QwenPaw, formerly CoPaw, is a free, open-source personal AI agent from the AgentScope team. Run it on your computer, in Docker, or on a server, then connect a model and give it work to do. It can research a topic, read documents, prepare a daily digest, help with code, or send a result to a chat channel.
The agent can keep context between sessions, use Skills and tools you approve, and return to scheduled jobs. You can connect a cloud model with an API key or run a local model if you want to keep inference on your own hardware.
Features
- Choose between QwenPaw Local, Ollama, LM Studio, and supported cloud providers. Local models do not need a cloud API key.
- Keep conversation history available for later recall and use long-term memory when the agent needs it.
- Use one installation through the Console, terminal UI, desktop app, or connected services such as Discord, DingTalk, Lark, Telegram, iMessage, WeChat, and QQ.
- Install Skills, plugins, and MCP tools for documents, web research, browser work, news summaries, and other jobs.
- Use Coding Mode for a project folder, or schedule a digest, reminder, report, or channel message.
- Set rules for tool use, file access, and Skills before the agent works with sensitive material.
Use Cases
- Publishers and researchers can gather sources, summarize notes, and receive a daily topic digest.
- People who handle documents can use Skills to read and prepare PDF, Word, Excel, and PowerPoint files.
- Developers can open a project in Coding Mode to inspect code, make an approved change, or review a pull request.
- Anyone with repetitive personal work can schedule a reminder, brief, or channel update instead of starting the same task by hand.
How to use QwenPaw
Start with the local Console. Get one model working and try a small task there before you add a chat channel, scheduled job, or extra Skill. That order keeps the first setup simple and makes problems easier to trace.
Table Of Contents
Install QwenPaw
The official installer is the easiest option on a personal computer. It installs QwenPaw and its required environment.
macOS or Linux:
curl -fsSL https://qwenpaw.agentscope.io/install.sh | bashWindows PowerShell:
irm https://qwenpaw.agentscope.io/install.ps1 | iexOpen a new terminal after installation, then run:
qwenpaw init --defaults
qwenpaw appOpen http://127.0.0.1:8088/. The Console is the main place to chat with QwenPaw, select a model, check its workspace, and manage Skills or channels. Use qwenpaw init if you want the setup prompts instead of the default configuration.
The beta desktop app is available for Windows and macOS. Use it if you do not want to install Python or run commands. If you already manage Python, QwenPaw supports Python 3.11 through 3.13 and installs with pip install qwenpaw.
Connect a model
Open Settings > Models. QwenPaw needs an active model before it can reply or use tools. A cloud model is usually the quicker choice if you already have an API key. Choose the provider, enter the key, select a model, and set it as active. DashScope users can also set DASHSCOPE_API_KEY in their shell or a private .env file. Do not commit that file or share it in screenshots.
For a local setup, choose QwenPaw Local, Ollama, or LM Studio. QwenPaw Local uses a built-in llama.cpp backend and can download compatible models from the Console. Ollama and LM Studio need to be installed and running first. Local models avoid a cloud API charge, but speed and quality depend on your computer and the model you choose.
qwenpaw models download Qwen/Qwen3-4B-GGUF
qwenpaw models local
qwenpaw models listQwenPaw command reference
Run these commands after QwenPaw is installed.
| Command | Use it for | Useful option |
|---|---|---|
qwenpaw init | Run the guided first-time setup. | --defaults uses default answers. --force replaces the existing configuration. |
qwenpaw app | Start the local Console. | Open http://127.0.0.1:8088/ after it starts. |
qwenpaw | Open the terminal chat interface. | qwenpaw tui --resume <id> resumes a previous session. |
qwenpaw . | Open Coding Mode for the current folder. | Run it from the project root. |
qwenpaw models config | Set a provider, API key, and active model. | models list shows the current model setup. |
qwenpaw models download <repo_id> | Download a local GGUF model. | --source modelscope changes the download source. |
qwenpaw channels config | Connect or change chat channels. | channels list shows channel status. |
qwenpaw skills install <URL> | Install a Skill from a supported URL. | skills list shows enabled Skills. |
qwenpaw uninstall | Remove the app and keep its data. | --purge also deletes configuration and data. |
Try one small task
Use the Console for the first task. Ask for something useful that does not need access to every account or folder. For example, you can ask QwenPaw to create a research brief from a short list of sources, summarize a folder of notes without editing it, or explain a codebase in Coding Mode without changing files.
If QwenPaw asks to run a tool or read a file, check the request before you approve it. A narrow first task gives you a clear picture of what the agent can do with the model and permissions you chose.
Add Skills and recurring tasks
Skills add work that a basic chat does not handle on its own, such as document processing, web research, browser tasks, or news summaries. Install one Skill, test it with a real task, and check its access before adding another.
Use qwenpaw skills list to see enabled Skills. To install one from a supported source, run qwenpaw skills install <URL>. When a task produces reliable results, you can create a scheduled task or heartbeat to run it again later.
Connect a chat channel
You do not need a chat channel to use QwenPaw. Add one when you want a result sent to Discord, DingTalk, Lark, Telegram, or another supported service. Each service needs its own app or bot credentials. Open Channels in the Console or run qwenpaw channels config, then send one test message before you use that channel for scheduled work.
Use the terminal or Docker
Run qwenpaw for the full-screen terminal chat. Run qwenpaw . inside a project folder to open Coding Mode for that project.
If you prefer Docker, keep QwenPaw’s workspace, credentials, and backups in separate volumes:
docker pull agentscope/qwenpaw:latest
docker run -p 127.0.0.1:8088:8088 \
-v qwenpaw-data:/app/working \
-v qwenpaw-secrets:/app/working.secret \
-v qwenpaw-backups:/app/working.backups \
agentscope/qwenpaw:latestWhen the container starts, open http://127.0.0.1:8088/. Add -e VAR=value or --env-file .env only if the container needs a provider key.
Pros
- Free and open source under Apache 2.0.
- Runs with local or cloud models.
- Works through the Console, TUI, desktop app, and chat channels.
Cons
- Needs an active model.
- Local models need capable hardware.
Alternatives & related resources
- 7 Best OpenClaw Alternatives for Safe & Local AI Agents
- NanoClaw: Lightweight OpenClaw Alternative for Safer Local AI Agents
- Nanobot: Lightweight Open-source AI Assistant Inspired by OpenClaw
- Openclaw: Your 24/7 Personal AI Employee
- Automate Anything: 10 Best & Open-source AI Agents
FAQs
Q: Is QwenPaw free and open source?
A: Yes. QwenPaw is licensed under Apache 2.0. You can run it with a local model without paying for a cloud model API, although your computer must have enough resources for that model.
Q: Can I run QwenPaw on Windows?
A: Yes. Use the official PowerShell installer, install with pip, run the Docker image, or use the Windows desktop app. QwenPaw Local, Ollama, and LM Studio are available local-model paths on Windows.
Q: Does QwenPaw require a paid API subscription?
A: No. A cloud provider needs its own API key, but QwenPaw can also run through QwenPaw Local, Ollama, or LM Studio.
Q: Can I connect QwenPaw to Discord?
A: Yes. Start the Console, open Channels, and configure Discord with the required bot credentials. The same QwenPaw instance can also be connected to other supported channels.
Q: How do QwenPaw Skills work?
A: Skills add capabilities such as document processing, web work, browser tasks, or scheduled jobs. Install them from a supported URL or marketplace, review their requested access, and enable them for the agent that should use them.
