Python 3.10+ Python 3.12- License: MIT Alpha

中文版 | English

Pywen Logo

Pywen is a full-stack Python Code Agent platform built for transparent engineering practice and reproducible research. It serves both as a practical coding assistant for developers and as a unified foundation where different agents can be compared, debugged, and evolved in a controllable environment.

Pywen is positioned as a unified foundation for the Code Agent ecosystem, with two complementary roles:

• Fair Arena (Research): Reproducible and comparable evaluation environment for different Code Agents
• Agent OS (Engineering): Stable and evolvable runtime for moving from prototype to production

Core design goals:

• Reproducibility: Reconstruct diverse agent behaviors under one Python framework
• Fair Comparison: Standardized tool interfaces and evaluation workflows
• Iterative Stability: Modular architecture for smooth model/tool evolution
• Production Readiness: Permission control, approval flow, and trajectory audit for governance

• 2026年01月25日: Launched the Skills system for reusable capability injection and modular extension. Post: https://mp.weixin.qq.com/s/8t6JtpT9ARB7hy3fow6xkA
• 2025年12月29日: Released Pywen 2.0 with a core refactor focused on maintainability and extensibility. Post: https://mp.weixin.qq.com/s/U1XBBNLGWObU5MOdTZcECQ
• 2025年11月26日: Refactored core agent naming from QwenAgent to PywenAgent for better clarity and consistency
• 2025年08月26日: Updated the /agent module by adding the Claude Code agent, with execution logic aligned to Claude Code, and introducing dedicated tools such as the task tool and todowrite tool. You can switch to the Claude Code agent using /agent claude.
• 2025年08月08日: Updated the /agent switch agent module, added DeepResearch agent, with execution logic aligned with Google's open-source DeepResearch LangGraph version. You can use /agent research to switch GeminiResearchDemo agent. Before you use it, please make sure you have the serper api key.
• 2025年08月06日: Released the first version of Pywen, with execution logic aligned with Gemini CLI/Qwen Code

Pywen was developed in response to common pain points in existing Code Agent tooling for Python users: fragmented tech stacks, limited observability, difficult customization, and poor cross-agent comparability.

From implementation perspective, Pywen references and learns from projects such as Qwen-Code and Codex, and can run with code-capable models including Qwen3-Coder. We acknowledge and appreciate these open-source contributions. These are development choices rather than the primary product positioning.

• 🚀 Code Specialization: Qwen3-Coder excels in code generation, understanding, and debugging
• ⚡ Efficient Inference: Optimized model architecture providing fast responses
• 🔧 Engineering Practicality: Specifically trained for real-world software development scenarios

Project Status: The project is still under active development. We welcome your help in improving Pywen.

Pywen is a Python-native CLI with strong ecosystem compatibility and a transparent modular architecture. It is designed for both practical delivery and scientific iteration: developers can inspect and control execution paths, while researchers can run fair comparisons and reproducible studies in a unified environment. This research-friendly and engineering-oriented design helps build an open, evolvable foundation for the Code Agent community.

• 🤖 Multi-Agent Support: Pywen Agent (Qwen3-Coder based), Claude Code Agent, Codex Agent, and Research Agent
• 🚀 Multi-Model Support: Supports Qwen as well as OpenAI- and Anthropic-compatible APIs, enabling seamless switching between different models
• 📦 Modular: Built on modular architecture, extensible and customizable
• 🛠️ Rich Tool Ecosystem: File editing, bash execution, web search, memory management, and more
• 📊 Trajectory Recording: Detailed logging of all Agent operations for debugging and analysis
• ⚙️ Smart Configuration: YAML-based configuration with environment variable support
• 📈 Session Statistics: Real-time tracking of API calls, tool usage, and token consumption
• 🔄 Agent Switching: Seamlessly switch between different agents using /agent command

pip install pywen

git clone https://github.com/PAMPAS-Lab/Pywen.git
cd Pywen
uv venv
uv sync --all-extras
# linux/macos
source .venv/bin/activate
# windows
.venv\Scripts\activate

Simply run the pywen command to start:

# Interactive mode (default)
pywen
# Single prompt mode
pywen "Create a Python hello world script"
# Specify agent type
pywen --agent pywen
pywen --agent claude
pywen --agent codex
# Specify model and API key via command line
pywen --model "Qwen/Qwen3-Coder-Plus" --api_key "your-key"

If it's your first run and there's no configuration file:

• Pywen will first try to locate a config file:
  • Default path: ~/.pywen/pywen_config.yaml
  • A pywen_config.yaml in the current working directory or any parent directory
• If none is found, it will try to locate an example config pywen_config.example.yaml and copy it to the default path:

# built-in example configuration located at `pywen/config/pywen_config.example.yaml`
cp pywen/config/pywen_config.example.yaml ~/.pywen/pywen_config.yaml

• If there is neither a real config file nor an example file, Pywen will exit with an error message and suggest two options:
  • Manually copy an example file to pywen_config.yaml and edit fields like api_key / model
  • Or provide configuration via environment variables (e.g. PYWEN_API_KEY, PYWEN_MODEL, PYWEN_BASE_URL)

Therefore, the recommended workflow is to prepare pywen_config.yaml (or the relevant environment variables) according to the example configuration before running pywen.

