LangSmith and Langfuse differ on lock-in posture: proprietary and hosted (self-host only on Enterprise) versus open-source and self-hostable. That matters more than storage or sync-async differences.
Imagine two doorbell apps. One runs on the company's servers (unless you buy its top plan), and you cannot inspect the code; if the company changes its pricing, you change vendors or you pay. The other 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.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Observability vendor selection looks like a tool comparison and turns out to be an ecosystem commitment. LangSmith and Langfuse cover the same surface area (spans, prompts, completions, costs, evals, datasets, prompt management) and look indistinguishable on a feature checklist. The architecturally important difference is invisible on the checklist: it lives in the licensing and the hosting model.
Both accept OpenTelemetry traces, so ingest is not where they split. The split is who runs the backend and under what licence, and that decides how hard it is to leave.
LangSmith: closed, hosted, LangChain-native
LangSmith is the tracing and evaluation product built and operated by LangChain. The core service is closed-source and hosted on LangChain's infrastructure, with self-hosting available as an add-on to the Enterprise plan. For LangChain and LangGraph apps, tracing is automatic: set LANGSMITH_TRACING=true and LANGSMITH_API_KEY, and every Runnable, LLM call, tool call, and agent step shows up as a run in the LangSmith UI.
Non-LangChain code can be instrumented via LangSmith's @traceable decorator (Python) or traceable() wrapper (TypeScript), the trace context manager, or the RunTree API; nested calls produce nested runs. LangSmith also accepts OpenTelemetry traces at its OTLP endpoint, and its docs list integrations for OpenAI, Anthropic, CrewAI, Vercel AI SDK and Pydantic AI. LangSmith does not tie you to LangChain as a framework.
The commercial story is the commitment. LangSmith pricing scales with trace volume, and trace history leaves through the API or bulk export. A team that has built six months of trace history, evals, and prompt experiments in LangSmith and then decides to migrate faces real work: export historical data, re-instrument code with a different vendor's SDK, rebuild dashboards. The work is bounded but not trivial.
None of this makes LangSmith a bad pick. It is the best fit for teams that are committed to LangChain and want the tightest possible integration. It is a weaker pick for teams that need self-hosting without an Enterprise contract or open licensing, because every commitment adds to the eventual cost of leaving.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- A team that leaves LangSmith must export its trace history (API or bulk export) and re-instrument, even if it keeps LangChain.
- A team moving from LangChain to thin SDK wrappers on critical paths can keep the same Langfuse instance, because Langfuse accepts OTel traces from any code.
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?
Two parts: (1) export historical traces from LangSmith (their API supports this but it is a project, not one-shot); (2) re-instrument code with the new vendor's SDK or OTel exporter, which means touching every chain config call site. The historical-data part is what makes the migration heavy; the re-instrumentation is bounded.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Picking the vendor on dashboard polish or eval-feature richness without checking the licensing and hosting model. A feature gap can close in months; the licensing and hosting model rarely changes.
60 second bullets to scan on the way to the call.
LangSmith's licensing and hosting model
Langfuse's licensing and hosting model
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.