Anthropic Adds OpenTelemetry Trace Export to the Claude Platform

Anthropic has added OpenTelemetry trace export to the Claude Platform, announcing on October 14, 2026 that developers can now send traces from Claude Managed Agents, the Messages API, and Claude SDKs to any observability tool that accepts the standard, with the feature available in public beta.

In Anthropic’s announcement, the company said that adding a destination in the Claude Console sends every agent turn, model call, and tool use to the user’s observability tool as a standard OpenTelemetry trace carrying latency, token counts, and IDs. Managed Agents and the Messages API require that Console destination, while Claude SDKs export traces automatically. Developers who already send Claude Code or Cowork traces to an observability tool can have agents built on the Claude Platform appear in the same place.

Anthropic said Claude traces follow the OpenTelemetry GenAI semantic conventions, so data from Claude and other providers works in the same tools and dashboards, and the traces also work with any OpenTelemetry tool that accepts a standard OTLP endpoint.

Two Setup Paths and One Trace Model

The Claude Platform documentation defines a trace as the full record of one request through a system and a span as one step in a trace, such as a model request or a tool use. It describes two setup paths. SDK tracing is configured in code by installing the SDK and registering a tracer provider, and it requires no special permission. API tracing is configured by adding a destination in the Claude Console, and it requires Admin permission.

Coverage differs between the two paths. API tracing captures each model run within a request, Managed Agents turns, and server tools run by the API. SDK tracing adds client tools run by the tool runner, tools in a self-hosted sandbox, and other API requests such as file uploads. Setting up both places a request’s SDK and API spans in a single trace; for requests to Claude Platform on AWS, the API spans appear in their own traces instead.

For Managed Agents, a session produces a span for each Sessions API request and a turn span for each agent turn, and each turn contains spans for each model request, each tool the API runs, and each outcome evaluation. The documentation directs Claude Code, Agent SDK, and Cowork users to separate monitoring and observability documentation for those products.

Content Modes, Limits, and Compliance Behavior

Each destination and each SDK client has a content mode that controls whether spans include model prompts and responses. Metadata only, the default, records identifiers, timing, status, token counts, and text the user chooses, such as the names and descriptions of agents and tools and the names of MCP servers. Full content adds model prompts and responses.

Under Full content, each span holds up to 0.75 MB of content by default, which the documentation states fits a 200k-token context window of English text and stays below the 1 MB request limit of many observability platforms and proxies; a full 1M-token context window needs about 4 MB. When content exceeds the limit, the SDK or the API shortens it and sets the anthropic.content.truncated attribute to true. Request spans drop the system prompt and the earliest messages first, and images and documents appear only as their type.

Organizations with a HIPAA, FERPA, or K-12 data-handling restriction with Anthropic export only metadata from destinations regardless of the configured content mode, though SDK spans are unaffected. Under zero data retention, Full content remains usable because the API holds exported content only in memory until it delivers it to the destination. Claude Managed Agents is not eligible for zero data retention, and its retention terms are unchanged when OpenTelemetry is in use.

For token accounting, the documentation instructs users to count each token once by summing spans from only one source per request, using the API spans when a request has both SDK and API spans, and it notes that spans can arrive more than once when a delivery is retried.

Named Integrations and Early Customer Statements

The announcement groups supported observability platforms by use case. For debugging agents alongside the rest of a software stack, it names Grafana Cloud, Honeycomb, Sentry, and Splunk, where agent traces sit beside logs and metrics so developers can see which service a slow or failed step came from. For improving agents with production data, it names Arize AI, Braintrust, Datadog, Langfuse, LangSmith by LangChain, New Relic, and PostHog, describing uses such as turning real traces into evals and catching quality issues. For tracking usage across an organization, it names Databricks Unity Gateway, where platform and security teams can see token usage by team and flag unusual activity.

Several early customers described their deployments in the announcement. Foxglove co-founder and CTO Roman Shtylman said the company builds agents that help robotics teams debug their robots, and that native OpenTelemetry export from the Claude Platform lets it analyze agent interactions in PostHog alongside product usage data so it can evaluate agent performance in practice and prioritize improvements for robotics teams.

Ramp senior software engineer David Traina said the company built Reflect, an agentic system that scores Langfuse traces, finds recurring failures, and proposes fixes. “On some agents, we’ve already cut token usage by 10–20% and made sessions 30% faster,” Traina said, adding that native OpenTelemetry export from the Claude API can feed instrumentation directly into that improvement loop with no extra steps.

Sterling Labs CTO Hyon Lee said native export streams every trace into Langfuse with no custom instrumentation for the firm’s agents analyzing live markets. FloQast head of infrastructure Ankush Sharma said the accounting software company’s agents run on the Claude Platform with telemetry in Grafana Cloud, putting each model call in the same trace as the reconciliation or journal-entry workflow it ran inside, so the company can answer auditor questions about what an agent did from one timeline.