Anthropic adds OpenTelemetry traces and a Desktop sessions view

On 14 October 2026 at 17:25 UTC, Anthropic’s ClaudeDevs account posted that Claude Platform can send OpenTelemetry traces, in public beta, from the Messages API, Claude Managed Agents, and the Claude SDKs to an observability stack. At 02:50 UTC on 15 October the same account linked the guide and linked a launch article dated 14 October.

The article says the export is in public beta. The docs page marks the feature Beta and says it is provided under the agreement with Anthropic, including the Beta Services section of the Service Specific Terms. The overview page does not list a separate price.

What gets a span

The overview says the API and the SDK can trace Messages API requests, tool uses, and Managed Agents sessions. A trace is one request. A span is one step, such as a model request or a tool use. Attributes follow the GenAI semantic conventions.

SDK tracing and API tracing can be turned on together, so one trace holds both. For Claude Platform on AWS, the docs say the API spans stay in their own traces. Messages and Sessions API requests, and token usage, appear on both paths. Model runs inside a request, Managed Agents turns, and server tools are API spans. Client tools, tools in a self-hosted sandbox, and other calls such as file uploads are SDK spans. API tracing needs an admin and a Console destination. SDK tracing needs code and covers clients that use the SDK.

The span reference says names and attributes can change during the beta. Examples it documents include anthropic.messages.create, anthropic.model_request, anthropic.tool_use, anthropic.mcp_tool_use, anthropic.custom_tool_use, and anthropic.session.turn. The API does not emit anthropic.messages.create for the OpenAI-compatible chat-completions endpoint.

How to turn it on

API tracing sends spans to an OTLP/HTTP endpoint that accepts protobuf on HTTPS port 443. An admin adds a destination in the Console: a listed provider or a custom endpoint, headers, a content mode, and the workspaces covered. Filter the backend on service.name set to claude-platform. Delivery is best effort. The overview says a span can be lost, and a caller can omit spans by turning off the sampled flag in the traceparent header. For usage totals it points to the Usage API.

SDK tracing uses the application’s tracer provider and exporter. The page says to pin the SDK version, because options and the span schema can change during the beta.

Content mode defaults to metadata only: identifiers, timing, status, token counts, and chosen text such as agent and tool names. Full content adds prompts and responses, up to 0.75 MB per span by default. Over that, the span is shortened and anthropic.content.truncated is set. Organizations with a HIPAA, FERPA, or K-12 restriction export metadata only from destinations. The docs say SDK spans are outside that restriction.

The launch article names Grafana Cloud, Honeycomb, Sentry, Splunk, Arize, Braintrust, Datadog, Langfuse, LangSmith, New Relic, PostHog, and Databricks, and says any standard OTLP endpoint works. Quotes there, including Ramp’s token and speed figures, are those customers’ statements.

Desktop sessions, the same afternoon

At 19:01 UTC on 14 October, ClaudeDevs posted a thread on the Claude Code desktop app. The opening post says the app is now quicker at starting sessions, easier for splitting work across a few, and shows what each session is doing. The replies, all timestamped 19:01 UTC, add the detail.

The next post says the new session screen keeps recent folders one click away, and that a saved environment can be a few folders, an SSH host, or the cloud, with a worktree per folder. Another says a parallel task starts a nested session that reports back into the main session. A third says /btw opens a side question in its own tab, and that a suggested follow-up can run as its own session in one click. The sidebar post says the sidebar shows which sessions are working and which need a person. Inside a session, a live to-do follows the plan, a recap card sums up a turn that needs an answer, and a side rail shows where the session runs, its context, and its changes.

The last reply says to update the desktop app and start a new session, and links the desktop quickstart. The desktop reference describes the same surfaces as they stand: parallel sessions in a sidebar, a side-by-side view (Cmd-click on macOS, Ctrl-click on Windows), an optional Git worktree, Local, Cloud, or SSH, and a filter by status, project, or environment. That page does not date those controls to 14 October.

The Claude Code changelog entry for 2.1.300, also dated 14 October 2026, does not itemize that session screen. It does add OpenTelemetry GenAI semantic-convention attributes to Claude Code’s own trace spans, including gen_ai.provider.name, gen_ai.conversation.id, gen_ai.tool.name, and error.type. That is the CLI’s telemetry, separate from the Platform export above.

For the Platform traces, an admin adds a destination, and an SDK user pins the client and chooses metadata or full content. For the desktop app, the thread’s instruction is to update and start a new session.