claude-code-from-scratch

Why we included this project

This is a teaching project for people who want to understand how a coding agent works internally, not just what it does for them. The real Claude Code is a few hundred thousand lines of code; here the core pieces get rebuilt in roughly five thousand lines of TypeScript and Python, spread across thirteen chapters that each add one component: the agent loop, the tool system, context compression, memory recall, skills, multi-agent setups, and MCP integration. Every chapter maps its code back to the Claude Code architecture so you can see what was simplified and why. The part that stands out for learners is that each chapter runs with a single command and no API key, using a local mock model, and a --diff flag shows exactly what that chapter adds. For developers or tech leads thinking about building or evaluating a lightweight agent of their own, it is a genuinely useful reference for how the pieces fit together.