Claude Code Py

中文说明

Python implementation of Claude Code, Anthropic's official AI coding assistant CLI. Reverse-engineered from the original TypeScript source (~500K lines) and rewritten in Python with full tool-use support.

Uses any OpenAI-compatible API (DashScope/Qwen by default) as the LLM backend.

Quick Start

# Install
git clone https://github.com/ZackZikaiXiao/claude-code-py.git
cd claude-code-py
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

# Set your API key
export DASHSCOPE_API_KEY=sk-xxx

# Run
claudepy -y

-y auto-approves tool calls so the AI can read/write files and run commands without asking every time.

What It Can Do

Like the original Claude Code, you talk to it in natural language and it uses tools to get things done:

> Help me create a Flask app with a /hello endpoint

  📝 Write  app.py (12 lines)
  ✓ Write  Created: app.py

  $ pip install flask
  $ python app.py

Done. Flask app running on http://localhost:5000/hello

20 built-in tools: Bash, Read, Edit, Write, Grep, Glob, Agent (sub-agents), WebSearch, WebFetch, NotebookEdit, Task management, Plan mode, and more.

Switching Models

Works with any OpenAI-compatible API. Built-in provider presets:

claudepy -y                            # Default: DashScope qwen-plus-latest
claudepy -y --provider openai          # OpenAI gpt-4o (needs OPENAI_API_KEY)
claudepy -y --provider deepseek        # DeepSeek (needs DEEPSEEK_API_KEY)
claudepy -y --provider anthropic       # Anthropic Claude (needs ANTHROPIC_API_KEY)
claudepy -y --provider openrouter      # OpenRouter (needs OPENROUTER_API_KEY)
claudepy -y --model qwen-max           # Use a specific model name

Switch at runtime with /provider openai or /model gpt-4o.

CLI Options

claudepy -y                     # Recommended: auto-approve tools
claudepy -y -p "fix the bug"   # One-shot mode (non-interactive)
claudepy --model qwen-max       # Specify model
claudepy --base-url https://... # Custom API endpoint
claudepy --max-tokens 16384     # Max output tokens

Slash Commands

Type these during a conversation:

Command	Description
/help	Show all commands
/model qwen-max	Switch model
/provider openai	Switch provider
/clear	Clear conversation
/compact	Compress context to save tokens
/cost	Show token usage
/tools	List available tools
/exit	Quit

Permission System

First time a write tool runs, you'll see:

  ⚠ Bash
  $ npm install
  Allow? Yes(y) / Always this session(a) / Always forever(!) / No(n):

• y — allow this once
• a — don't ask again this session
• ! — never ask again (saved to ~/.claude/settings.json)
• n — deny

Or just use claudepy -y to skip all of this.

How It Works

This is a Python rewrite based on the Claude Code TypeScript source. The core architecture is preserved:

1. Query Loop — send messages + tool schemas to LLM → parse streaming response → execute tool calls → loop until done
2. Tool System — 20 tools, each with JSON schema, permission rules, and async execution
3. Concurrent Execution — read-only tools (Grep, Glob, Read) run in parallel; write tools run serially
4. Auto-Compact — when conversation approaches context limit, older messages are summarized by the LLM
5. Persistent Permissions — tool approvals saved to disk, no repeated prompts

Install from GitHub

pip install git+https://github.com/ZackZikaiXiao/claude-code-py.git

Requirements

• Python >= 3.9
• openai, rich, prompt_toolkit, httpx
• Optional: markdownify, pypdf

License

MIT