🔓 Claude Code — Complete Decompiled Source Code

The full internal source code of Anthropic's Claude Code CLI (v1.0.x), reverse-engineered and decompiled from the official npm package.

---

⚠️ Disclaimer

This repository contains decompiled source code from Anthropic's proprietary Claude Code product. It is published for educational and research purposes only. All intellectual property belongs to Anthropic, PBC. This code is not licensed for redistribution, modification, or commercial use. Use responsibly.

---

📌 What is Claude Code?

Claude Code is Anthropic's agentic AI coding assistant — a terminal-based CLI tool that integrates the Claude AI model directly into your development workflow. It can:

• Read, write, and edit files autonomously
• Execute shell commands with intelligent sandboxing
• Search codebases using grep, glob, and ripgrep
• Interact with MCP (Model Context Protocol) servers
• Manage multi-agent workflows ("swarms")
• Connect to IDEs (VS Code, JetBrains, etc.)
• Resume, branch, and fork conversations
• Operate in interactive (REPL) and headless (SDK) modes

This repository contains the complete decompiled src/ directory — every internal module, every tool, every prompt, every hook.

---

🏗️ Architecture Overview

src/
├── main.tsx                 # CLI entrypoint — argument parsing, migrations, session bootstrap
├── query.ts                 # Core agentic query loop — the heart of Claude Code
├── QueryEngine.ts           # Headless query engine for SDK/non-interactive modes
├── Tool.ts                  # Tool type system — defines the interface for ALL tools
├── commands.ts              # Slash-command registry and loader
├── context.ts               # System & user context builders
├── cost-tracker.ts          # Token usage & cost tracking
├── history.ts               # Session history management
├── setup.ts                 # First-run setup and onboarding
├── tools.ts                 # Tool pool assembly and filtering
├── tasks.ts                 # Background task definitions
│
├── screens/                 # Full-screen TUI screens
│   ├── REPL.tsx             # The main REPL (895KB, 5,006 lines — the largest file)
│   ├── Doctor.tsx           # Diagnostic screen
│   └── ResumeConversation.tsx
│
├── tools/                   # All 40+ built-in tools
│   ├── BashTool/            # Shell command execution (with sandbox)
│   ├── FileEditTool/        # AI-powered file editing
│   ├── FileReadTool/        # File reading with range support
│   ├── FileWriteTool/       # Full file creation/overwrite
│   ├── GrepTool/            # Ripgrep-powered code search
│   ├── GlobTool/            # File pattern matching
│   ├── AgentTool/           # Sub-agent spawning (multi-agent)
│   ├── WebSearchTool/       # Web search capability
│   ├── WebFetchTool/        # URL content fetching
│   ├── MCPTool/             # MCP server tool proxying
│   ├── NotebookEditTool/    # Jupyter notebook editing
│   ├── REPLTool/            # REPL tool execution
│   ├── LSPTool/             # Language Server Protocol integration
│   ├── TaskCreateTool/      # Background task creation
│   ├── TeamCreateTool/      # Multi-agent team creation
│   ├── TodoWriteTool/       # TODO list management
│   ├── SkillTool/           # Skill/workflow invocation
│   ├── SleepTool/           # Timed delay tool
│   ├── ScheduleCronTool/    # Cron job scheduling
│   ├── SendMessageTool/     # Inter-agent messaging
│   └── ... (40+ tools total)
│
├── components/              # 140+ Ink (React-for-CLI) UI components
│   ├── Messages.tsx         # Conversation message renderer
│   ├── PromptInput/         # User input with autocomplete, vim mode
│   ├── StatusLine.tsx       # Bottom status bar
│   ├── Spinner.tsx          # Activity spinner with tips
│   ├── ModelPicker.tsx      # Model selection dialog
│   ├── ThemePicker.tsx      # Theme customization
│   ├── PermissionRequest.tsx # Tool permission dialogs
│   ├── StructuredDiff/      # Side-by-side diff viewer
│   ├── VirtualMessageList.tsx # Virtualized scrolling message list
│   ├── ScrollKeybindingHandler.tsx # Keyboard navigation
│   ├── ConsoleOAuthFlow.tsx # OAuth authentication flow
│   └── ... (140+ components)
│
├── hooks/                   # 85+ React hooks
│   ├── useCanUseTool.tsx    # Tool permission checking
│   ├── useReplBridge.tsx    # Remote control bridge (115KB!)
│   ├── useVoiceIntegration.tsx # Voice mode (99KB!)
│   ├── useTypeahead.tsx     # Autocomplete engine (212KB!)
│   ├── useGlobalKeybindings.tsx # Keyboard shortcut system
│   └── ... (85+ hooks)
│
├── utils/                   # 300+ utility modules
│   ├── auth.ts              # Authentication (OAuth, API keys, keychain)
│   ├── config.ts            # Configuration management
│   ├── hooks.ts             # Hook lifecycle system (159KB!)
