What Is QwenPaw

QwenPaw (formerly CoPaw) is an open-source personal AI agent workbench from Alibaba's AgentScope team. It uses a local-first architecture: conversations, memory and files are stored on your own device by default, and the data belongs to you. QwenPaw ships with the ReMe self-evolving memory system and supports full-channel access across DingTalk, Feishu, WeChat and more, so a single instance can host multiple agents with independent memories. It includes a five-layer security sandbox, works with the local QwenPaw-Flash models plus 14+ cloud providers, and extends through Skills and MCP to automate coding, scheduled tasks and document processing.

Main Capabilities

• Multi-model support: compatible with the self-developed QwenPaw-Flash local models and 14+ cloud LLMs, and able to run fully offline.
• ReMe memory: a three-layer self-evolving memory system (live context + full history + Markdown knowledge base) that extracts notes and produces periodic summaries, with manually editable memory files.
• Full-channel access: one instance connects to 7 major platforms such as DingTalk, Feishu, WeChat, Discord and Telegram, with memory shared across them.
• Skill extensions: built-in scheduled tasks, document processing and browser operations, plus external tools through the MCP protocol.
• Multi-agent collaboration: one instance hosts multiple independent agents, orchestrating complex tasks across systems via the ACP protocol.
• Security sandbox: a five-layer defense system that intercepts dangerous commands, with allow/deny/ask approval levels.
• Flexible deployment: local pip installation, a free cloud sandbox, and one-click Alibaba Cloud Compute Nest deployment.

How to Use QwenPaw

1. Visit the site: open https://qwenpaw.agentscope.io/ to review the project.
2. Get the source: click Download, or clone the GitHub repository locally.
3. Install the program: run pip install qwenpaw, or install from source.
4. Initialize the workspace: run qwenpaw init --defaults to generate default configuration and directories.
5. Configure a model: set a local model path or enter a cloud API key in ~/.qwenpaw/.
6. Start the service: run qwenpaw app for the web console, or qwenpaw for the TUI terminal.
7. Connect channels: add bot credentials for DingTalk, Feishu, WeChat and other platforms in the configuration.
8. Create an agent: add a new agent through the interface or a config file, and assign it an independent memory space and skill permissions.
9. Invoke skills: trigger built-in skills with natural-language instructions, or install MCP plugins to extend external tooling.
10. Manage memory: view and edit Markdown memory files directly in the memory/ directory; indexing and evolution happen automatically.

Use Cases

• Personal productivity automation: schedule a Cron task to pull news RSS feeds, generate a daily briefing with ReMe memory, and push it to DingTalk, Feishu or WeChat.
• Team collaboration hub: connect WeCom, DingTalk and Discord to one instance, answer customer questions in a unified way, and keep context synchronized across platforms.
• Development assistance: in Coding Mode, read pull requests, compare code and draft review comments, with a three-panel web IDE for live diff previews.
• Content creation and operations: draft an article, adapt the format automatically, and publish it to WeChat Official Accounts, Discord, Telegram and more.
• Security-sensitive scenarios: run the QwenPaw-Flash local model fully offline for sensitive documents and personal data, keeping data on-device for compliance.

Limitations and Notes

• Because of the local-first architecture, full capability depends on your local runtime and a configured model or API key.
• Offline use requires a local model (QwenPaw-Flash 2B/4B/9B), and results depend on hardware performance.
• Channel integrations require each platform's own bot credentials and permissions.
• Specific features, model lists and deployment options are subject to the latest documentation on the official site and repository.

FAQ

Q1: How is QwenPaw different from a normal chatbot such as ChatGPT?

A: ChatGPT is a cloud chat tool, whereas QwenPaw is a personal agent workbench that runs on your own machine, with self-evolving memory, a five-layer security sandbox, full-channel access and multi-agent collaboration.

Q2: Will my data be uploaded to the cloud?

A: No. QwenPaw is local-first: all conversations, memory and files are stored on your local device by default, it can run fully offline, and the data belongs to you.

Q3: Can I use it without programming experience? Is installation complex?

A: Yes. Running pip install qwenpaw and qwenpaw init --defaults is enough to start, and a web console and desktop app let you configure everything without code.

Q4: Which platforms does QwenPaw support? Can it connect to WeChat or DingTalk?

A: Yes. A single instance can connect to DingTalk, Feishu, WeChat, QQ, Discord, Telegram, iMessage and other major platforms, all sharing one memory.

Q5: Can I use it without internet or without paying for an API?

A: Yes. The built-in QwenPaw-Flash (2B/4B/9B) local models run fully offline with no API key, and they run smoothly on an ordinary laptop.