Claude Code Source Code Reconstruction

Complete TypeScript source code of the Claude Code CLI, reconstructed from source maps and runnable locally.

Warning

This repository is an unofficial version, reconstructed from the source map of the public npm release package. It is for research and learning purposes only and does not represent the internal development repository structure of Anthropic. Some modules have been replaced with compatible shims.

---

Requirements

• Bun ≥ 1.3.5
• Node.js ≥ 24

Quickstart

$ bun install       # Install dependencies
$ bun run dev       # Start CLI (Interactive)
$ bun run version   # Verify version number

---

Project Structure

.
├── src/                # Core source code (~2,006 TS/TSX files)
│   ├── entrypoints/    # CLI entry points
│   ├── main.tsx        # Main initialization (auth / MCP / settings / feature flags)
│   ├── dev-entry.ts    # Development entry point
│   ├── QueryEngine.ts  # Core engine (~1,295 lines, LLM API loop, persistence)
│   │
│   ├── tools/          # Tool implementations (53 items: Bash, Read, Edit, Agent...)
│   ├── commands/       # Slash commands (87 items)
│   ├── services/       # Backend services (API, MCP, OAuth, telemetry/Datadog)
│   ├── utils/          # Utility functions (git, permissions, model, token budget)
│   │
│   ├── components/     # Terminal UI components (~406 files, React + Ink)
│   ├── hooks/          # Custom React Hooks
│   ├── ink/            # Ink terminal renderer (custom branch)
│   ├── vim/            # Vim mode engine
│   ├── keybindings/    # Keybindings
│   │
│   ├── coordinator/    # Multi-Agent orchestration and worker coordination
│   ├── bridge/         # IDE bidirectional communication & remote bridge control
│   ├── remote/         # Remote session teleportation & management
│   ├── server/         # IDE direct connection server
│   ├── skills/         # Reusable workflow & skill system
│   ├── plugins/        # Plugin system
│   ├── memdir/         # Persistent memory system (5-tier memory)
│   ├── voice/          # Voice interaction (streaming STT, unreleased)
│   ├── buddy/          # Gacha companion sprite system (Easter egg)
│   └── assistant/      # "KAIROS" always-running daemon mode (unreleased)
│
├── shims/              # Native module compatibility alternatives
├── vendor/             # Native binding source code
├── package.json
├── tsconfig.json
└── bun.lock

---

Architecture

Claude Code is built on top of a highly optimized and robust architecture designed for LLM API interaction, token efficiency, and advanced execution boundaries.

Boot Sequence

dev-entry.ts → entrypoints/cli.tsx → main.tsx → REPL (React/Ink)
  │                │          │
  │                │          └─ Full Initialization: Auth → GrowthBook (Feature Flags) → MCP → Settings → Commander.js
  │                └─ Fast Path: --version / daemon / ps / logs
  └─ Startup Gate: scans for missing relative imports; blocks boot until all resolve

Core Engine & Token Optimization

Token efficiency is critical for survival in Claude Code. The architecture employs industry-leading token saving techniques:

• QueryEngine.ts: The central engine (~1,295 lines) managing the LLM API loop, session lifecycle, and automatic tool execution.
• 3-Tier Compaction System:
  1. Microcompact: Uses the cache_edits API to remove messages from the server cache without invalidating the prompt cache context (zero API cost).
  2. Session Memory: Uses pre-extracted session memory as a summary to avoid LLM calls during mid-level compaction.
  3. Full Compact: Instructs a sub-agent to summarize the conversation into a structured 9-section format, employing <analysis> tag stripping to reduce token usage while maintaining quality.
• Advanced Optimizations:
  • FILE_UNCHANGED_STUB: Returns a brief 30-word stub for re-read files.
  • Dynamic max output caps (8K default with 64K retry) preventing slot-reservation waste.
  • Caching latches to prevent UI toggles limits (e.g., Shift+Tab) from busting 70K context.
  • Circuit breakers preventing wasted API calls on consecutive compaction failures.

Harness Engineering (Permissions & Security)

The "Harness" safely controls LLM operations within the local environment:

• Permission Modes: Features 6 primary modes (acceptEdits, bypassPermissions, default, dontAsk, plan) plus internal designations like auto (yoloClassifier) and bubble (sub-agent propagation).
• Security Checkers: Incorporates PowerShell-specific security analysis to detect command injection, download cradles, and privilege escalation, as well as redundant path validations.
• Architectural Bypasses: Specific environments intentionally bypass checks (e.g., CLAUDE_CODE_SIMPLE clears system prompts), while failing schema parsing can inadvertently circumvent standard permissions.

Teams & Multi-Agent Orchestration

• Agents: Orchestrated via AgentTool, created with three distinct paths: Teammate (tmux or in-process), Fork (inheriting context), and Normal (fresh context).
• Coordinator Mode: A designated coordinator delegates exact coding tasks to worker agents (Agent, SendMessage, TaskStop), effectively isolating high-level reasoning from raw file execution.

Memory System (5-Tier Architecture)

Designed to persist AI knowledge across sessions and agents:

1. Memdir: Project-level indices and topic files (MEMORY.md).
2. Auto Extract: Fire-and-forget forked agent that consolidates memory post-session.
3. Session Memory: Real-time context tracking without extra LLM overhead.
4. Team Memory: Shared remote state leveraging SHA-256 delta uploads and git-leaks-based secret extraction guards.
5. Agent Memory: Agent-specific knowledge scoped to local, project, or user levels.

Unreleased Subsystems & Future Directions

Hidden behind 88+ compile-time feature flags and 700+ GrowthBook runtime gates:

• KAIROS: An always-running background daemon featuring a "Dream" mode (autonomous memory consolidation during idle time).
• Computer Use ("Chicago"): macOS desktop control MCP (mouse, keyboard, screenshot capabilities).
• Voice Mode: Microphone control utilizing streaming STT.
• ULTRAPLAN: Capable of executing multi-agent planning over 30-minute CCR sessions.
• Web Browser ("Bagel") & Teleport: Integrated web navigation and remote session context teleportation.

UI Architecture

Terminal UI based on React + Ink:

• ink/ — Custom Ink branch (layout / focus / ANSI / virtual scrolling / click detection)
• components/ (~406 files) — Messages, inputs, diffs, permission dialogs, status bar
• hooks/ — Tools / voice / IDE / vim / sessions / tasks
• vim/ — Full Vim keybinding engine (motions, operators, text objects)

---

Reconstruction Notes

Source maps cannot 100% reconstruct the original repository. The following may be missing or degraded:

Type	Description
Type-only files	Type-only .d.ts files may be missing
Build artifacts	Code generated during the build is not in the source map
Native bindings	Private native modules are replaced with shims/
Dynamic resources	Dynamic imports and resource files may be incomplete

---

Patches

This fork includes the following modifications from the original source:

Patch	Description	Related Issue
Welcome banner toggle	Added showWelcomeBanner setting to disable the startup banner (LogoV2). Set "showWelcomeBanner": false in ~/.claude/settings.json to hide.	#2254

---

Disclaimer

• The source code copyright belongs to Anthropic.
• This is for technical research and learning purposes only. Please do not use it for commercial purposes.