Openllmetry

OpenLLMetry instruments LLM applications with standard OpenTelemetry — one Traceloop.init() call, and every LLM invocation (OpenAI, Anthropic, Cohere, Bedrock, etc.) and framework call (LangChain chains, LlamaIndex retrievers, CrewAI agents) emits OTel spans to your existing observability stack. For agents: zero code changes to existing LLM calls after init, export traces to Grafana Tempo, Jaeger, Datadog, Honeycomb, or Traceloop Cloud, add workflow spans with @workflow decorator for multi-step agent traces. Standard OTel spans mean traces integrate with non-LLM infrastructure traces in the same dashboard. Confidence is docs-derived.

Python SDK: pip install opentelemetry-sdk traceloop-sdk. Init: from traceloop.sdk import Traceloop; Traceloop.init(app_name="my-agent", api_key="tlp_...") — all LLM calls auto-traced from this point. Manual spans: from traceloop.sdk.decorators import workflow, task; @workflow(name="rag-pipeline") def rag(query): retriever_result = retrieve(query); return generate(retriever_result). @task(name="retrieve") def retrieve(query): ... Auto-instrumentation list: OpenAI, Anthropic, Azure OpenAI, Cohere, Bedrock, Mistral, HuggingFace, LangChain, LlamaIndex, CrewAI, Haystack, DSPy, Transformers, ChromaDB, Qdrant, Pinecone, Weaviate. OTel export: OTEL_EXPORTER_OTLP_ENDPOINT env var → any OTel-compatible backend. Traceloop Cloud: default export destination when api_key set. JS SDK: npm install @traceloop/node-server-sdk.

Traceloop API key auth (for Cloud export): api_key parameter in Traceloop.init() or TRACELOOP_API_KEY env var. Keys from app.traceloop.com → Settings. HTTPS enforced. Self-hosted OTel export: set OTEL_EXPORTER_OTLP_ENDPOINT to your OTel collector — no Traceloop API key needed. No OAuth2. Data: prompts/completions exported to configured OTel backend — for Traceloop Cloud, review data policy. Prompt redaction: configure disable_batch=True or custom span processors to redact sensitive prompt content before export.

SDK overhead: OTel instrumentation adds ~1–3ms per traced call. Async-safe: SDK works in async Python. Batch export: spans buffered and exported in batches — configurable export interval. Export failures: OTel SDK retries on network errors; does not block LLM calls on export failure. LLM token counts: extracted from LLM API responses and added as OTel span attributes (llm.usage.prompt_tokens, llm.usage.completion_tokens). Framework instrumentation: LangChain/LlamaIndex spans include retrieval context and document metadata. New providers: community-contributed instrumentors; check GitHub for latest support list.

www.traceloop.com/docs covers SDK reference, auto-instrumentation list, OTel export configuration, and Traceloop Cloud setup. Getting started: pip install traceloop-sdk, Traceloop.init() — first LLM trace in 2 minutes. Traceloop Cloud: free tier (1M spans/month). Self-hosted: point to any OTel backend. GitHub: traceloop/openllmetry (4k+ stars). Community via Slack (active). Good documentation; OTel export guide is particularly thorough for teams with existing OTel infrastructure.