AnoSys SDK for Claude Agents

Automatically trace Claude Agent SDK runs and send them to AnoSys: agents, tools, subagents, tokens and cost.

Installation

pip install anosys-sdk-claude-agents

Quick Start

import asyncio
from anosys_sdk_claude_agents import AnosysClaudeAgentsLogger

# Initialize once, before importing query (instruments query() and ClaudeSDKClient)
AnosysClaudeAgentsLogger()

from claude_agent_sdk import ClaudeAgentOptions, query


async def main():
    options = ClaudeAgentOptions(allowed_tools=["Bash", "Read"])
    async for message in query(prompt="List the Python files here", options=options):
        print(message)

asyncio.run(main())

ClaudeSDKClient is traced too: each receive_response() produces one agent run.

If query was imported before the logger was created, wrap it explicitly:

from anosys_sdk_claude_agents import wrap_query
query = wrap_query(query)

For other message streams, use trace_messages(stream, prompt=..., options=...).

Options

AnosysClaudeAgentsLogger(
    get_user_context=lambda: {"session_id": "abc", "token": "user-token"},
    agent_name="support_agent",   # gen_ai.agent.name (default "claude_agent")
    capture_content=True,          # prompts, results, tool input/output (default)
    emit_llm_spans=False,          # one chat span per assistant response
    use_batch_processor=False,
)

What Data is Captured?

The Agent SDK runs the model inside a Claude Code subprocess, so the integration builds spans from the messages the run yields:

Token and cost totals live on the agent span. When emit_llm_spans is on, chat spans also carry per-call usage, so sum one or the other, not both.

The run span is the active span while the SDK works, so the Claude Code subprocess inherits the trace context (TRACEPARENT).

Records are tagged cvs200 = "ClaudeAgentsTelemetry" and use the same columns as anosys-sdk-anthropic. See MAPPINGS.md.

Environment Variables

License

Apache 2.0