[D] We reimplemented Claude Code entirely in Python — open source, works with local models

Hey everyone,

We just released Claw Code Agent — a full Python reimplementation of the Claude Code agent architecture, based on the reverse-engineering work shared in this tweet:

https://x.com/Fried_rice/status/2038894956459290963

Why?

The original Claude Code is npm/TypeScript/Rust. If you're a Python developer, good luck reading or extending it. We rebuilt the whole thing in pure Python so anyone can understand it, modify it, and

run it with local open-source models.

What it does:

• Full agentic coding loop with tool calling
• Core tools: file read/write/edit, glob, grep, shell
• Slash commands: /help, /context, /tools, /memory, /status, /model
• Context engine with CLAUDE.md discovery
• Session persistence — save and resume agent runs
• Tiered permissions: read-only → write → shell → unsafe

Works with any OpenAI-compatible backend:

• vLLM (documented path)
• Ollama
• LiteLLM Proxy

Recommended model: Qwen3-Coder-30B-A3B-Instruct — runs fully local, fully free.

Repo: https://github.com/HarnessLab/claw-code-agent

We're actively working on this and happy to add features or take PRs. If something is missing or broken, open an issue — we want to make this useful for the community.

Would love to hear your feedback.