bozhoudev/14days-build-claude-code-cli

This is a 14-day tutorial project for building a code agent from scratch in Python. A code agent is a program that can read files, edit code, run terminal commands, and talk to an AI model to complete programming tasks. The goal here is not to produce a polished product but to understand how the outer layer around an AI model, called a harness, is actually structured. The tutorial uses Claude Code as the design reference. The author studied a publicly available source snapshot of Claude Code to understand how it handles things like tool calls, permissions, file editing, session memory, and subagents, then rebuilt those ideas in Python so learners can run and modify each piece themselves. The AI model used in the examples is configurable, by default it points to a cost-friendly DeepSeek endpoint that is compatible with the Anthropic API format, so you can follow along without paying for Claude API calls, though switching to Claude or any other compatible service is straightforward. The 14 days are split into two halves. The first seven days build a working single-agent CLI that can accept prompts, call the model, use tools, read and edit files safely, run bash commands with a permission system, and save sessions. The second seven days extend it with plan mode, skills, subagents, context compression, a simple multi-agent coordinator, isolated work directories called worktrees, and a plugin system called MCP that lets you add external tools. Each day has a tutorial document and a reference snapshot directory under packages/ with a working implementation you can run directly. The learning path involves building in your own project folder while consulting those snapshots, not copying each day's folder wholesale. A web-based version of the tutorial also lives in the repository under agent-code-learn/ and can be previewed locally with npm. The project is not affiliated with Anthropic and is released under the MIT license.