Claude Platform opens OpenTelemetry tracing beta: Messages API, Managed Agents and SDK spans flow to your own observability stack

Anthropic opened a public beta of OpenTelemetry tracing for Claude Platform on 2026-10-14. There are two layers. Server-side, an admin adds a telemetry destination in the Claude Console or via the Admin API (OTLP/HTTP with protobuf over HTTPS on port 443), and the API exports spans for Messages API requests and Claude Managed Agents sessions with no code changes. Client-side, the Claude SDKs (Python, TypeScript, Go, Java 2.72.0+ and others) emit spans for each request and each tool the tool runner executes. Only server-side tracing shows Managed Agents turns and Anthropic-run tools; only SDK tracing shows tools in your own code or a self-hosted sandbox, so Anthropic recommends enabling both to get one end-to-end trace.

Spans follow the OpenTelemetry gen_ai conventions and default to metadata such as duration, status, token counts, model and tool names; a full-content mode also exports prompts, responses, tool inputs and results. Zero-data-retention customers can still export full content for Messages API requests because the content is held only in memory until delivery, while Managed Agents is not ZDR-eligible. Partner cards in the docs include Datadog, Grafana Cloud, Honeycomb and Langfuse. No performance or overhead benchmarks were published. During the beta, span names and SDK options may change in breaking ways, so pin SDK versions, and when counting tokens sum only the API-side spans for each request to avoid double counting.