Observability and Evals for Claude Platform (OTel)

Set up Langfuse

1. Sign up for Langfuse Cloud or self-host Langfuse.
2. Create or select a project. Open Settings → API Keys, create a project key pair, and note the project's data region.

Add a Langfuse destination

Open OpenTelemetry settings in the Claude Console and select Add destination.

• Select Langfuse as the provider.
• Select your Langfuse project's region. The Console offers US, EU, JP, HIPAA, and self-hosted.
• Enter the project's public key and secret key.
• Select Full content to capture message and tool content when your organization permits it.
• Under Workspaces, choose All, Selected, or All except.
• Save the destination.

Send a request and find its trace

Messages API

Set ANTHROPIC_API_KEY in your local environment. Set ANTHROPIC_MODEL to a model available in your Claude workspace. This example requires curl and Python 3. It prints the response headers so you can match request-id to the trace.

#!/usr/bin/env bash
set -euo pipefail

: "${ANTHROPIC_API_KEY:?Set your workspace API key}"
: "${ANTHROPIC_MODEL:?Set an available model}"

payload=$(python3 - <<'PY'
import json, os
print(json.dumps({
    "model": os.environ["ANTHROPIC_MODEL"],
    "max_tokens": 64,
    "messages": [{
        "role": "user",
        "content": "Reply with exactly: Langfuse OTel check passed",
    }],
}))
PY
)

curl --fail-with-body --silent --show-error --include \
  https://api.anthropic.com/v1/messages \
  -H "x-api-key: ${ANTHROPIC_API_KEY}" \
  -H 'anthropic-version: 2023-06-01' \
  -H 'content-type: application/json' \
  --data "$payload"

Allow about a minute after enabling export. Check Delivery in the Claude Console. In Langfuse, open Tracing and find the request. Match anthropic.request.id to the response request-id. Platform spans have service.name set to claude-platform.

Claude Managed Agents

Run a session in the enabled workspace. Its agent turns, model requests, and platform-run tools appear as observations in Langfuse.

Read your trace

These examples use platform export and optional SDK tracing.

Messages API

Select the platform anthropic.messages.create span. With Full content enabled, this platform request span contains the prompt and response. A response can contain a tool request rather than the final answer. Follow the next request to see the result after the tool runs.

The child anthropic.model_request spans show model usage and timing, but do not contain the prompt or response. To read that content, select the parent anthropic.messages.create server span. An empty model preview does not mean that content export failed.

Explore the Messages API example trace.

Claude Managed Agents

Select the agent's turn observation, such as returns-reviewer, to read its input and output. Its child observations show model requests and tool activity. The platform does not export the internal model prompt and response on anthropic.model_request spans. An empty model preview does not mean that content export failed.

Open the session link to see related traces.

Explore the Managed Agents example trace.

The example uses synthetic order data. A coordinator looks up an order and return policy. A separate reviewer checks the proposed refund, writes an audit note, and reads it back. No refund is issued. Select the reviewer's turn to read its content, then select its child model and tool observations to inspect the individual steps.

Count tokens

Messages API token and cost totals in Langfuse can include repeated usage across SDK, server, and model spans. Use the counting rules below to interpret usage.

Anthropic's counting rules say which spans to include:

• Messages API: count one source per request. With platform spans, count either the server request or its model children.
• Managed Agents: count model-request and outcome-evaluation spans.
• For Messages API requests, count compaction usage once. Cache tokens are already included in input tokens; reasoning tokens are included in output tokens.