OpenTelemetry in Claude: Anthropic Adds Native Traces
Anthropic rolled out native OpenTelemetry traces in Claude, exporting agent latency, tool calls, and token metrics directly to enterprise OTLP collectors.
New capability streams execution latency, tool calls, and token metrics from Claude Managed Agents and the Messages API to enterprise OTLP collectors.
Anthropic rolled out native support for OpenTelemetry in Claude, introducing structured trace collection across Claude Managed Agents, the Messages API, and official client SDKs. Documented by the provider on October 14, 2026, the update allows engineering teams to export runtime telemetry directly to enterprise monitoring suites or self-hosted collectors over standard OTLP/HTTP.
Client collection and direct server export
The telemetry architecture operates across two layers. On the application side, official SDKs emit spans for client-driven API requests and tool executions. On the cloud backend, Anthropic servers handle automated export for Claude Managed Agents and Messages API sessions, delivering server-side traces without requiring changes to client application code.
Telemetry destinations and console setup
To route outgoing data, Anthropic introduced the enterprise Telemetry Destinations API under the telemetry-destinations-2026-08-11 beta header. Organizations can configure up to ten distinct OTLP/HTTP endpoints, filter exports by workspace, and verify connectivity by dispatching synthetic test spans labeled anthropic.test.
Teams can also manage these integrations through a graphical interface in the Claude Console. Located at /settings/telemetry-destinations, the settings panel lets administrators register third-party endpoints and monitor transport health through a dedicated delivery status column.
Metadata, tools, and privacy controls
Each span recorded by the platform captures execution latency, HTTP status codes, target model identifiers, input and output token consumption, and the specific tool names executed by the agent.
To help organizations meet corporate governance and privacy requirements, Anthropic introduced two data exposure modes. The metadata-only setting (metadata_only) allows teams to track latency and token spend without transmitting prompts or message bodies to observability platforms. When internal compliance policies require complete oversight, teams can enable full mirroring to capture complete input and completion text.
Why OpenTelemetry in Claude matters for agents
Moving from isolated prompts to autonomous agent systems changes how teams debug AI workloads. In agentic workflows, Claude executes multi-step decision loops, inspects local files, and triggers external tools asynchronously. Without standardized distributed tracing, diagnosing latency spikes in intermediate steps, runaway loops, or unexpected token spikes proved difficult.
By adopting OpenTelemetry, a vendor-neutral standard governed by the Cloud Native Computing Foundation (CNCF), Anthropic avoids proprietary logging silos. Site reliability engineering (SRE) teams can route Claude traces into existing observability stacks such as Grafana, Splunk, Databricks, and LangSmith, viewing agent workflows right alongside standard backend services.
Beta limits and ingestion costs
Server-side export is currently available for organization-level accounts, with an initial quota of ten destinations per organization. Anthropic has not announced extra charges for emitting OTLP packets, though standard ingestion fees apply depending on contracts with downstream observability vendors. The provider has not shared a target date for promoting the destinations API from beta to general availability.