AnoSys SDK for Anthropic

Automatically capture and send Anthropic (Claude) API calls to AnoSys for monitoring, analytics, and observability.

Features

✨ Native Instrumentation - Uses the OpenTelemetry spans the Anthropic SDK emits itself; no third-party instrumentor
✨ Streaming Support - messages.stream() and stream=True calls are captured with final usage
✨ OpenTelemetry Semantic Conventions - Follows Gen AI standards, same columns as the OpenAI integration
✨ Prompt Caching & Thinking - Cache read/write tokens and extended thinking are tracked
✨ Zero Configuration - Works out of the box with just your API key

Installation

pip install anosys-sdk-anthropic

Requires anthropic>=2.0.0.

Quick Start

import os
import anthropic
from anosys_sdk_anthropic import AnosysAnthropicLogger

os.environ["ANTHROPIC_API_KEY"] = "your-anthropic-key"
os.environ["ANOSYS_API_KEY"] = "your-anosys-key"

# Initialize (do once at startup, before creating clients)
AnosysAnthropicLogger()

# Use Anthropic normally - all calls are automatically logged
client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    system="You are a helpful assistant.",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(message.content[0].text)

Options

AnosysAnthropicLogger(
    get_user_context=lambda: {"session_id": "abc", "token": "user-token"},
    capture_content=True,       # record prompts and completions (default)
    use_batch_processor=False,  # BatchSpanProcessor instead of SimpleSpanProcessor
)

capture_content sets ANTHROPIC_OPEN_TELEMETRY_TRACES_CONTENT_MODE=content when it is not already set. The SDK reads it when a client is constructed, so create clients after the logger. A client given its own open_telemetry option uses that option instead, and the SDK treats it as complete: open_telemetry={"tracer_provider": provider} alone silently switches the client to metadata-only. When you need to pass an option, start from get_anthropic_open_telemetry_options(), which keeps the active provider and your capture_content choice:

from anosys_sdk_anthropic import get_anthropic_open_telemetry_options

client = anthropic.Anthropic(open_telemetry={**get_anthropic_open_telemetry_options(), "conversation_id": "abc"})

If your application already registers an OpenTelemetry TracerProvider, the logger attaches to it. It sends the Anthropic SDK's spans plus any GenAI spans your code creates around them (spans with a gen_ai.operation.name, such as an agent loop or tool calls, as in the support-copilot notebook); other spans on the provider are skipped.

What Data is Captured?

Following OpenTelemetry Gen AI standards:

• gen_ai.system / gen_ai.provider.name - Always "anthropic"
• gen_ai.request.model, gen_ai.request.max_tokens, gen_ai.request.stop_sequences
• gen_ai.response.model, gen_ai.response.id, gen_ai.response.finish_reasons (stop reason)
• gen_ai.usage.input_tokens (including cached tokens), gen_ai.usage.output_tokens, gen_ai.usage.total_tokens
• Prompt cache tokens (cvn4 read, cvn5 write) and extended thinking (cvb3)
• System instructions, input and output messages
• Timestamps, duration, errors

Records are tagged cvs200 = "AnthropicTelemetry". See MAPPINGS.md for the column map.

Environment Variables

License

Apache 2.0