Anthropic Adds OpenTelemetry Tracing to Claude

Anthropic has launched a public beta adding OpenTelemetry tracing support to the Claude API, Managed Agents, and official SDKs. By exporting telemetry via standard OTLP endpoints, developers can monitor model requests and tool usage within existing observability tools, including Datadog, Honeycomb, Grafana Cloud, Splunk, Arize AI, Braintrust, Sentry, and Langfuse. Self-hosted OpenTelemetry Collectors can also accept exports via HTTPS on port 443 using protobuf encoding.

The system helps engineers identify latency bottlenecks and execution failures across complex AI agent workflows. Spans follow standard OpenTelemetry gen_ai conventions. In the default metadata-only mode, traces collect operational data including duration, status, token counts, model names, and tool names. Switching to full-content mode logs entire prompts, responses, tool inputs, and tool outputs, constrained by a default max content size of 0.75 MB per span.

To trace requests end-to-end, the Claude SDK injects W3C traceparent headers into outgoing API calls. This context propagation extends through Anthropic's Model Context Protocol connector, allowing a single distributed trace to span application code, Claude model execution, and downstream tool interactions. Developers can also pass a conversation_id string across multi-turn exchanges to group related agent interactions together.

The feature is offered at no additional charge from Anthropic, though normal backend ingest and storage costs apply. Platform limits constrain organizations to 10 OTLP destinations and 1,000 selected workspaces per destination. For software engineers, native OpenTelemetry integration simplifies agent debugging by bringing model-level execution metrics into unified operational dashboards without site-level code modifications.