Your LLM app, traced like any other microservice

What it does

OpenLLMetry is a collection of OpenTelemetry instrumentations and an optional SDK that capture telemetry from LLM providers, vector databases, and AI frameworks. It emits standard OTel data, which means traces from an OpenAI request or a Pinecone lookup show up in the same pipeline as your existing database and HTTP instrumentation. The project is maintained by Traceloop and licensed under Apache 2.0.

The interesting bit

The real value is normalization, not novelty. Instead of building yet another LLM-specific observability silo, OpenLLMetry piggybacks on OpenTelemetry’s semantic conventions—which the project helped get adopted officially—to make your LLM infrastructure look like ordinary microservice infrastructure to any compatible backend.

Key highlights

• Supports 15+ LLM providers (OpenAI, Anthropic, Bedrock, Gemini, etc.), 7+ vector DBs (Pinecone, Weaviate, Qdrant, etc.), and frameworks like LangChain, LlamaIndex, and CrewAI.
• Exports to 25+ existing observability platforms including Datadog, Honeycomb, Grafana, New Relic, and Splunk via standard OTel protocols.
• Can be used as standalone instrumentations if you already have OpenTelemetry wired up, or via the Traceloop SDK for a quick start.
• The project’s semantic conventions are now officially part of OpenTelemetry.
• Telemetry collection in the SDK was removed entirely as of v0.49.2.

Caveats

• The README is essentially a feature list and quickstart; it doesn’t discuss performance overhead, sampling strategies, or cost implications of high-cardinality LLM traces.
• If you don’t already use an OpenTelemetry-compatible backend, the “plug into existing tools” pitch is moot—you’ll need to set one up first.

Verdict

Worth a look if you’re running Python LLM workloads in production and want unified tracing without adopting a proprietary agent. Skip it if you’re looking for a hosted, all-in-one LLM evaluation and observability suite; this is plumbing, not a product.