Loading repository data…
dadiaomengmeimei / repository
Reverse-engineering Claude Code's 512K+ lines of TypeScript — architecture, design decisions, and 11 transferable patterns for building AI Agents
Reverse-engineering Claude Code's 512K+ lines of TypeScript — extracting the architecture, design decisions, and 11 transferable patterns that make a production AI Agent work.
Not just a code walkthrough. An architectural analysis that explains why each design choice was made, with alternatives compared.
buildTool() factory + Zod schema as prompt, 40+ tools with 18+ feature flags/simplify 3-agent parallel review| # | Topic | Core Source Files | Time |
|---|---|---|---|
| 00 | Index & Architecture Overview | — | 10min |
| 01 | Global Architecture | main.tsx, App.tsx, QueryEngine.ts | 30min |
| 02 | Tool System | Tool.ts, tools.ts, GlobTool.ts | 45min |
| 03 | Permission & Security | permissions.ts, filesystem.ts | 45min |
| 04 | Query Loop & API ⭐ | query.ts, StreamingToolExecutor.ts | 50min |
| 05 | Multi-Agent + Coordinator ⭐ | AgentTool.tsx, runAgent.ts | 50min |
| 06 | MCP, Skills & Extensions | mcp/client.ts, skills/ | 45min |
| 07 | Prompt Engineering | prompts.ts, BashTool/prompt.ts | 60min |
| 08 | Voice Mode & Buddy | voiceStreamSTT.ts, buddy/ | 30min |
Total ~5.5 hours for a systematic understanding of Claude Code internals.
Course content available in Chinese (zh-CN/) and English (en/).
| Path | Chapters | Time |
|---|---|---|
| Quick Start | 00 → 01 → 02 → 04 | 2h |
| Security Focus | 00 → 03 → 02 | 1.5h |
| Agent Research | 00 → 05 → 04 → 06 | 2.5h |
| Prompt Research | 00 → 07 | 1h |
| Full Course | 00 → 01 → ... → 08 | 5.5h |
| Chapter | Most Valuable Finding |
|---|---|
| 01 | React/Ink terminal UI — CLI can use React |
| 02 | buildTool() factory + Zod schema doubles as prompt |
| 03 | 5-layer defense; bypass mode still has safety floor (.git/.claude protected) |
| 04 | StreamingToolExecutor concurrent execution, 5-layer compression pipeline |
| 05 | Coordinator star topology, fork prompt cache sharing, 10-step finally cleanup |
| 06 | Skills 3-layer architecture, /simplify 3-agent parallel review |
| 07 | 40+ prompt files with static/dynamic separation, Meta-Prompt generates Agents |
| 08 | Record-first-connect-later eliminates latency, Buddy Bones not persisted |
Extracted from Ch.04 & Ch.05 — applicable to any Agent system:
From Query Loop: Optimistic Recovery · Layered Degradation · State Machine + Transition Log · Read-Write Lock Concurrency · Immutable Config Snapshot · Hierarchical Cancellation
From Multi-Agent: Capability-based Security · Cache-Friendly Forking · Deterministic Cleanup · Star Topology Orchestration · Monotonic Permission Narrowing
Each pattern includes Claude Code implementation details and universal applications (DB, K8s, microservices, etc.) in the chapter notes.
1,646 lines of TypeScript · 15 files · 4 dependencies
nano-claude-code rewrites Claude Code's core as a minimal, hackable implementation you can actually run and modify.
| Claude Code | nano-claude-code | |
|---|---|---|
| Files | ~1,900 | 15 |
| Lines | 512,000+ | 1,646 |
| Dependencies | 50+ | 4 |
| Tools | 40+ | 6 |
Read the notebook to understand the architecture. Run nano-claude-code to build your own.
👉 github.com/dadiaomengmeimei/nano-claude-code
MIT. The analyzed Claude Code source code is copyrighted by Anthropic. This project is for educational and research purposes only.
If this helps you, give it a ⭐! Follow @dadiaomengmeimei for more AI source code analysis.
Selected from shared topics, language and repository description—not editorial ratings.
washi4 /
Interactive course teaching how AI coding agents work — reverse-engineering Claude Code's architecture. Generated from shareAI-lab/learn-claude-code + zarazhangrui/codebase-to-course.