Claude Platform Now Exports OpenTelemetry Traces From Managed Agents, the Messages API and 7 SDKs, but Delivery Is Best Effort and Setup Is Admin Only

Anthropic has added OpenTelemetry export to the Claude Platform. Since 14 October 2026, traces from Claude Managed Agents, the Messages API and the Claude SDKs can flow into Grafana Cloud, Datadog, Langfuse or any other tool that accepts a standard OTLP endpoint. The feature is in public beta.

The pitch, in Anthropic’s announcement post, is simple: every agent turn, model call and tool use shows up as a standard trace with latency, tokens and IDs, next to the rest of your stack. If you already send Claude Code or Cowork traces somewhere, your own Claude Platform agents can now land in the same place.

We read the docs line by line. The feature is real and useful, but it is a beta with an admin gate, best effort delivery and a few ways to get your numbers wrong.

What Anthropic actually shipped

There are two halves to this launch, and they work differently.

API tracing covers Messages API requests and Managed Agents sessions. The API itself emits the spans, so you change no code. You add a “destination” in the Claude Console that points at your observability platform’s OTLP/HTTP endpoint, and the API starts sending spans for the workspaces you pick.

SDK tracing runs inside your application. According to the 14 October release notes, the Claude SDK emits spans in Python, TypeScript, C#, Go, Java, PHP and Ruby. You register an OpenTelemetry tracer provider before your first call and the SDK exports spans through it automatically.

The OpenTelemetry overview says you can run both at once, in which case a request’s SDK span and API span sit in one trace. That is the setup that shows you the full picture, from your own POST /chat handler down to each model request and tool call Claude ran on the server.

Spans follow the OpenTelemetry GenAI semantic conventions, which means platforms that already understand those conventions can parse Claude traces without custom mapping.

SDK tracing vs API tracing at a glance

Question	SDK tracing	API tracing
Setup	Code: register a tracer provider	Claude Console destination
Permission needed	None	Admin role in the Console organization
Coverage	Clients using the Claude SDK	Any client, including raw HTTP
Sees server tools and model runs inside a request	No	Yes
Sees your own client tools run by the tool runner	Yes	No
Claude Platform on AWS	Supported	Supported, but API spans start their own traces

Which observability tools it works with

Anthropic names a long partner list. For debugging next to logs and metrics, it points to Grafana Cloud, Honeycomb, Sentry and Splunk. For turning production traces into evals, it lists Arize AI, Braintrust, Datadog, Langfuse, LangSmith by LangChain, New Relic and PostHog. For organization-wide token tracking, it names Databricks Unity Gateway.

The partner label mostly changes the Console form. Pick a listed provider and the Console asks for its API key and sets a content size limit that platform accepts. Pick Custom and you enter the endpoint URL and headers yourself. Anything that speaks OTLP/HTTP with protobuf over HTTPS on port 443 should work.

That port detail will catch self-hosters. The API tracing guide notes that an OpenTelemetry Collector listens on port 4318 by default, so you need to serve HTTPS on 443 or put a load balancer or reverse proxy in front of it.

Metadata only or full content: the privacy switch

Each destination and each SDK client has a content mode. The default, metadata_only, sends identifiers, timing, status, token counts and names you chose, such as agent, tool and MCP server names. Switch to content (Full content in the Console) and spans also carry the system prompt, input messages and model responses.

Full content is where the debugging value is, and also where the risk is. Your prompts, user messages and tool results end up stored by a third party under that platform’s own retention and access rules. Anthropic’s docs say as much, and a destination’s terms note that after delivery the data sits outside Anthropic.

A few compliance rules are baked in:

• Organizations with a HIPAA, FERPA or K-12 data-handling restriction get metadata only from destinations, whatever the setting. SDK spans are not affected by that restriction.
• With zero data retention, Full content still works for Messages API requests. The API holds the exported content in memory only until it delivers it.
• Claude Managed Agents is not eligible for zero data retention, and turning on OpenTelemetry does not change that.

With Full content, each span holds up to 0.75 MB of content by default. Anthropic says that fits a 200k-token English context and stays under the 1 MB request cap of many platforms and proxies. Over the limit, the content is shortened and the span gets anthropic.content.truncated set to true. A full 1M-token window needs about 4 MB, so raise the limit if your backend allows it.

The caveats Anthropic puts in the fine print

This is the part most launch coverage skips.

Delivery is best effort. The overview page says exported spans are not guaranteed to be a complete record of API usage, and a span can be lost. A caller can also drop spans by turning off the sampled flag in the traceparent header. Anthropic tells you to use the Usage API, not traces, to track each API key’s usage or spot a leaked key.

