// the find
traceloop/openllmetry
Open-source observability for your GenAI or LLM application, based on OpenTelemetry
OpenLLMetry adds OpenTelemetry instrumentation to LLM providers (OpenAI, Anthropic, Bedrock, etc.) and vector DBs, so your existing observability stack gets LLM-aware traces without a separate tool. It's for teams already running Datadog/Honeycomb/Grafana who want LLM call spans to appear in the same traces as their DB queries and HTTP calls. The semantic conventions it pioneered are now being standardized upstream in OpenTelemetry itself.
Built on real OpenTelemetry primitives — spans, attributes, exporters — so you're not locked into Traceloop's SaaS and can ship traces anywhere OTLP goes. Coverage is unusually wide: 15+ LLM providers, 7 vector DBs, and most of the major agent frameworks (LangChain, LlamaIndex, CrewAI, OpenAI Agents). Each provider is its own installable package, so you don't pull in instrumentation for things you don't use. Tests use recorded HTTP cassettes rather than live API calls, which makes the CI actually reproducible when providers change their response shapes.
Each package is versioned independently and the cassettes quickly go stale when providers ship API changes — there are already open issues about broken instrumentation after provider updates. The monorepo uses Nx but each package has its own lockfile, which creates dependency drift across packages and makes coordinated upgrades painful. Cost tracking (token counts, model pricing) is surfaced as span attributes but there's no built-in aggregation or alerting — you have to wire that up yourself in your observability backend, which varies wildly by destination. The `traceloop-sdk` wrapper adds a layer of magic on top of bare OpenTelemetry that makes debugging initialization failures harder than it should be.