Start Claude Code

Run Claude Code v2.1.88 from leaked source — one command setup. Both interactive TUI and non-interactive modes work.

GitHub: https://github.com/JiaranI/start-claude-code

Quick Start

git clone https://github.com/JiaranI/start-claude-code.git
cd start-claude-code

# Set API key (or use existing OAuth login from official Claude Code)
export ANTHROPIC_API_KEY="sk-ant-xxx"

# Run!
./start.sh

First run will automatically install Bun, dependencies, and configure everything.

---

Usage

Interactive TUI (full terminal UI)

# With API key
export ANTHROPIC_API_KEY="sk-ant-xxx"
./start.sh --dangerously-skip-permissions

# Or with existing OAuth login (if you've used official Claude Code before)
./start.sh --dangerously-skip-permissions

Non-interactive mode (scripts/pipes)

export ANTHROPIC_API_KEY="sk-ant-xxx"
./start.sh -p "explain this code" --dangerously-skip-permissions < /dev/null

Use a third-party proxy

export ANTHROPIC_BASE_URL="https://your-proxy.com"   # Don't include /v1
export ANTHROPIC_API_KEY="your-key"
export DISABLE_PROMPT_CACHING=1
export DISABLE_INTERLEAVED_THINKING=1
export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1

# Non-interactive (recommended for proxies)
./start.sh -p "hello" --model claude-sonnet-4-20250514 \
  --dangerously-skip-permissions --no-session-persistence < /dev/null

# Interactive TUI with proxy
./start.sh --model claude-sonnet-4-20250514 --dangerously-skip-permissions --bare

Specify model

./start.sh --model claude-sonnet-4-20250514
./start.sh --model claude-opus-4-20250514
./start.sh --model claude-haiku-4-5-20241022

---

Manual Setup

If you prefer to set up step by step:

# 1. Install Bun
curl -fsSL https://bun.sh/install | bash
export PATH="$HOME/.bun/bin:$PATH"

# 2. Run setup
node scripts/setup.mjs

# 3. Run
bun src/entrypoints/cli.tsx --dangerously-skip-permissions

---

Environment Variables

Variable	Description	Default
ANTHROPIC_API_KEY	API key (not needed if using OAuth)	-
ANTHROPIC_BASE_URL	API base URL (no /v1 suffix)	https://api.anthropic.com
ANTHROPIC_MODEL	Default model	claude-sonnet-4-6
DISABLE_PROMPT_CACHING	Disable prompt caching (needed for most proxies)	0
DISABLE_INTERLEAVED_THINKING	Disable interleaved thinking (needed for some proxies)	0
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS	Disable experimental beta headers (needed for some proxies)	0
CLAUDE_CODE_FORCE_FULL_LOGO	Show full welcome screen with border and tips	0

---

Status

Feature	Status
--version / --help	Working
-p non-interactive mode	Working (full API call + tool use)
Interactive TUI	Working (full UI rendering + input)
OAuth login	Working (reuses existing official Claude Code login)
Third-party proxy	Working (set ANTHROPIC_BASE_URL)

---

What is this?

On March 31, 2026, the full source code of Anthropic's Claude Code CLI was leaked via a .map file in their npm registry. This repository contains:

• The complete src/ directory (~1,900 files, 512,000+ lines of TypeScript)
• Build shims and stubs to make it runnable from source
• A one-click setup script

Source authenticity: Verified by comparing all 1,902 source files against @anthropic-ai/claude-code@2.1.88's source map — 100% byte-identical match, zero differences.

What was leaked vs what we had to reconstruct

The leak only contains src/ — raw TypeScript before compilation. Running from source required reverse-engineering the missing build infrastructure:

Missing piece	What it is	Our workaround
bun:bundle	Compile-time feature flag API (dead code elimination)	Runtime shim returning false for all flags
MACRO.*	Build-time constant injection (MACRO.VERSION etc.)	Global variable definition in preload
package.json	Dependency declarations (94 packages)	Reverse-engineered from import statements
Private npm packages	@anthropic-ai/sandbox-runtime, @ant/*, @anthropic-ai/mcpb	Empty stub modules with fake exports
Generated files	coreTypes.generated.ts etc. (built by internal scripts)	Manually created type stubs
Feature-gated files	Files deleted at compile time (connectorText.ts, TungstenTool/*)	Empty stub files
Keybinding dispatch	registerHandler callbacks gated behind chord-only check	Removed wasInChord guard

Tech Stack

Component	Technology
Runtime	Bun
Language	TypeScript
Terminal UI	React + Ink (custom fork)
CLI Parser	Commander.js
API	Anthropic SDK
Code Search	ripgrep (bundled)
Protocols	MCP, LSP

Source Structure

src/
├── main.tsx              # CLI entrypoint
├── QueryEngine.ts        # Core LLM API caller
├── tools/                # ~40 agent tools (Bash, Edit, Read, etc.)
├── commands/             # ~50 slash commands (/commit, /review, etc.)
├── components/           # ~140 Ink UI components
├── services/             # API, MCP, OAuth, analytics
├── bridge/               # IDE integration (VS Code, JetBrains)
├── coordinator/          # Multi-agent orchestration
├── skills/               # Skill system
├── plugins/              # Plugin system
├── memdir/               # Persistent memory
├── buddy/                # Companion sprites (19 animals!)
└── ...

---

Troubleshooting

"API Error: 400 ... invalid beta flag"

Your proxy doesn't support Claude beta headers. Set:

export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1
export DISABLE_INTERLEAVED_THINKING=1

Or use --bare mode which skips most features.

"API Error: 400 ... cache_control"

export DISABLE_PROMPT_CACHING=1

"API Error: 400 ... /v1/v1/messages"

Your ANTHROPIC_BASE_URL should NOT include /v1:

# Wrong
export ANTHROPIC_BASE_URL="https://proxy.com/v1"

# Correct
export ANTHROPIC_BASE_URL="https://proxy.com"

"API Error: 401" in interactive mode with proxy

Interactive mode may use OAuth tokens from a previous official Claude Code login. Use --bare to force API key auth:

./start.sh --dangerously-skip-permissions --bare

Stuck / no output in non-interactive mode

Always pipe /dev/null to stdin:

./start.sh -p "hello" --dangerously-skip-permissions < /dev/null

---

Disclaimer

This repository archives source code that was leaked from Anthropic's npm registry. All original source code is the property of Anthropic.