Claude Code — Leaked Source (2026-03-31)

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

Chaofan Shou (@Fried_rice) discovered the leak and posted it publicly:

"Claude code source code has been leaked via a map file in their npm registry!"

— @Fried_rice, March 31, 2026

The source map file in the published npm package contained a reference to the full, unobfuscated TypeScript source, which was downloadable as a zip archive from Anthropic's R2 storage bucket.

Claude Code is Anthropic's official CLI tool that lets you interact with Claude directly from the terminal to perform software engineering tasks — editing files, running commands, searching codebases, managing git workflows, and more.

This repository contains the leaked src/ directory.

• Leaked on: 2026-03-31
• Language: TypeScript
• Runtime: Bun
• Terminal UI: React + Ink (React for CLI)
• Scale: ~1,900 files, 512,000+ lines of code

src/
├── main.tsx                 # Entrypoint (Commander.js-based CLI parser)
├── commands.ts              # Command registry
├── tools.ts                 # Tool registry
├── Tool.ts                  # Tool type definitions
├── QueryEngine.ts           # LLM query engine (core Anthropic API caller)
├── context.ts               # System/user context collection
├── cost-tracker.ts          # Token cost tracking
│
├── commands/                # Slash command implementations (~50)
├── tools/                   # Agent tool implementations (~40)
├── components/              # Ink UI components (~140)
├── hooks/                   # React hooks
├── services/                # External service integrations
├── screens/                 # Full-screen UIs (Doctor, REPL, Resume)
├── types/                   # TypeScript type definitions
├── utils/                   # Utility functions
│
├── bridge/                  # IDE integration bridge (VS Code, JetBrains)
├── coordinator/             # Multi-agent coordinator
├── plugins/                 # Plugin system
├── skills/                  # Skill system
├── keybindings/             # Keybinding configuration
├── vim/                     # Vim mode
├── voice/                   # Voice input
├── remote/                  # Remote sessions
├── server/                  # Server mode
├── memdir/                  # Memory directory (persistent memory)
├── tasks/                   # Task management
├── state/                   # State management
├── migrations/              # Config migrations
├── schemas/                 # Config schemas (Zod)
├── entrypoints/             # Initialization logic
├── ink/                     # Ink renderer wrapper
├── buddy/                   # Companion sprite (Easter egg)
├── native-ts/               # Native TypeScript utils
├── outputStyles/            # Output styling
├── query/                   # Query pipeline
└── upstreamproxy/           # Proxy configuration

Every tool Claude Code can invoke is implemented as a self-contained module. Each tool defines its input schema, permission model, and execution logic.

User-facing slash commands invoked with / prefix.

A bidirectional communication layer connecting IDE extensions (VS Code, JetBrains) with the Claude Code CLI.

• bridgeMain.ts — Bridge main loop
• bridgeMessaging.ts — Message protocol
• bridgePermissionCallbacks.ts — Permission callbacks
• replBridge.ts — REPL session bridge
• jwtUtils.ts — JWT-based authentication
• sessionRunner.ts — Session execution management

Checks permissions on every tool invocation. Either prompts the user for approval/denial or automatically resolves based on the configured permission mode (default, plan, bypassPermissions, auto, etc.).

Dead code elimination via Bun's bun:bundle feature flags:

import { feature } from 'bun:bundle'

// Inactive code is completely stripped at build time
const voiceCommand = feature('VOICE_MODE')
  ? require('./commands/voice/index.js').default
  : null

Notable flags: PROACTIVE, KAIROS, BRIDGE_MODE, DAEMON, VOICE_MODE, AGENT_TRIGGERS, MONITOR_TOOL

The core engine for LLM API calls. Handles streaming responses, tool-call loops, thinking mode, retry logic, and token counting.

Defines base types and interfaces for all tools — input schemas, permission models, and progress state types.

Manages registration and execution of all slash commands. Uses conditional imports to load different command sets per environment.

Commander.js-based CLI parser + React/Ink renderer initialization. At startup, parallelizes MDM settings, keychain prefetch, and GrowthBook initialization for faster boot.

Startup time is optimized by prefetching MDM settings, keychain reads, and API preconnect in parallel — before heavy module evaluation begins.

// main.tsx — fired as side-effects before other imports
startMdmRawRead()
startKeychainPrefetch()

Heavy modules (OpenTelemetry ~400KB, gRPC ~700KB) are deferred via dynamic import() until actually needed.

Sub-agents are spawned via AgentTool, with coordinator/ handling multi-agent orchestration. TeamCreateTool enables team-level parallel work.

Reusable workflows defined in skills/ and executed through SkillTool. Users can add custom skills.

Built-in and third-party plugins are loaded through the plugins/ subsystem.

This repository archives source code that was leaked from Anthropic's npm registry on 2026-03-31. All original source code is the property of Anthropic.