Skip to content

Claude Agent SDK notebook

Open in Colab ยท View on GitHub

A two-tool support agent (classify_email backed by an LLM, lookup_policy) runs 20 synthetic emails with a scripted model (no API key, offline), counting classification LLM calls. Then Pecca's route_rfi is added as the first tool and the same emails run again: classification LLM calls drop from 20 to 0, the agent still answers every email, and a per-email table shows Pecca's confidence and fallback flag.

The Claude Agent SDK drives the Claude Code CLI, so this notebook points ANTHROPIC_BASE_URL at a small local server that replays the scripted plan; the tools still run through the SDK's in-process MCP server.

Run it locally (the notebook installs its own dependencies): 

pip install pecca claude-agent-sdk
jupyter lab examples/notebooks/agent_claude_agent_sdk.ipynb
The committed notebook has its outputs, so GitHub renders the results without running anything. Runtime: about 30 seconds. Needs Docker: no.