Claude Agent SDK → OpenObserve

Capture agent query timing, token usage, turn counts, and error status for every agent run built with Anthropic's Claude Agent SDK. The SDK does not include a built-in OTel instrumentor, so traces are created by wrapping query() calls in manual spans and extracting usage data from ResultMessage.

Prerequisites

Python 3.10+
Claude Code CLI installed (npm install -g @anthropic-ai/claude-code or existing Claude Code installation)
An OpenObserve account (cloud or self-hosted)
Your OpenObserve organisation ID and Base64-encoded auth token
An Anthropic API key

Installation

pip install openobserve-telemetry-sdk claude-agent-sdk python-dotenv

Configuration

Create a .env file in your project root:

# OpenObserve instance URL
# Default for self-hosted: http://localhost:5080
OPENOBSERVE_URL=https://api.openobserve.ai/

# Your OpenObserve organisation slug or ID
OPENOBSERVE_ORG=your_org_id

# Basic auth token — Base64-encoded "email:password"
OPENOBSERVE_AUTH_TOKEN=Basic <your_base64_token>

# Anthropic API key
ANTHROPIC_API_KEY=your-anthropic-key

Instrumentation

The Claude Agent SDK runs Claude Code CLI as a subprocess and has no built-in OTel instrumentation. Call openobserve_init() to set up the tracer provider, then wrap each query() call in a manual span and extract usage data from the ResultMessage.

import asyncio
from dotenv import load_dotenv
load_dotenv()

from openobserve import openobserve_init
openobserve_init()

from opentelemetry import trace
from claude_agent_sdk import query, ClaudeAgentOptions, ResultMessage

tracer = trace.get_tracer(__name__)


async def run_agent(prompt: str):
    options = ClaudeAgentOptions(
        allowed_tools=["Read", "Glob"],
        permission_mode="dontAsk",
        max_turns=3,
    )
    with tracer.start_as_current_span("claude_agent.query") as span:
        span.set_attribute("claude_agent.prompt", prompt[:100])
        async for message in query(prompt=prompt, options=options):
            if isinstance(message, ResultMessage):
                span.set_attribute("claude_agent.subtype", message.subtype)
                span.set_attribute("claude_agent.num_turns", message.num_turns)
                span.set_attribute("claude_agent.is_error", message.is_error)
                span.set_attribute("claude_agent.duration_api_ms", message.duration_api_ms)
                if message.usage:
                    span.set_attribute("claude_agent.input_tokens", message.usage.get("input_tokens", 0))
                    span.set_attribute("claude_agent.output_tokens", message.usage.get("output_tokens", 0))
                    span.set_attribute("claude_agent.cache_read_input_tokens", message.usage.get("cache_read_input_tokens", 0))


asyncio.run(run_agent("List the Python files in this directory."))

If Claude Code CLI is not on your PATH, pass its location explicitly:

options = ClaudeAgentOptions(
    cli_path="/path/to/claude",
    permission_mode="dontAsk",
    max_turns=3,
)

What Gets Captured

Each query() call produces one span containing the following attributes sourced from ResultMessage.

Attribute	Description
`claude_agent_prompt`	First 100 characters of the prompt
`claude_agent_subtype`	`success` or `error_during_execution`
`claude_agent_num_turns`	Number of agentic turns executed
`claude_agent_is_error`	`true` if the agent encountered an error
`claude_agent_duration_api_ms`	Time spent in API calls (milliseconds)
`claude_agent_input_tokens`	Total input tokens consumed
`claude_agent_output_tokens`	Total output tokens generated
`claude_agent_cache_read_input_tokens`	Input tokens served from prompt cache
`duration`	End-to-end span latency

Viewing Traces

Log in to OpenObserve and navigate to Traces
Filter by operation name claude_agent.query to find agent spans
Click any span to inspect token usage, turn count, and duration
Use claude_agent_is_error = true to filter for failed runs

Claude Agent SDK trace in OpenObserve

Next Steps

With Claude Agent SDK instrumented, every agent run is recorded in OpenObserve. From here you can track token usage per prompt type, measure multi-turn agent costs, and alert on error runs.