The full source of Anthropic’s Claude Code CLI leaked through an npm map file 3 weeks ago. 512,000+ lines of production TypeScript. Public now.
A rare look inside how a billion-dollar AI agent is actually engineered — not a tutorial, the real thing.
1,900 files. 512K+ lines of TypeScript. 40+ tools. 50+ slash commands. An MCP server.
Prompt engineers, this is your Rosetta Stone. Agent builders, this is your textbook.
leaked-march-31-2026 · public-domain · 0/63 on VT · read-only study
claude-code/
├── src/
│ ├── main.tsx # Entrypoint — Commander + React + Ink
│ ├── QueryEngine.ts # Core LLM API caller (~46K lines)
│ ├── tools/ # 40+ agent tool implementations
│ ├── commands/ # 50+ slash commands
│ ├── bridges/ # VS Code · Cursor · JetBrains
│ ├── prompts/ # 🏆 THE SYSTEM PROMPTS
│ ├── hooks/ # React hooks incl. permission checks
│ └── mcp-server/ # Installable MCP integration
└── SKILL.md # How the codebase works
🏆 Why This Is Actually Interesting
The prompts/ folder — the literal system prompts Anthropic ships. How Claude Code is told to behave, sequenced, and constrained. You won’t find this level of production prompt engineering on Medium or YouTube — ever.
The tool system — every tool Claude Code uses (BashTool, FileEditTool, search_source, etc.) with its own schema, permission model, and execution logic. Blueprint for anyone building their own agent.
The bridges — real integration code for VS Code, Cursor, and JetBrains. Shows exactly how a CLI agent hooks into IDEs.
The SKILL.md at the repo root — Anthropic’s own internal guide for how the codebase is organized. Reading this first = instant orientation.
Runnable MCP server — the repo ships claude-code-explorer-mcp on npm. Use Claude Desktop or Cursor to explore the leaked Claude source. Meta loop.
⚙️ The Tech Stack
TypeScript (strict) · Bun · React · Ink · MCP SDK · ~1,900 files
Published: 2026-03-31 (made public domain by an Anthropic employee after discovery)
How it got out: the published npm package included a .map file pointing to the full unobfuscated TypeScript source in Anthropic’s R2 storage bucket. @Fried_rice caught it. Anthropic then released it officially.
📥 Grab It
| Source | Link | Status |
|---|---|---|
| Repo (main) | github.com/tjcrims0nx/claude-code | Re-assembled, browsable |
| Backup branch | backup branch | Original unmodified source |
| ZIP mirror | Google Drive | 9.48 MB |
| VirusTotal | 0 / 63 detections | Clean |
The repo is mirror-at-your-own-risk. Clone it locally if you plan to study it — GitHub DMCA’d copies can vanish overnight.
🧪 What You Can Actually Do With It
Read it — it’s TypeScript, strict mode, well-commented. You’ll learn more in an afternoon than six months of agent-building tutorials.
Fork it — for personal study, for research papers, for your own experiments.
Prompt-mine it — extract the system prompts and study how Anthropic phrases constraints, priorities, and tool-use instructions.
Plug it in — run the included MCP server, point Claude Desktop or Cursor at the source, ask it to explain its own architecture.
What you shouldn’t do — repackage it as a commercial product, strip the attribution, or publish a “rebranded Claude Code.” That’s how mirrors get nuked fast.
Quick Hits
| Want | Do |
|---|---|
 Browse the source |
Gittea repo |
 Download ZIP |
Drive mirror |
 Verify clean |
VirusTotal 0/63 |
 Start here |
Read SKILL.md → src/main.tsx → prompts/ |
 Archive it local |
git clone before any DMCA |
You’re not “downloading a leak.” You’re sitting in the cockpit of an AI agent that shipped to millions. Read the manuals.
!