Claude Code Source Leak Reimplementations | Generated by AI

Home 2026.05

Question: What are the best Claude Code open source projects/reimplementations right now?

Answer:

Big story here: on March 31, 2026, Anthropic accidentally shipped source maps inside the Claude Code npm package, exposing the full TypeScript source (~1,884 files). Chaofan Shou discovered that the source map files embedded in the published npm package contained the original TypeScript source under the sourcesContent key in the JSON. This triggered an explosion of reimplementations. Here’s the landscape:

The Leaked Source

yasasbanukaofficial/claude-code The raw leaked TypeScript source — the actual Claude Code CLI codebase including tool-calling, agentic workflows, and terminal UI. Archived for educational purposes. Note: this is just the skeleton, not the model/brain.

chauncygu/collection-claude-code-source-code A collection hub that archives the raw leaked TypeScript source and houses multiple subprojects studying Claude Code from multiple angles — architecture analysis, research reports, and clean-room reimplementations.

Best Reimplementations

1. nano-claude-code (Python) — Best for learning internals

A minimal, fully-runnable Python reimplementation that has grown from a ~900-line v1.0 prototype to a ~5,000-line v3.0 with multi-agent orchestration, persistent memory, and a skill system. Unlike architectural studies, it’s a real coding assistant you can use immediately. Supports 20+ models: Anthropic, OpenAI, Gemini, DeepSeek, Ollama, LM Studio.

Tool coverage: Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch, MemorySave/Delete/Search, Agent, SendMessage, Skill, SkillList — plus built-in /commit, /review, and custom markdown skills with argument substitution and fork/inline execution.

This is the best one to study if you want to understand the agent loop architecture.

2. ruvnet/open-claude-code (TypeScript) — Best clean-room drop-in replacement

A clean-room implementation (no leaked source used) that mirrors the actual Claude Code architecture: async generator agent loop, 25 tools, 4 MCP transports (stdio, SSE, Streamable HTTP, WebSocket), 6 permission modes, hooks, settings chain, sessions. Supports Anthropic, OpenAI, Gemini, AWS Bedrock, Google Vertex. Automated nightly releases with 903+ tests.

npx @ruvnet/open-claude-code "explain this codebase"

3. openclaw/openclaw — Referenced widely for architecture comparison

Mentioned in multiple analyses as the closest peer to the original. Used in cross-system comparisons alongside Claude Code and Hermes-Agent to analyze agent design principles.

The hottest repos right now by stars:

What to Study as an AI Engineer

Repo Why it matters
nano-claude-code Clean Python agent loop — read the source top-to-bottom
Leaked TypeScript source See the actual system prompts, tool schemas, self-healing mechanisms
ruvnet/open-claude-code Production-quality clean-room; good reference for MCP transport impl
VILA-Lab/Dive-into-Claude-Code Systematic analytical framework tracing design values → principles → implementation — good for understanding why architectural choices were made, not just what they are

The biggest insight from the leak: cross-cutting integrative mechanisms — not modular features — are the true locus of engineering complexity in Claude Code. The tool loop, permission system, and context management all bleed into each other. Worth studying that specifically.

References:

Back Donate