. Langfuse exposes both a LangChain callback handler AND an OpenTelemetry GenAI exporter, so it ingests traces from any framework or directly from the raw provider SDK.
The two products look similar on the dashboard: spans, prompts, completions, costs, evals, datasets, prompt management. The architecturally important difference is lock-in posture. One pins you to LangChain and to LangChain Inc.'s pricing; the other keeps your exits open. That is the load-bearing distinction, not the storage engine or the sync vs async support.
**The axis that actually matters: lock-in posture.**
LangSmith is closed-source, owned and operated by LangChain Inc., and tightly coupled to LangChain as a framework. Its ingestion path is the LangChain callback handler; non-LangChain code can be instrumented via a manual API but it is a second-class path. Picking LangSmith pins your framework choice (LangChain or LangGraph) and your tracing vendor in one decision.
Langfuse is open-source (MIT-licensed), self-hostable, and built around OpenTelemetry GenAI conventions. The core ingest path is an OTel exporter that ingests traces from any framework (LangChain, LlamaIndex, raw SDK, Vercel AI SDK) or any language with an OTel SDK. The hosted Langfuse Cloud is convenient but not required; you can run the whole stack on your own infrastructure.
**Why the other options are distractor-shaped.** Prompt storage versus hashing is a configuration concern, not an architectural one. Storage engine (Postgres versus ClickHouse) is an implementation detail that any vendor could change. Sync versus async support is feature parity that closes in months.
**The decision implication.** If your team has explicitly decided to use LangChain for the long term and wants the tightest possible integration with the framework, LangSmith is sharper. If your team values portability, polyglot stacks, OTel-native infrastructure, or the option to eject from LangChain later, Langfuse is the structurally right pick.
Quick reference
TL;DR
LangSmith and Langfuse differ on lock-in posture: proprietary + LangChain-coupled versus open-source + OTel-based; that axis dwarfs storage or sync-async differences.
Imagine two doorbell apps. One only works with one brand of doorbell, runs on the company's servers, and you cannot inspect the code; if the company changes its pricing, you change vendors or you pay. The other works with any doorbell, can run on your own machine, and the code is open; if the company changes, you keep running it yourself. The two apps may show you the same notifications today, but they leave you in very different places tomorrow. LangSmith is the first; Langfuse is the second.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Teams that are committed to LangChain long-term and value the tightest possible UI integration LangSmith's deep coupling is a real benefit if the commitment is genuine. Forcing OTel layers in for portability that is never exercised is wasted overhead.
Compliance environments that require traces never leave the company's network LangSmith Cloud is a non-starter; self-hosted LangSmith is enterprise-priced. Langfuse OSS self-host is the cost-effective compliant pick.
Polyglot stacks with Python LangChain and TypeScript Vercel AI SDK both in production LangSmith's coverage of non-LangChain code is second-class; Langfuse via OTel sees both stacks consistently.
Quick scan · 2–4 min
Everything important, quickly.
5 min: licensing and hosting model contrast + OTel GenAI as the architectural axis + why storage and sync/async are distractors + Arize Phoenix as third OSS option + Helicone as a gateway pattern + ejection mechanics.
Real-world examples
Real products, models, and research that use this idea.
1Teams that committed to LangChain plus LangSmith early (2023-2024) report that ejection in 2026 means rewriting both the framework AND the observability stack at once, the cost is real.
2Klarna's customer-support stack picked Langfuse explicitly to keep the OTel exit open as they migrated from LangChain to thin SDK wrappers on critical paths.
3Mastra ships its own first-class OTel GenAI integration in 2026, which makes Langfuse the natural pick for Mastra users.
4Cursor reportedly runs its own internal observability on OTel-native primitives, sidestepping both LangSmith and Langfuse but validating the OTel framing.
5Arize Phoenix is the third major OSS LLM observability product in 2026; it occupies similar ground to Langfuse with a stronger emphasis on RAG-specific evals.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat is the eject from LangSmith story when a team decides to migrate?
Sign in or go Pro to see the approach for this follow-up.
QHow does Arize Phoenix differ from Langfuse if both are OSS OTel-based?
Sign in or go Pro to see the approach for this follow-up.
QWhy does Helicone occupy a different category than either LangSmith or Langfuse?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Picking the vendor on dashboard polish or eval-feature richness without checking the licensing and OTel posture. Today's feature gap closes in months; lock-in posture is permanent.