Instrument an LLM or agent application with LangSmith tracing (SDK wrappers, @traceable, or OpenTelemetry ingestion)

Documented steps

1. Create a LangSmith API key in Settings at smith.langchain.com. Note your deployment region, because the ingest host differs: GCP US (default) https://api.smith.langchain.com, GCP EU https://eu.api.smith.langchain.com, GCP APAC https://apac.api.smith.langchain.com, AWS US https://aws.api.smith.langchain.com, self-hosted https://<your-domain>/api/v1.
2. Set environment variables: LANGSMITH_TRACING=true, LANGSMITH_API_KEY=<key>, LANGSMITH_PROJECT=<project name>, and LANGSMITH_ENDPOINT=<region URL> if you are not on the default GCP US host. Multi-workspace keys additionally accept LANGSMITH_WORKSPACE_ID.
3. Install the SDK. Python: pip install -U langsmith openai (add anthropic if you trace Anthropic). TypeScript: npm install langsmith openai (plus @anthropic-ai/sdk for Anthropic).
4. Wrap the OpenAI client so every model call becomes a traced run. Python: from langsmith.wrappers import wrap_openai; client = wrap_openai(OpenAI()). TypeScript: import { wrapOpenAI } from 'langsmith/wrappers'; const client = wrapOpenAI(new OpenAI()).
5. For Anthropic use wrap_anthropic(anthropic.Anthropic()) in Python, or import { wrapAnthropic } from 'langsmith/wrappers/anthropic' in TypeScript.
6. Decorate your own functions, tools and chain steps so they appear as named spans. Python: @traceable(run_type='tool') above the function. TypeScript: traceable(async function fn(){...}, { run_type: 'tool' }).
7. Run the app and confirm traces land in the target project in the LangSmith UI before adding further instrumentation.
8. If you need explicit control over trace structure outside the decorator flow, use the RunTree class directly: construct the root run with RunTree(name=..., run_type=..., inputs=...), call .post(), create children with .create_child(...), then .end(outputs=...) and .patch(). RunTree only needs LANGSMITH_API_KEY; it does not require LANGSMITH_TRACING.
9. For non-LangChain stacks or other languages, use the native OpenTelemetry path instead: set OTEL_EXPORTER_OTLP_ENDPOINT=https://api.smith.langchain.com/otel and OTEL_EXPORTER_OTLP_HEADERS='x-api-key=<key>' (optionally append ,Langsmith-Project=<name>). Install opentelemetry-sdk and opentelemetry-exporter-otlp and export with BatchSpanProcessor(OTLPSpanExporter(timeout=10)). Self-hosted OTel endpoint is https://<your-domain>/api/v1/otel.
10. To make a LangChain or LangGraph app route its own tracing through OpenTelemetry internally, add LANGSMITH_OTEL_ENABLED=true alongside the standard LANGSMITH_* variables.
11. In short-lived processes (CLI scripts, Lambda, serverless handlers) call the SDK flush() method before exit, and consider LANGCHAIN_CALLBACKS_BACKGROUND=false so callbacks run synchronously. Leave it true in long-running services to keep request latency low.
12. Only fall back to the REST API (POST /runs, PATCH /runs/{run_id}, POST /runs/multipart for batched ingest, header x-api-key and optional x-tenant-id) if no SDK is available for your language. The docs explicitly recommend the SDKs instead because synchronous REST calls add latency to your app.