Claude Code InternalsHow an AI coding agent actually works, explained in plain language

Based on Claude Code v2.1.88 source reconstruction. 8 chapters from zero to understanding the agent core.

🔄

The Core Loop

Claude Code's architecture is a single while loop. No complex frameworks, no vector search — surprisingly simple.

🛠️

8 core tools let AI read files, write code, and run commands. Learn how they're defined, registered, and invoked.

🛡️

A five-layer permission model makes AI both powerful and controllable. From tool-level checks to user hooks.

🚀

50 lines of Python captures the agent core. Once you understand the principles, you can build your own AI coding tool.