Production-grade AI Agent internals, dissected for learning.
This repository contains the extracted source code of Claude Code (Anthropic's AI programming CLI) along with 18 original deep-analysis articles (bilingual: English & Chinese) dissecting its architecture. The source was discovered from publicly available repositories and is provided strictly for educational and research purposes.
All intellectual property rights of the source code belong exclusively to Anthropic, PBC.
Disclaimer · Why Study This · Architecture · Deep Analysis · Run from Source · Tech Stack · Structure · License
IMPORTANT: READ BEFORE USE
This repository is strictly for educational and research purposes only.
- Source: The source code was discovered from publicly available sources, including instructkr/claw-code. The run-from-source setup is based on JiaranI/start-claude-code. It contains extracted source code of Claude Code, a product by Anthropic.
- Ownership: All intellectual property rights belong exclusively to Anthropic, PBC. This repository claims no ownership, authorship, or rights over the original code.
- Purpose: Provided solely as learning material for understanding AI agent system architecture and engineering patterns — for academic study, technical research, and educational discussion.
- Prohibited uses:
- Commercial purposes of any kind
- Building competing products or services
- Redistribution or repackaging of the source code
- Any purpose that violates Anthropic's Terms of Service or applicable laws
- Analysis articles: The deep analysis articles (
claude-code-deep-analysis/) are original commentary and analysis by the repository maintainers. Code snippets are included for commentary, criticism, and education purposes. These articles do not constitute legal advice. - DMCA / Takedown: If you represent Anthropic or believe this repository infringes on intellectual property rights, please contact us — we will promptly remove the infringing content. You may also file a DMCA takedown notice through GitHub's DMCA process.
- No warranty: Provided "as is" without warranty of any kind.
Claude Code is one of the most sophisticated production AI agent systems publicly observable. Unlike toy frameworks and demo agents, it is engineered for real-world, hours-long coding sessions with enterprise-grade reliability. Studying it reveals hard-won engineering decisions that no tutorial covers:
| Design Decision | Insight |
|---|---|
| Loop over Graph | A while(true) loop replaces DAG-based workflow orchestration — simpler, more flexible, easier to reason about at runtime |
| Recursion over Orchestration | Sub-agents recursively call query(), inheriting compression, error recovery, and streaming for free |
| Model decides, Framework executes | The framework enforces safety constraints (concurrency, side effects, permissions); all reasoning stays in the model |
| Built for 4-hour sessions | Four-layer context compression and three-level 413 error recovery — invisible in demos, critical in production |
| Immutability as cost optimization | Immutable API messages maximize prompt caching hits, reducing long-session costs by ~80% |
CLI / VS Code / IDE Extension
│
▼
┌───────────────────────┐
│ EntryPoint (cli.tsx) │
│ Fast-path routing │
└───────────┬───────────┘
│
┌───────────▼───────────┐
│ QueryEngine │
│ Session management │
│ API client wrapper │
└───────────┬───────────┘
│
┌───────────────────▼───────────────────┐
│ query() — Core Agent Loop │
│ ┌─────────────────────────────────┐ │
│ │ while (true) { │ │
│ │ messages = buildPrompt() │ │
│ │ response = stream(messages) │ │
│ │ tools = extractToolCalls() │ │
│ │ if (!tools) break │ │
│ │ results = executePar(tools) │ │
│ │ messages.push(results) │ │
│ │ maybeCompress(messages) │ │
│ │ } │ │
│ └─────────────────────────────────┘ │
└──────┬──────────┬──────────┬──────────┘
│ │ │
┌────────▼──┐ ┌────▼────┐ ┌───▼────────┐
│ 45+ Tools │ │ Permis- │ │ Context │
│ Bash, Edit │ │ sion │ │ Compression│
│ Glob, Grep │ │ System │ │ 4 layers │
│ MCP, LSP… │ │ 5 modes │ │ + 413 heal │
└────────────┘ └─────────┘ └────────────┘
| Metric | Count |
|---|---|
Core agent loop (query.ts) |
1,729 lines |
Main UI component (main.tsx) |
4,683 lines |
| Tool implementations | 45+ |
| Slash commands | 87+ |
| React UI components | 146+ |
| Utility functions | 564+ |
| React hooks | 85+ |
| Service modules | 38 |
| Feature flags | 89 |
System prompt assembly (prompts.ts) |
54.3 KB |
We produced 18 original articles with source-level deep dives into every major subsystem. Each article is available in both English and Chinese.
