LangSmith Documentation

What LangSmith does

LangSmith is a platform for seeing and measuring what your LLM application does. For agents that means every run recorded as a trace, with each model call, tool call, input, output, token count and latency, plus the tools to turn those traces into evaluations and monitoring.

The documentation covers:

• Observability: tracing from Python or TypeScript, dashboards and alerts, and automatic detection of recurring issues in traces.
• Evaluation: datasets of test cases, offline experiments comparing prompt or model versions, evaluators including LLM-as-judge, and online evaluation of live traffic.
• Feedback and annotation: collecting user feedback and reviewing traces in queues.
• Prompt engineering: a playground and prompt versioning.
• Deployment: running agents built with LangGraph.
• Integrations with OpenAI, Anthropic, CrewAI, the Vercel AI SDK, Pydantic AI and others, plus cloud, hybrid and self-hosted options.

Do I need LangChain to use it?

No. LangSmith is framework-agnostic: you can trace a hand-written agent loop with its SDK, or send traces from other frameworks through its integrations. It is naturally the smoothest with LangGraph and LangChain, where tracing is close to automatic.

Is it worth it for a student or small project?

The free Developer plan, one seat with up to 5,000 base traces a month and pay-as-you-go beyond that according to LangChain's pricing page, is enough to learn tracing and evals on a portfolio agent. For a team, compare it with open source options such as Langfuse, which you can self-host. Whichever you pick, the skills carry over: reading traces, building datasets from real failures and running experiments before shipping a change.

Prerequisites and time

A working agent or LLM app and a LangSmith account. The first trace takes under an hour; a useful evaluation workflow takes a few days of building datasets and evaluators.

Free or paid?

Free Developer plan as above. The Plus plan is listed at $39 per seat a month with up to 10,000 base traces, with usage-based charges above the included amount. Check the pricing page for current details before committing a team.

What you can do after it

Trace every step of an agent, find the failing step in a bad run, build a dataset from real traces, compare two versions in an experiment and set up alerts on quality and cost.

Common questions

Is my data sent to LangChain's cloud? On the cloud plans, yes; hybrid and self-hosted deployments are available for stricter requirements. Avoid sending personal data in traces either way.