│   ├── messages.ts          # Message construction & normalization (193KB!)
│   ├── sessionStorage.ts    # Session persistence (180KB!)
│   ├── teleport.tsx         # Session teleportation (cloud sync, 175KB!)
│   ├── attachments.ts       # File/memory attachments (127KB!)
│   ├── worktree.ts          # Git worktree operations
│   ├── permissions/         # Permission system
│   ├── model/               # Model selection & capabilities
│   ├── settings/            # Settings management
│   ├── swarm/               # Multi-agent swarm system
│   ├── mcp/                 # MCP client utilities
│   ├── telemetry/           # Analytics & tracing
│   ├── sandbox/             # Command sandboxing
│   └── ... (300+ files)
│
├── services/                # Backend service integrations
│   ├── mcp/                 # Model Context Protocol client
│   ├── oauth/               # OAuth 2.0 flow
│   ├── analytics/           # GrowthBook, Statsig, telemetry
│   ├── lsp/                 # Language Server Protocol client
│   ├── compact/             # Auto-compaction (context management)
│   ├── api/                 # Anthropic API client
│   ├── plugins/             # Plugin system
│   ├── tips/                # User tips & suggestions
│   └── voice.ts             # Voice input/STT
│
├── commands/                # 100+ slash commands (/command)
│   ├── init.ts              # /init — project initialization
│   ├── commit.ts            # /commit — git commit
│   ├── review.ts            # /review — code review
│   ├── compact/             # /compact — context compaction
│   ├── doctor/              # /doctor — diagnostics
│   ├── mcp/                 # /mcp — MCP server management
│   ├── model/               # /model — model switching
│   ├── teleport/            # /teleport — session sync
│   ├── ultraplan.tsx         # /ultraplan — advanced planning
│   ├── insights.ts          # /insights — usage analytics (113KB!)
│   └── ... (100+ commands)
│
├── bridge/                  # Remote Control Bridge (mobile/web)
│   ├── replBridge.ts        # Bridge core (100KB!)
│   ├── bridgeMain.ts        # Bridge lifecycle (115KB!)
│   ├── remoteBridgeCore.ts  # Remote session bridging
│   ├── jwtUtils.ts          # JWT authentication
│   └── trustedDevice.ts     # Device trust management
│
├── state/                   # Global state management
│   ├── AppState.tsx         # React context-based state
│   ├── AppStateStore.ts     # State store definition
│   └── store.ts             # Zustand-like store
│
├── types/                   # TypeScript type definitions
│   ├── message.ts           # Message types
│   ├── permissions.ts       # Permission types
│   ├── hooks.ts             # Hook event types
│   └── plugin.ts            # Plugin types
│
├── entrypoints/             # Multiple entry points
│   ├── cli.tsx              # Interactive CLI entrypoint
│   ├── mcp.ts               # MCP server mode
│   ├── sdk/                 # SDK entrypoint
│   └── init.ts              # Initialization
│
├── coordinator/             # Multi-agent coordinator mode
│   └── coordinatorMode.ts   # Orchestrates agent teams
│
├── plugins/                 # Plugin system
│   ├── builtinPlugins.ts    # Built-in plugin loader
│   └── bundled/             # Bundled plugins
│
├── skills/                  # Skill system
│   ├── bundledSkills.ts     # Built-in skills registry
│   ├── loadSkillsDir.ts     # Dynamic skill loader
│   └── bundled/             # Bundled skill definitions
│
├── migrations/              # Settings/config migrations
│   ├── migrateSonnet45ToSonnet46.ts
│   ├── migrateOpusToOpus1m.ts
│   └── ... (11 migration scripts)
│
├── vim/                     # Vim mode implementation
│   ├── motions.ts           # Cursor motions
│   ├── operators.ts         # Vim operators
│   ├── textObjects.ts       # Text objects
│   └── transitions.ts       # Mode transitions
│
├── buddy/                   # Companion sprite system (Easter egg)
│   ├── CompanionSprite.tsx  # Animated ASCII companion
│   ├── sprites.ts           # Sprite definitions
│   └── companion.ts         # Companion logic
│
├── voice/                   # Voice mode
│   └── voiceModeEnabled.ts  # Voice feature gate
│
├── ink/                     # Custom Ink (React terminal) extensions
├── keybindings/             # Keybinding system
├── outputStyles/            # Output formatting styles
├── schemas/                 # Zod schemas
├── constants/               # Constants and configuration
├── context/                 # React contexts
├── native-ts/               # Native TypeScript helpers
├── server/                  # Direct-connect server
├── remote/                  # Remote session management
├── upstreamproxy/           # Upstream proxy support
├── moreright/               # Layout system
└── memdir/                  # In-memory directory abstraction