Enter the analysis series (EN) → ｜ 进入分析系列 (中文) →
| # | Topic | English | 中文 |
|---|---|---|---|
| 00 | Core Conclusions | EN | 中文 |
| 01 | Entry Point | EN | 中文 |
| 02 | Main Loop | EN | 中文 |
| 03 | Streaming | EN | 中文 |
| 04 | Tool Orchestration | EN | 中文 |
| 05 | Permission System | EN | 中文 |
| 06 | Sub-Agents | EN | 中文 |
| 07 | Context Window | EN | 中文 |
| 08 | Message Types | EN | 中文 |
| 09 | Immutable Messages | EN | 中文 |
| 10 | Architecture Diagram | EN | 中文 |
| 11 | Design Philosophy | EN | 中文 |
| # | Topic | English | 中文 |
|---|---|---|---|
| 12 | MCP Integration | EN | 中文 |
| 13 | Memory System | EN | 中文 |
| 14 | System Prompt | EN | 中文 |
| 15 | Session & Bridge | EN | 中文 |
| 16 | Tool Implementations | EN | 中文 |
| 17 | Hook System | EN | 中文 |
This repository supports running Claude Code locally from source (with reduced functionality). Verified working with version 2.1.88.
| Dependency | Notes |
|---|---|
| Node.js >= 18 | For the setup script |
| Bun >= 1.0 | Runtime for Claude Code (auto-installed by setup) |
| Authentication | Claude Pro/Max/Team subscription (OAuth) or API key (sk-ant-xxx) |
# 1. Install dependencies and generate shims
node scripts/setup.mjs
# 2. Login with your Claude subscription (opens browser)
./start.sh login
# 3. Launch
./start.sh
Running from source with Claude Pro/Max subscription — Opus 4.6 (1M context) ✓
Alternative: Use an API key instead
export ANTHROPIC_API_KEY="sk-ant-xxx"
./start.shWindows users
rem Use API key on Windows (OAuth login requires start.sh on macOS/Linux)
set ANTHROPIC_API_KEY=sk-ant-xxx
bun src/entrypoints/cli.tsx- ripgrep: The
tarextraction in the setup script may fail on Windows. Install ripgrep manually and ensurergis on your PATH. - Bun path: After installing Bun, restart your terminal or add
%USERPROFILE%\.bun\binto PATH. - OAuth login:
start.sh loginis macOS/Linux only. On Windows, runbun src/entrypoints/cli.tsx auth login --claudeai.
These files are not part of the original source — they were created to make the source runnable:
| File | Purpose |
|---|---|
package.json |
100+ dependencies reverse-engineered from import statements |
tsconfig.json |
TypeScript config with baseUrl and .js → .ts/.tsx path resolution |
bunfig.toml |
Bun config specifying the preload plugin |
preload.ts |
Core shim: bun:bundle mock (feature() returns false), MACRO.* global injection |
scripts/setup.mjs |
One-command setup: dependency install, private package stubs, missing file generation, ripgrep download |
start.sh |
macOS/Linux launcher with OAuth login (./start.sh login), credential detection (Keychain/file/env), proxy auto-detection |
| Challenge | Solution |
|---|---|
bun:bundle compile-time API |
preload.ts provides a runtime shim; feature() returns false for all flags |
| 89 feature flags | All disabled — feature-gated code paths do not execute |
MACRO.* compile-time macros |
Defined as globalThis.MACRO in preload.ts |
from 'src/...' imports |
Source under src/; tsconfig.json baseUrl: "." for natural resolution |
No package.json |
Reverse-engineered 100+ dependencies from import statements |
@ant/* private packages |
scripts/setup.mjs creates empty stubs |
| Feature | Missing Package | Description |
|---|---|---|
| Computer Use | @ant/computer-use-mcp |
Screenshots, mouse clicks, keyboard input |
| Native input | @ant/computer-use-input |
Rust/enigo native bindings |
| macOS screen/window | @ant/computer-use-swift |
Swift native bindings (macOS only) |
| Chrome integration | @ant/claude-for-chrome-mcp |
Chrome extension MCP server |
| Sandbox runtime | @anthropic-ai/sandbox-runtime |
Command execution sandbox |
| MCP Bridge | @anthropic-ai/mcpb |
MCP protocol bridge |
All 89 feature flags (all disabled)
Since feature() returns false at runtime, all feature-gated code paths are inactive:
ABLATION_BASELINE AGENT_MEMORY_SNAPSHOT AGENT_TRIGGERS BRIDGE_MODE BUDDY BUILDING_CLAUDE_APPS CCR_AUTO_CONNECT COORDINATOR_MODE DAEMON DIRECT_CONNECT DUMP_SYSTEM_PROMPT FORK_SUBAGENT HISTORY_PICKER KAIROS MCP_SKILLS MONITOR_TOOL NATIVE_CLIPBOARD_IMAGE PERFETTO_TRACING QUICK_SEARCH SSH_REMOTE STREAMLINED_OUTPUT TEAMMEM TEMPLATES TERMINAL_PANEL TORCH ULTRAPLAN ULTRATHINK VOICE_MODE WEB_BROWSER_TOOL WORKFLOW_SCRIPTS and 59 more.
