OpenTelemetry Anthropic Instrumentation

Warnings

• breaking Breaking changes were introduced in version 0.54.0 to conform to the OpenTelemetry Generative AI Semantic Conventions (GenAI SemConv). This may alter span attribute names and structure.
  Fix: Review the OpenTelemetry GenAI Semantic Conventions documentation for the updated attribute naming and structure. Your observability backend might require updates to dashboards or alerts.
• gotcha By default, this instrumentation logs the full content of prompts, completions, and embeddings to span attributes. This data may contain sensitive information.
  Fix: To disable logging of sensitive content, set the environment variable `TRACELOOP_TRACE_CONTENT` to `false`.
• gotcha When using `client.messages.stream()` with base64 encoded images, particularly with Langfuse, input token counts might be inflated in traces.
  Fix: This is a known upstream issue (traceloop/openllmetry#3949). Monitor the project's GitHub for a fix. Consider alternative methods for calculating token usage or use non-streaming calls if accurate token counts are critical for images.
• gotcha Instrumentation itself only creates spans. An OpenTelemetry SDK (TracerProvider, SpanProcessor, Exporter) must be explicitly configured to process and send telemetry data to a backend.
  Fix: Always initialize a `TracerProvider`, add at least one `SpanProcessor` (e.g., `BatchSpanProcessor`), and configure an appropriate `SpanExporter` (e.g., `OTLPSpanExporter`) before calling `instrument()` on the Anthropic instrumentor.

Install

• pip install opentelemetry-instrumentation-anthropic anthropic opentelemetry-sdk opentelemetry-exporter-otlp Install with core OTel SDK and OTLP exporter

Imports

• AnthropicInstrumentor

  from opentelemetry.instrumentation.anthropic import AnthropicInstrumentor

Quickstart

This quickstart demonstrates how to set up OpenTelemetry with the Anthropic instrumentation, configure an OTLP HTTP exporter, and make a traced Anthropic API call. Ensure `ANTHROPIC_API_KEY` is set in your environment. An OpenTelemetry Collector or compatible backend should be running to receive traces.

import os
from anthropic import Anthropic
from opentelemetry.instrumentation.anthropic import AnthropicInstrumentor
from opentelemetry import trace
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter

# --- OpenTelemetry Setup ---
# 1. Configure the OpenTelemetry TracerProvider
resource = Resource.create({"service.name": "anthropic-llm-app"})
provider = TracerProvider(resource=resource)
trace.set_tracer_provider(provider)

# 2. Configure an OTLP exporter to send traces (e.g., to an OTLP collector or a service like SigNoz/Arize)
# Default OTLP HTTP endpoint is http://localhost:4318/v1/traces
otlp_exporter = OTLPSpanExporter()
span_processor = BatchSpanProcessor(otlp_exporter)
provider.add_span_processor(span_processor)

# 3. Instrument the Anthropic library
AnthropicInstrumentor().instrument()

# --- Anthropic API Call ---
ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY", "YOUR_ANTHROPIC_API_KEY")

if ANTHROPIC_API_KEY == "YOUR_ANTHROPIC_API_KEY":
    print("WARNING: ANTHROPIC_API_KEY not set or placeholder. API calls will fail without a valid key.")
else:
    try:
        client = Anthropic(api_key=ANTHROPIC_API_KEY)

        print("Making an Anthropic API call...")
        response = client.messages.create(
            model="claude-3-opus-20240229", # Or another suitable model
            max_tokens=100,
            messages=[
                {"role": "user", "content": "Explain the concept of quantum entanglement in a sentence."}
            ],
        )

        print("Anthropic API call successful.")
        print(f"Response: {response.content[0].text[:50]}...")

    except Exception as e:
        print(f"Error during Anthropic API call: {e}")

# Ensure all spans are exported before the application exits
provider.force_flush()