Learn Claude Code

Resources & lists

20-lesson tutorial reverse-engineering Claude Code's harness: agent loop, tools, subagents, MCP, teams, in runnable Python.

by shareAI-labshareAI-lab/learn-claude-code

Stars79k

Forks12.6k

Watchers312

Open issues69

LicenseMIT

Last push23 days ago

Install

install

$ git clone https://github.com/shareAI-lab/learn-claude-code

Update

update

$ git pull / re-visit the repo

Tutorial repo -- no install; clone and run the chapter code, or bookmark and revisit.

When to use it

You want to understand how Claude Code actually works internally -- agent loop, tool dispatch, subagents, context compaction, MCP routing.
You're building your own coding-agent harness and want runnable, minimal reference implementations to learn from.
You want to teach or onboard engineers on harness engineering versus model training (agency = model + harness).
You need a structured, chapter-by-chapter path from one-loop-plus-bash up to multi-agent teams and worktree isolation.

Commands & flags

git clone https://github.com/shareAI-lab/learn-claude-code: Clone the repo to read and run all 20 chapters locally.
pip install -r requirements.txt && cp .env.example .env: Install deps and set ANTHROPIC_API_KEY before running any lesson.
python s01_agent_loop/code.py: Run lesson 1: the minimal agent loop with a single bash tool.
python s08_context_compact/code.py: Run the context-compaction lesson.
python s20_comprehensive/code.py: Run the endpoint chapter assembling every mechanism around one loop.
cd web && npm install && npm run dev: Launch the Next.js web platform at localhost:3000.

How to use it

This is a tutorial/reference repo, not a skill or plugin. Clone it and work through the chapters in order:

git clone https://github.com/shareAI-lab/learn-claude-code
cd learn-claude-code
pip install -r requirements.txt
cp .env.example .env   # set ANTHROPIC_API_KEY
python s01_agent_loop/code.py

The canonical track is the root-level s01_agent_loop/ through s20_comprehensive/ folders. Each chapter has a narrative README (zh default, plus README.en.md / README.ja.md), a standalone code.py, and an images/ dir. Read s01 -> s20 in order; each adds exactly one harness mechanism (permissions, hooks, todo, subagents, skills, compaction, memory, tasks, teams, worktree isolation, MCP) on top of the unchanged core loop. A legacy 12-lesson track lives under agents/ (e.g. python agents/s01_agent_loop.py), and a Next.js web platform runs via cd web && npm install && npm run dev.

Why it earns its place

Trending on GitHub (Trendshift-featured) and authored by shareAI-lab, the team behind the Kode agent CLI (@shareai-lab/kode) and kode-agent-sdk. It is one of the clearest end-to-end breakdowns of agent harness engineering: 20 minimal, runnable Python chapters that strip Claude Code down to its essence -- one loop plus tools, on-demand skill loading, context compaction, subagents, task graphs, teams, worktree isolation, permissions, hooks, memory, and MCP routing. Its thesis -- agency comes from model training, the harness is the vehicle -- reframes "building an agent" as building a harness, directly applicable to anyone extending Claude Code.

Try it

claude

> Walk me through s06_subagent/code.py and explain how it keeps the subagent's context isolated from the main loop.