Token totals are easy to double count. Usage appears on several spans per request. When you run SDK and API tracing together, both emit a span named anthropic.messages.create, told apart only by kind (CLIENT vs SERVER). The docs say to sum one source per request, and warn that retried deliveries can duplicate spans. Build a cost dashboard without reading that section and your numbers will be inflated.

The schema can break. The SDK guide warns that during the beta a new SDK version can rename a span, attribute, option or environment variable. Pin your SDK version if dashboards depend on these names.

Hard limits. Each organization gets 10 destinations, disabled ones included. The API disables a destination after 24 hours of failed deliveries, a 3xx redirect, or an endpoint that resolves to a non-public IP.

Confirmed vs not stated

Claim	Status
Public beta launched 14 Oct 2026	Confirmed in announcement and release notes
Traces from Managed Agents, Messages API and SDKs	Confirmed
Admin API endpoint with beta header telemetry-destinations-2026-08-11	Confirmed in the API reference
Separate USD fee for OTel export	Not stated anywhere in the docs we checked
General availability date	Not announced
Guaranteed, complete span delivery	No. Best effort by design

On pricing, Anthropic published no separate charge for the export itself and no new Claude model price came with this launch. Your observability vendor will still bill you for ingesting spans, and Full content traces are much heavier than metadata.

Who is already using it

Anthropic quoted four customers. Roman Shtylman, co-founder and CTO of Foxglove, said native export lets the company analyze agent interactions in PostHog alongside product usage data.

David Traina, a senior software engineer at Ramp, described an internal system called Reflect that scores Langfuse traces and proposes fixes. He said it has already cut token usage by 10 to 20 percent and made sessions 30 percent faster on some agents. That figure comes from Reflect, not from the export feature alone.

Hyon Lee, CTO of Sterling Labs, said traces now stream into Langfuse with no custom instrumentation. Ankush Sharma, Head of Infrastructure at FloQast, said Claude model calls now sit in the same Grafana Cloud trace as the reconciliation or journal-entry workflow they ran inside, which helps answer auditors.

The AWS angle and one more release note

Claude Platform on AWS gets API tracing too, with differences. The API spans for AWS requests are always in their own traces because that platform does not support the traceparent header yet. Destinations are managed through the AWS endpoint at aws-external-anthropic.{region}.api.aws with your IAM policy deciding who can change them.

The same 14 October release notes also say Claude Managed Agents scheduled deployments are now available on Claude Platform on AWS. Paired with tracing, that matters for teams running agents on a timer with no one watching. If you missed it, our piece on Managed Agents dynamic workflows that run up to 1,000 agents explains why visibility into multi-agent runs is becoming a real need.

What it means for Indian developers

The admin requirement is the first hurdle. Many Indian startups and agency teams share one Claude Console organization where only a founder or a client holds the admin role. Developers with the developer role can view destinations but cannot add them, so expect a ticket or a call before API tracing works.

SDK tracing needs no special permission, so it is the faster start. Langfuse is open source and can be self-hosted, which suits teams that want trace data on their own infrastructure. Just remember the HTTPS on port 443 rule for any Collector you run.

Be careful with Full content if your product handles customer data from Indian users. Prompts and responses copied into an overseas SaaS backend can raise questions under the Digital Personal Data Protection Act and your client contracts. Metadata only is the safer default until you have checked where your observability vendor stores data.

This also fits a wider push toward agent tooling that we have been tracking, from Claude Code’s parallel cloud threads to DigitalOcean’s agent learning services. Anthropic’s regional moves, like its new Singapore office, suggest Asia-Pacific enterprise buyers are part of the audience for features like this.

FAQ

Is Claude OpenTelemetry export free?

Anthropic’s docs list no separate fee for the export. You still pay normal Claude API usage, and your observability platform charges for the spans it ingests and stores.

Do I need to change my code to get traces?

Not for API tracing. An admin adds a destination in the Claude Console and the API sends spans for Messages API requests and Managed Agents sessions. SDK tracing needs a tracer provider registered in your app.

Which languages does the Claude SDK support for tracing?

Python, TypeScript, C#, Go, Java, PHP and Ruby, according to the 14 October 2026 release notes.

Will my prompts be sent to the observability tool?

Only if you choose Full content. The default is metadata only, which sends IDs, timing, status and token counts. HIPAA, FERPA and K-12 restricted organizations get metadata only from destinations regardless.

Can I use these traces for billing?

Anthropic says no. Delivery is best effort and spans can be lost or skipped, so use the Usage API for accurate usage and leaked key checks.