TypeScript source
│
├─ Bun bundler (bun build)
│ ├─ Inject MACRO.* constants (--define)
│ ├─ Resolve feature() calls → set 89 feature flags true/false
│ ├─ Dead Code Elimination → strip disabled feature branches
│ └─ Bundle into single JS file
│
├─ Optional: Bun compile → single-file executable binary
│
└─ Publish to npm (@anthropic-ai/claude-code)
| Error | Fix |
|---|---|
Cannot find module 'src/...' |
Verify source is under src/ and bunfig.toml exists |
Missing 'default' export in module '*.md' |
Run node scripts/setup.mjs to regenerate stubs |
Cannot find package '@ant/...' |
Run node scripts/setup.mjs to recreate stubs |
bun: command not found |
Install Bun: curl -fsSL https://bun.sh/install | bash |
No authentication found |
Run ./start.sh login for OAuth, or export ANTHROPIC_API_KEY="sk-ant-xxx" for API key |
| Using a non-Anthropic proxy | start.sh auto-detects; manual: set DISABLE_PROMPT_CACHING=1 and DISABLE_INTERLEAVED_THINKING=1 |
npm install -g @anthropic-ai/claude-code| Layer | Technologies |
|---|---|
| Runtime | Node.js >= 18, Bun >= 1.0 |
| Language | TypeScript with React JSX |
| UI | React 19, custom terminal reconciler, 146+ components |
| CLI | Commander.js 12 |
| AI/LLM | Anthropic SDK, Claude Agent SDK, AWS Bedrock, Azure Identity |
| Protocols | Model Context Protocol (MCP, 6 transports), LSP, OAuth/XAA |
| Observability | OpenTelemetry (traces, metrics, logs) with OTLP exporters |
| Code tools | ripgrep, Sharp (image processing), Marked, Turndown, Diff |
claude-code-source/
├── README.md # This document (English)
├── README.zh.md # 中文版文档
├── assets/ # Images
│ ├── anthropic-claude-code.webp
│ └── screenshot.png
├── package.json # Dependencies (reverse-engineered)
├── tsconfig.json # TypeScript configuration
├── bunfig.toml # Bun preload configuration
├── preload.ts # Runtime shim (feature flags, macros)
├── start.sh # Launcher script
├── scripts/setup.mjs # One-command setup
│
├── claude-code-deep-analysis/ # 18 original analysis articles (EN + 中文)
│ ├── README.en.md # Series index (English)
│ ├── README.md # Series index (中文)
│ ├── 00-core-conclusion.en.md # Each article has .en.md + .md
│ ├── 00-core-conclusion.md
│ ├── ...
│ └── 17-hook-system.en.md
│
└── src/ # Claude Code source (Anthropic)
├── query.ts # Core agent loop (1,729 lines)
├── QueryEngine.ts # Session management (46.6 KB)
├── Tool.ts # Tool interface & registry (29.5 KB)
├── main.tsx # Main React component (4,683 lines)
│
├── entrypoints/ # Entry points (CLI, MCP, SDK types)
├── commands/ # 87+ slash commands
├── tools/ # 45+ tool implementations
│ ├── BashTool/ # Shell execution (430 KB of safety code)
│ ├── AgentTool/ # Sub-agent execution
│ ├── FileEditTool/ # File editing with diff matching
│ ├── MCPTool/ # Model Context Protocol
│ └── ... # 30+ more tools
│
├── services/ # 38 service modules
│ ├── api/ # Anthropic API integration
│ ├── mcp/ # MCP protocol (25 subdirs)
│ ├── compact/ # Context compression
│ ├── SessionMemory/ # Cross-session memory
│ └── ...
│
├── components/ # 146+ React UI components
├── hooks/ # 85+ React hooks
├── utils/ # 564+ utility functions
├── permissions/ # Permission system (5 modes)
├── bridge/ # CLI ↔ VS Code integration
├── constants/ # System prompt assembly & config
├── memdir/ # Cross-session memory system
├── skills/ # Skill/plugin system
├── voice/ # Voice input handling
└── ...
- Source code: All rights belong to Anthropic, PBC. No license is granted.
- Analysis articles (
claude-code-deep-analysis/): Original commentary by the repository maintainers. Code snippets are for educational purposes only.
Contact: For any questions, please reach out via GitHub Issues.