Auto-instrument an LLM application with Traceloop OpenLLMetry and export OpenTelemetry spans to any OTLP endpoint

Documented steps

1. Install the SDK: pip install traceloop-sdk.
2. Initialize once, early in application startup: from traceloop.sdk import Traceloop; Traceloop.init().
3. In local development call Traceloop.init(disable_batch=True) so spans appear immediately instead of being batched.
4. If your app uses a supported library, no further code is needed. Auto-instrumented providers are OpenAI and Azure OpenAI, Anthropic, Google Generative AI (Gemini), Cohere, Groq, HuggingFace, Mistral AI, Ollama, Replicate, AWS Bedrock, AWS SageMaker, GCP Vertex AI, IBM watsonx.ai, Together AI, Aleph Alpha, and Writer. Vector stores: Chroma, Pinecone, Qdrant, Weaviate, LanceDB, Milvus, Marqo. Frameworks: LangChain, LlamaIndex, LiteLLM, CrewAI, LangGraph, Haystack, Langflow, Agno.
5. Wrap your own orchestration code with the decorators from traceloop.sdk.decorators so business logic is visible alongside model calls: @workflow(name='my_workflow'), @task(name='my_task'), @agent(name='joke_translation'), @tool(name='history_jokes'). The name argument is optional on all four and @workflow also accepts version=.
6. To decorate a class rather than a function, pass method_name to indicate which method runs the unit of work, for example @agent(name='base_joke_generator', method_name='generate_joke').
7. To send data to Traceloop Cloud, set TRACELOOP_API_KEY (generate it at app.traceloop.com under Settings > API keys). No endpoint configuration is needed; the default base URL is https://api.traceloop.com.
8. To send to your own collector or any OTLP backend instead, set TRACELOOP_BASE_URL to the endpoint, for example http://<collector-host>:4318 for OTLP over HTTP.
9. If the destination requires authentication, set TRACELOOP_HEADERS in W3C format, key1=value1,key2=value2. Confirm the exact delimiter syntax against the installed SDK before production use.
10. For exporters other than OTLP HTTP or gRPC, such as Zipkin, pass a configured OpenTelemetry exporter object directly: Traceloop.init(exporter=<your_exporter_instance>).
11. Narrow what gets instrumented with the allowlist and blocklist parameters on init: Traceloop.init(instruments=[...]) or Traceloop.init(block_instruments=[...]). Other documented init parameters are app_name, resource_attributes, and should_enrich_metrics.
12. For PII and compliance, set TRACELOOP_TRACE_CONTENT=false to stop logging prompt, completion, and embedding content. Content logging is ON by default. The TypeScript equivalent is Traceloop.initialize({ traceContent: false }).
13. Disable Traceloop's own anonymous SDK telemetry, which is separate from your trace data, with TRACELOOP_TELEMETRY=false.
14. Deploy and confirm spans arrive at the configured destination before relying on dashboards.
15. When building downstream dashboards, map against the documented attributes: gen_ai.system, gen_ai.request.model, gen_ai.response.model, gen_ai.prompt, gen_ai.completion, gen_ai.usage.prompt_tokens, gen_ai.usage.completion_tokens, gen_ai.usage.total_tokens. OpenLLMetry also emits non-standard extensions llm.request.type (completion or chat), llm.user, traceloop.span.kind (workflow, task, agent, tool), traceloop.workflow.name, and for vector stores db.system and db.vector.query.top_k.