OpenInference Instrumentation for Claude Agent SDK

OpenTelemetry-based instrumentation for the Claude Agent SDK (@anthropic-ai/claude-agent-sdk). Produces AGENT and TOOL spans following OpenInference semantic conventions.

Installation

npm install @arizeai/openinference-instrumentation-claude-agent-sdk

Usage

Auto-instrumentation (CommonJS)

import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { ClaudeAgentSDKInstrumentation } from "@arizeai/openinference-instrumentation-claude-agent-sdk";

const provider = new NodeTracerProvider();
provider.register();

const instrumentation = new ClaudeAgentSDKInstrumentation();
instrumentation.setTracerProvider(provider);

Manual instrumentation (ESM)

import * as ClaudeAgentSDK from "@anthropic-ai/claude-agent-sdk";
import { ClaudeAgentSDKInstrumentation } from "@arizeai/openinference-instrumentation-claude-agent-sdk";

const instrumentation = new ClaudeAgentSDKInstrumentation({
  tracerProvider: provider,
});
const instrumentedClaudeAgentSDK = instrumentation.manuallyInstrument(ClaudeAgentSDK);

for await (const message of instrumentedClaudeAgentSDK.query({
  prompt: "Write a haiku about telemetry.",
})) {
  console.log(message);
}

With trace configuration (mask sensitive data)

const instrumentation = new ClaudeAgentSDKInstrumentation({
  tracerProvider: provider,
  traceConfig: {
    hideInputs: true,
    hideOutputs: true,
  },
});

Instrumented APIs

V1: query()

The query() function is wrapped to produce:

• An AGENT span for the entire query lifecycle
• TOOL child spans for each tool call (via hook injection)

V2 (unstable): unstable_v2_prompt(), unstable_v2_createSession(), unstable_v2_resumeSession()

• unstable_v2_prompt() produces an AGENT span
• Session methods produce per-turn AGENT spans with TOOL child spans

Span Attributes

AGENT spans

TOOL spans