Once you enter the Pywen command-line interface, you can:

# File operations
> Create a Python script to calculate Fibonacci sequence
> Refactor functions in main.py to make them more efficient
# Code analysis and debugging
> Fix bugs in this project and add unit tests
> Analyze performance bottlenecks in my code
# Project management
> Create a new Flask web application with proper structure
> Add comprehensive documentation to this codebase

# System commands
/about show version info
/agent switch between different agents (pywen/claude/codex/research)
/clear clear the screen and conversation history
/help for help on pywen code
/model view and manage model configurations
/stats check session stats
/tools list available Pywen tools 
/bug submit a bug report
/quit exit the cli
# Special commands
!<command> - Execute shell command
# Keyboard shortcuts
Ctrl+Y - Toggle YOLO mode (auto-approve all operations - use with caution!)
# Direct input of task descriptions to execute agent

Pywen supports multiple specialized agents:

# List available agents
/agent
# Switch to Pywen Agent (default, Qwen3-Coder based)
/agent pywen
# Switch to Claude Code Agent
/agent claude
# Switch to Codex Agent (OpenAI GPT-5 Codex)
/agent codex
# Switch to Research Agent (Gemini-based)
/agent research

Available Agents:

• Pywen Agent (pywen): General-purpose coding assistant based on Qwen3-Coder
• Claude Code Agent (claude): Advanced file operations and project understanding
• Codex Agent (codex): OpenAI Codex-based coding assistant
• Research Agent (research): Multi-step research agent for comprehensive information gathering

Use with caution:

• Press Ctrl+Y to toggle YOLO mode
• Default behavior: All tool calls require user confirmation for safety
• In YOLO mode, all tool calls are automatically approved without user confirmation
• This speeds up execution but removes safety checks
• Mode status is displayed in the interface
• You can also type 'a' (always) when prompted for tool confirmation to enable YOLO mode

Pywen uses YAML configuration files. The default configuration file is located at ~/.pywen/pywen_config.yaml.

Example Configuration:

# Default agent to use
default_agent: pywen
# Model configurations
agents:
 # Pywen Agent (Qwen3-Coder)
 - agent_name: pywen
 model: "Qwen/Qwen3-Coder-Plus"
 api_key: "your-api-key"
 base_url: "https://api-inference.modelscope.cn/v1"
 provider: openai
 wire_api: chat
 # Claude Code Agent
 - agent_name: claude
 provider: anthropic
 model: "claude-3.5-sonnet"
 api_key: "your-anthropic-key"
 base_url: "https://api.anthropic.com/v1"
 wire_api: chat
 # Codex Agent
 - agent_name: codex
 provider: openai
 model: "gpt-5.1"
 api_key: "your-openai-key"
 base_url: "https://api.openai.com/v1/"
 wire_api: responses
# Permission level: locked / edit_only / planning / yolo
permission_level: locked
# Maximum conversation turns
max_turns: 10
# Memory monitor settings
memory_monitor:
 check_interval: 3
 maximum_capacity: 100000
 model: "Qwen/Qwen3-235B-A22B-Instruct-2507"

Configuration Priority:

1. Command line arguments (highest)
2. Configuration file values
3. Environment variables
4. Default values (lowest)

Configuration File Location: Pywen ships an embedded example configuration; on first run it will copy it to the default config path.

• Default: ~/.pywen/pywen_config.yaml
• You can specify a custom path with --config flag

You can set API keys through environment variables. Pywen supports agent-specific environment variables:

# Pywen Agent (Qwen3-Coder)
export PYWEN_PYWEN_API_KEY="your-api-key"
export PYWEN_PYWEN_BASE_URL="https://api-inference.modelscope.cn/v1"
export PYWEN_PYWEN_MODEL="Qwen/Qwen3-Coder-Plus"
# Claude Agent
export PYWEN_CLAUDE_API_KEY="your-anthropic-key"
export PYWEN_CLAUDE_BASE_URL="https://api.anthropic.com/v1"
export PYWEN_CLAUDE_MODEL="claude-3.5-sonnet"
# Codex Agent
export PYWEN_CODEX_API_KEY="your-openai-key"
export PYWEN_CODEX_BASE_URL="https://api.openai.com/v1/"
export PYWEN_CODEX_MODEL="gpt-5.1"
# Generic fallback (if agent-specific not set)
export PYWEN_API_KEY="your-api-key"
export PYWEN_BASE_URL="https://api-inference.modelscope.cn/v1"
# Tool API Keys (optional but recommended)
export SERPER_API_KEY="your-serper-api-key" # For web search
export JINA_API_KEY="your-jina-api-key" # For content reading