---

🔑 Key Findings & Revelations

1. 🧠 The Query Loop (query.ts)

The core of Claude Code is an async generator function query() that:

• Manages the agentic loop (model call → tool execution → model call → ...)
• Handles auto-compaction when context gets too large
• Implements reactive compaction for prompt_too_long errors
• Supports max_output_tokens recovery with automatic continuation
• Manages microcompact (cache-aware context shrinking)
• Tracks token budgets (task_budget API parameter)
• Handles streaming tool execution (parallel tool calls)

2. 🔒 Anti-Debugging Protection (main.tsx)

Claude Code actively detects and terminates if running under a debugger:

if ("external" !== 'ant' && isBeingDebugged()) {
  process.exit(1);
}

It checks for --inspect, --debug, Node.js inspector API, and NODE_OPTIONS environment variables. Internal Anthropic builds ("ant") bypass this check.

3. 🏷️ Internal vs External Builds

The codebase uses bun:bundle feature flags extensively:

• feature('COORDINATOR_MODE') — Multi-agent orchestration
• feature('KAIROS') — Assistant mode (proactive agent)
• feature('VOICE_MODE') — Voice input
• feature('BRIDGE_MODE') — Mobile/web remote control
• feature('SSH_REMOTE') — SSH remote execution
• feature('DIRECT_CONNECT') — Direct server connections
• feature('ULTRAPLAN') — Advanced planning mode
• feature('BUDDY') — Companion sprite Easter egg
• feature('FORK_SUBAGENT') — Agent forking
• feature('TRANSCRIPT_CLASSIFIER') — Auto-mode safety classifier
• feature('EXPERIMENTAL_SKILL_SEARCH') — Skill discovery
• feature('WEB_BROWSER_TOOL') — Browser automation tool

Code marked "external" === 'ant' is Anthropic-internal only and dead-code-eliminated from public builds.

4. 🔐 Authentication System (utils/auth.ts)

Claude Code supports multiple authentication methods:

• OAuth 2.0 via claude.ai subscription
• API Keys from ANTHROPIC_API_KEY environment variable
• macOS Keychain integration for secure key storage
• API Key Helper — custom scripts that provide API keys on demand
• File Descriptor — keys passed via fd (for subprocess isolation)
• AWS STS — Bedrock authentication with credential refresh
• GCP — Vertex AI authentication
• Foundry — Azure Foundry authentication

5. 🛡️ Permission System (utils/permissions/)

A sophisticated multi-level permission system:

• Permission Modes: default, plan, auto, bypassPermissions
• Auto-Mode Classifier: A transcript-based safety classifier that auto-approves safe actions
• Hook-Based Permissions: External scripts can approve/deny tool usage
• Always Allow/Deny Rules: Configurable per-tool permission patterns
• Sandbox: Command sandboxing using OS-level primitives
• Trust Dialog: First-run workspace trust confirmation

6. 📡 Remote Control Bridge (bridge/)

Claude Code can be controlled remotely from mobile/web:

• WebSocket-based real-time communication
• JWT authentication for device trust
• Capacity wake system for mobile push notifications
• Message bridging between local and remote sessions
• QR code generation for mobile connection

7. 🤖 Multi-Agent System (utils/swarm/, tools/AgentTool/)

Full multi-agent orchestration support:

• Agent Types: Custom agents via .claude/agents/ directory
• Team Creation: TeamCreateTool spawns agent teams
• Inter-Agent Communication: SendMessageTool for mailbox messaging
• Coordinator Mode: Central orchestrator that delegates to workers
• In-Process Teammates: Agents run in the same process (lightweight)
• tmux/iTerm2 Backends: Agents in separate terminal panes

8. 💰 Cost & Token Tracking (cost-tracker.ts)

Detailed financial tracking:

• Per-model token usage (input, output, cache read, cache creation)
• USD cost calculation using internal pricing tables
• Web search request counting
• Session cost persistence and restoration on resume
• Cost threshold dialogs (warns at spending limits)

9. 🔄 Context Management (services/compact/)

Four-tier context management to stay within model limits:

1. Snip Compact — Removes old messages from the middle
2. Microcompact — Cache-aware content shrinking (preserves cache keys)
3. Auto Compact — Full summarization of conversation history
4. Reactive Compact — Emergency compaction on API prompt_too_long errors
5. Context Collapse — Staged collapsing of tool-use sequences

10. 📝 Session Storage (utils/sessionStorage.ts, 180KB!)

