ClaudeDevs announced on October 14, 2026, that Claude Platform OpenTelemetry tracing had entered public beta, covering Messages API requests, Claude Managed Agents sessions and Claude SDKs. The two routes serve different needs: SDK tracing records activity in your application, while a Console destination adds traces from Anthropic’s service. ClaudeDevs’ announcement and Anthropic’s overview describe the feature as a beta.
What each tracing path captures
A trace links related operations; its individual records, called spans, can show how a request moves through an application or service. The SDK and Console paths capture different parts of that journey, so choosing between them depends on where you need visibility.
| Decision | SDK tracing | Console API tracing |
| Where it is configured | In your application, using its OpenTelemetry exporter | In Claude Console, by adding an observability destination |
| Permission | No special Console role | An organization admin must add or change destinations |
| What it captures | Application-side spans, including client tools run by the tool runner | Messages API requests, individual model runs, Managed Agents turns and server-run tools |
| Shared coverage | Messages API requests and token usage | Messages API requests and token usage |
Anthropic’s OpenTelemetry overview, SDK guide and API guide detail the two paths. If you need to inspect tools running inside your own application, SDK tracing covers that client-side activity; Console API tracing adds visibility into work performed inside Claude Platform. The routes can be combined. With context propagation configured, supported Claude Platform API spans can join application traces; Anthropic says AWS API spans begin separate traces because AWS does not yet support the traceparent header.
The documented beta scope also differs by route. Anthropic lists SDK tracing for Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry. Console-configured API tracing is listed for Claude API and Claude Platform on AWS.
Set up SDK tracing before the first request
For SDK tracing, register an OpenTelemetry tracer provider before the first SDK call you want to capture, then use the application’s exporter to send spans to your observability service. Registering it later means earlier calls will not have been traced. Anthropic’s SDK guide also notes that SDK options and span schemas may change incompatibly during beta; pinning the SDK version lets you control when you adopt those changes.
Streaming affects when a span finishes: it ends once your application has read or closed the stream. In full-content mode, response data is included when the stream is read with messages.stream(). That makes stream consumption part of the tracing picture, not just the API call itself.
Choose what content leaves your system
Metadata-only is the default for both routes. Full-content mode adds model prompts and responses, so selecting it changes what reaches the destination. The receiving observability provider then applies its own access and retention controls to exported data. Anthropic’s documentation on content modes sets a default maximum of 0.75 MB per full-content span. Oversized spans are shortened and marked anthropic.content.truncated=true; request spans drop the system prompt and earliest messages first, while tool spans drop result blocks from the end. Images and documents appear only as their type.
Organizations with a HIPAA, FERPA or K–12 data-handling restriction with Anthropic can export metadata only through Console destinations, regardless of the configured content mode. Anthropic says that destination restriction does not apply to SDK spans. For organizations using zero data retention, full content remains available for Messages API tracing, and Anthropic holds exported content only in memory until delivery. Claude Managed Agents is not eligible for zero data retention, and its retention is unchanged by OpenTelemetry.
Anthropic also ties the beta to the customer’s agreement, including the Beta Services section of its Service Specific Terms. That contractual qualification sits alongside the practical choice of content mode and destination controls.
Configure the Console destination
Console API tracing does not require an application code change to export API-side spans, but an organization admin must configure the destination. The endpoint must accept OTLP/HTTP protobuf over HTTPS on port 443. Anthropic appends /v1/traces to the base endpoint URL unless you set a separate traces endpoint. Each organization can configure up to 10 destinations, with up to 1,000 selected workspaces and 20 request headers per destination. See Anthropic’s API tracing guide for the configuration requirements.
The destination also has to meet separate delivery deadlines: the TLS handshake must finish within 5 seconds, response headers must arrive within 5 seconds, and each delivery times out after 10 seconds total.
Traces are not a complete usage record
Anthropic describes delivery as best effort and at least once. Spans may be lost or duplicated, so trace export is not a complete record of API usage; Anthropic directs customers to the Usage API for usage tracking.
Claude Code monitoring is a separate telemetry path with its own documentation. The public beta announced on October 14 concerns Claude Platform’s Messages API, Managed Agents and SDK tracing.