Environment Variable Format:

• Agent-specific: PYWEN_<AGENT_NAME>_<FIELD> (e.g., PYWEN_PYWEN_API_KEY)
• Generic: PYWEN_<FIELD> (fallback if agent-specific not set)

Serper API (Web Search):

1. Visit serper.dev
2. Sign up for a free account
3. Get your API key from the dashboard

Jina API (Content Reading):

1. Visit jina.ai
2. Sign up for a free account
3. Get your API key from the dashboard

Pywen provides a comprehensive toolkit for software development. Different agents may have access to different tools:

Common Tools (Available to most agents):

• File Operations: read_file, write_file, edit, read_many_files
• File System: ls, glob, grep
• Bash Execution: bash - Run shell commands and scripts
• Web Operations: web_search, web_fetch
• Memory Management: memory - Store and retrieve information

Agent-Specific Tools:

• Claude Agent: task, todo - Task planning and management
• Codex Agent: update_plan, apply_patch - Codex-specific operations
• Research Agent: Specialized research workflow tools

For detailed information about all available tools and their capabilities, see docs/tools.md.

Pywen also supports MCP (Model Context Protocol) to connect external tools and services such as playwright.

1. Open your configuration file:

   ~/.pywen/pywen_config.yaml

2. Locate the mcp section and enable it:

   mcp:
    enabled: true
    isolated: false
    servers:
    - name: "playwright"
    command: "npx"
    args:
    - "@playwright/mcp@latest"
    enabled: true
    include:
    - "browser_*"
    save_images_dir: "./outputs/playwright"

Ensure that you have Node.js installed. You can verify it with the following command:

node -v

If you don’t have it installed, please follow the Node.js installation guide.

If your device does not have a browser installed, install browser for Playwright with:

npx playwright install --with-deps

After enabling MCP and installing the required browser, Pywen will be able to call the playwright MCP server for tasks like browser automation, screenshot capture, and web interaction.

Pywen automatically records detailed execution trajectories for debugging and analysis:

# Trajectory files are automatically saved to trajectories/ directory
trajectories/trajectory_xxxxxx.json

Trajectory files contain:

• LLM Interactions: All messages, responses, and tool calls
• Agent Steps: State transitions and decision points
• Tool Usage: Which tools were called and their results
• Metadata: Timestamps, token usage, and execution metrics

Monitor your usage with real-time statistics:

> /stats

Tracks:

• API calls and token consumption
• Tool usage patterns
• Session duration
• Model performance metrics

We welcome contributions to Pywen! Here's how to get started:

1. Fork the repository
2. Set up the development environment:

   git clone https://github.com/your-username/Pywen.git
   cd Pywen
   uv venv
   uv sync --all-extras
   # linux/macos
   source .venv/bin/activate
   # windows
   .venv\Scripts\activate

3. Create a feature branch
4. Make your changes and add tests
5. Submit a pull request

• Follow PEP 8 style guidelines
• Add tests for new features
• Update documentation as needed
• Use type hints appropriately
• Ensure all tests pass before submitting

• Python 3.10+,<3.13
• API key for your chosen agent:
  • Pywen Agent: ModelScope API key or Qwen API key
  • Claude Agent: Anthropic API key
  • Codex Agent: OpenAI API key
  • Research Agent: Google API key (and Serper API key for web search)
• Internet connection for API access
• (Optional) Node.js for MCP server support

Configuration Issues:

# re-create config from the built-in example template
rm ~/.pywen/pywen_config.yaml
pywen

API Key Issues:

# Verify your API key is set (for Pywen agent)
echo $PYWEN_PYWEN_API_KEY
# Or check generic fallback
echo $PYWEN_API_KEY
# Check configuration in Pywen
> /model

Agent Switching Issues:

# List available agents
> /agent
# Check current agent type
> /stats

We thank:

• Google for the Gemini CLI project, which provided agent execution logic and rich tool ecosystem libraries for this project
• Alibaba Cloud Qwen Team for providing the powerful Qwen3-Coder model and Qwen-Code reference implementation
• OpenAI for the Codex project, which provided valuable ideas and references for the Codex-style agent design in this project
• ByteDance for the trae-agent project, which provided valuable foundational architecture for this project
• Yuyz0112 for the claude-code-reverse project, and shareAI-lab for the Kode project, both of which provided inspiration for the Claude code agent development in this project

This project is licensed under the MIT License - see the LICENSE file for details.

Pywen - Making the power of Code Agent accessible for intelligent software development! 🚀

PAMPAS-Lab - Dedicated to breakthroughs in large model agent frameworks, bridging AI research and applications! 🚀