Comprehensive session persistence:

• JSONL-based transcript storage
• Session search by title, content, and metadata
• File history snapshots for undo/redo
• Attribution tracking for git commit decoration
• Content replacement records for resumed sessions
• Background session management (concurrent sessions)

11. 🎨 UI System (Ink/React for Terminal)

A full React-based terminal UI built on Ink:

• Virtual scrolling message list
• Structured diff viewer
• Model picker dialog
• Theme system with multiple themes
• Vim mode with full motion/operator/text-object support
• Fullscreen layout with alternate screen buffer
• Animated spinner with contextual tips

12. 🔌 Plugin System (utils/plugins/)

Extensible plugin architecture:

• DXT format — Declarative extension toolkit
• MCP servers — Model Context Protocol integration
• Bundled plugins — Built-in extensions (GitHub, Slack, etc.)
• Marketplace — Official plugin marketplace support
• Version management — Automatic updates and orphan cleanup

13. 🎙️ Voice Mode (services/voice.ts, hooks/useVoiceIntegration.tsx)

Full voice input support:

• Speech-to-Text streaming
• Voice keyterm detection
• Integration with the prompt input system
• 99KB of voice integration hooks

14. 🌐 Teleport System (utils/teleport.tsx, 175KB!)

Cloud-based session synchronization:

• Transfer sessions between machines
• Git-based session state serialization
• Branch checkout for teleported sessions
• Error recovery and retry logic

15. 📊 Analytics & Telemetry (services/analytics/)

Comprehensive analytics infrastructure:

• GrowthBook — Feature flag & A/B testing
• Statsig — Event logging
• OpenTelemetry — Session tracing with Perfetto export
• BigQuery exporter — Internal data pipeline
• Diagnostic logging — PII-free diagnostic events

16. 🏗️ Model Migrations (migrations/)

Automated model string migrations:

• Sonnet 4.5 → Sonnet 4.6
• Opus → Opus 1M
• Fennec → Opus (internal codename)
• Legacy Opus → Current Opus
• Model-specific config updates

---

📈 Codebase Statistics

Metric	Value
Total Files	1,902
TypeScript/TSX Files	1,884
Total Lines of Code	512,664
Largest File	screens/REPL.tsx (895 KB, 5,006 lines)
Main Entry Point	main.tsx (803 KB, 4,684 lines)
Built-in Tools	40+
Slash Commands	100+
UI Components	140+
React Hooks	85+
Utility Modules	300+

---

🛠️ Technology Stack

Technology	Purpose
TypeScript	Primary language
Bun	Runtime & bundler (via bun:bundle feature flags)
React / Ink	Terminal UI framework
Anthropic SDK	Claude API client
Commander.js	CLI argument parsing
Zod	Runtime type validation
chalk	Terminal coloring
execa	Child process execution
lodash-es	Utility functions
MCP SDK	Model Context Protocol
GrowthBook	Feature flags

---

🔍 Internal Codenames Discovered

Codename	Meaning
Tengu	Analytics event prefix (tengu_*) — Claude Code's internal name
Kairos	Assistant/proactive mode
Fennec	An earlier Claude model variant
Lodestone	Deep link / URL protocol handler
Tungsten	An experimental tool type
Walrus	Reactive compact feature
Torch	An experimental command
Ultraplan	Advanced multi-step planning
Grove	A UI component system
Buddy	Animated companion sprite

---

🚀 How This Code Was Obtained

Claude Code is distributed as a compiled JavaScript bundle via npm (@anthropic-ai/claude-code). The source was obtained by:

1. Installing the official npm package
2. Extracting the bundled JavaScript
3. Decompiling/reformatting the minified source
4. Reconstructing the module structure from import paths
5. Recovering TypeScript types from the build artifacts

The code retains original comments, function names, and module structure — Anthropic did not obfuscate the source beyond standard bundling.

---

⚖️ Legal Notice

This code is the intellectual property of Anthropic, PBC. It is published here for:

• Security research and vulnerability analysis
• Understanding AI agent architectures
• Educational purposes

Do NOT use this code to:

• Build competing products
• Bypass security or licensing mechanisms
• Distribute modified versions commercially

If Anthropic requests removal, this repository will be taken down immediately.

---

🤝 Contributing

This is a static snapshot of the decompiled source. No contributions to the source code are accepted. Issues documenting findings, architecture insights, or security observations are welcome.

---

📜 License

This code is NOT open source. It is proprietary software owned by Anthropic, PBC. No license is granted for use, modification, or distribution.

---

⭐ Star this repo if you found it insightful
Understanding how AI agents work internally is key to building better ones.