The OpenTelemetry GenAI Semantic Conventions define spans, metrics and events for generative AI — LLM client inference, agents, tool execution, embeddings, retrieval, evaluation and Model Context Protocol operations — plus provider-specific conventions such as OpenAI. They were split out of the main semantic conventions into their own repository, created 5 May 2026, and depend on the core conventions through the Weaver tool. The documents carry Development status and the repository has no releases yet.
OpenTelemetry GenAI Semantic Conventions
OpenTelemetry GenAI Semantic Conventions name what an AI workload does when it is traced: the model
call, the agent step, the tool invocation, the retrieval, the MCP request. They give gen_ai.*, mcp.*
and openai.* attribute namespaces a shared meaning so AI telemetry from different frameworks and vendors
lines up.
- Own repository - Split out of the main semantic conventions into a repository created on 5 May 2026; shared, non-GenAI conventions stay upstream and are linked at a pinned version.
- Scope - Client inference, agents, tool execution, evaluation, embeddings, retrieval and Model Context Protocol operations, plus provider-specific conventions.
- Weaver-managed - Dependencies on the core conventions are managed with OpenTelemetry Weaver, and the
docs are generated from YAML in
model/. - Reference implementations - A
reference/tree holds implementations and a Python compliance matrix with per-signal support reports. - Development status - The GenAI document is marked Development; the schema URL is still a TODO in the README and no release has been cut. Apache 2.0.
Where it sits. This is OpenTelemetry applied to the AI layer. Its MCP conventions make Model Context Protocol traffic observable in the same traces as the HTTP calls underneath, and its agent and tool spans describe the runtime that specs like Agent2Agent coordinate.
