Anthropic opens public beta for OpenTelemetry tracing on the Claude platform

On 14 October 2026, Anthropic opened a public beta that lets developers export OpenTelemetry traces from Claude Managed Agents, the Messages API, and the Claude SDKs directly to any OTLP-compatible observability backend. One destination added in the Claude Console is all it takes, and from that point every agent turn, model call, and tool use lands in your existing observability stack as a standard OpenTelemetry trace.

If you have been building on the Claude platform and watching your agent behaviour through logs or custom instrumentation, this gives you something more structured: a vendor-neutral, standardised trace for every step Claude takes.

What is actually being exported

OpenTelemetry is the industry-standard, vendor-neutral framework for collecting traces, metrics, and logs. A trace is made up of spans, where each span is a timed record of one operation. OTLP is the protocol used to ship those spans to a backend. The reason this matters is that you do not need to build a separate observability layer for the model layer. Claude can speak the same language as the rest of your application.

The tracing beta covers three surfaces:

• The Messages API, which emits spans for each request once you register an OTLP destination in the Claude Console.
• Claude Managed Agents, which adds per-session and per-turn tracing including tool use and calls to Model Context Protocol (MCP) servers.
• Claude SDKs in Python, TypeScript, Go, Java, C#, PHP, and Ruby, which emit client-side spans and propagate trace context through to the API.

Spans follow the OpenTelemetry GenAI semantic conventions, so Claude trace data works in the same tools and dashboards you use for any other provider.

Where the traces can go

The list of supported OTLP-compatible backends is broad. For debugging alongside application logs and metrics, traces work with Grafana Cloud, Honeycomb, Sentry, and Splunk. For evaluation and quality measurement from production data, they work with Arize AI, Braintrust, Datadog, Langfuse, LangSmith by LangChain, New Relic, and PostHog. For organisation-wide usage tracking and security review, you can route traces to Databricks Unity Gateway.

The point is that you choose the tool. Anthropic ships the trace; you decide where it lands.

Two ways to set it up

There are two integration routes, and they serve different needs.

Via the Claude Console: An organisation admin registers an OTLP/HTTP destination (reachable over HTTPS on port 443), selects which workspaces it applies to, and that is it. No application code changes are required. This route covers API and Managed Agents spans.

Via the Claude SDK: A developer registers an OpenTelemetry tracer provider in their application code. This creates client-side spans and propagates trace context into the API, giving you a single continuous trace that spans your application and Claude together. No special admin permission is needed for this route.

MCP connectors can also be instrumented separately to attach downstream tool spans to the same trace, giving you end-to-end visibility from the user request through Claude to any external tool it called.

Controlling what goes into a span

Each destination and each SDK client has a content mode setting that determines how much detail appears in a span.

Metadata only is the default. Spans carry identifiers, timing, status, token counts, model name, and text you choose to include such as agent names, tool names, and MCP server descriptions. This is enough to debug latency and spot failures without exposing sensitive content.

Full content adds model prompts and responses, tool inputs, and tool results. Each span holds up to 0.75 MB of content by default, which Anthropic says is sufficient for a 200k-token context window of English text and sits below the 1 MB request limit common across observability platforms. When content exceeds the limit, the SDK or API shortens it and sets anthropic.content.truncated to true. For requests, the system prompt and earliest messages are dropped first; for tool spans, blocks are dropped from the end of the result.

There is one important restriction for regulated environments. Organisations with a HIPAA, FERPA, or K-12 data-handling agreement with Anthropic export only metadata from Console destinations regardless of the configured content mode. SDK spans are not subject to that restriction. Under zero data retention, full content remains usable because the API holds exported content only in memory until it delivers it to the destination.

What this means for teams building with Claude

For developers, the practical benefit is that agent traces sit alongside your application logs and metrics in the tools you already use. When a step is slow or fails, you can see whether the problem originated in your code, in a model call, or in a downstream tool, all from one timeline.

For teams trying to improve agent quality over time, having structured production traces fed into evaluation tools like Langfuse or Braintrust means you can measure whether a prompt change actually helped, catch regressions, and build evals from real behaviour rather than synthetic test cases.

For platform and security teams, routing traces to a tool like Databricks Unity Gateway means token usage becomes auditable by team, and unusual activity can be flagged through the same processes you use for the rest of your infrastructure.

One practitioner quoted in Anthropic’s announcement put the regulated-industry case plainly:

“Agents that touch a company’s books have to show their work, the same way an accountant does. Our agents run on the Claude Platform and our telemetry lives in Grafana Cloud, and native OpenTelemetry export from Claude puts each model call in the same trace as the reconciliation or journal-entry workflow it ran inside. We see the prompt version, tool calls, latency, and token spend alongside the rest of our stack, so when an auditor asks what an agent did and why, we can answer from one timeline.”

A few things to note before you enable it

Delivery is best-effort, not guaranteed. Spans can be lost, which is standard behaviour for telemetry pipelines but worth knowing if you are building compliance workflows on top of this data.

The Console setup requires the organisation admin role. SDK setup does not.

Anthropic has not announced pricing for the tracing feature or a timeline for general availability. The beta is covered by the Beta Services section of Anthropic’s service terms.

If you already instrument Claude Code or Cowork with OpenTelemetry, agents built on the Claude platform can be configured to appear in the same dashboards without additional work.

Full setup documentation is available in the Claude Platform OpenTelemetry docs.