QwenPaw: Open-Source Personal AI Assistant

1. Project Overview

QwenPaw is an open-source, extensible personal AI assistant platform that runs locally or in the cloud, giving global users a private, always-available digital partner that can chat, automate tasks, write code, and connect to the messaging apps they already use.

2. Background & Positioning

QwenPaw comes from the AgentScope team and stands for "Qwen Personal Agent Workstation." Its core mission is to move AI assistants away from being disposable, stateless chat tools and toward being a warm, continuous companion that remembers context, learns over time, and respects user privacy through local-first deployment.

What sets QwenPaw apart from typical chatbot wrappers:

Persistent, structured memory instead of a context window that resets every session — conversations are turned into readable, editable, and searchable Markdown memory.
Deployment flexibility — it can run entirely offline with small local models (no API key required) or connect to 14+ cloud LLM providers, giving users full control over cost, privacy, and capability trade-offs.
Security-first architecture — sandboxing, tool inspection, and skill scanning are built in rather than bolted on, which matters for an assistant that can execute shell commands and touch the filesystem.
Omnichannel presence — a single agent instance can be reached from DingTalk, Lark, WeChat, Discord, Telegram, iMessage, and QQ, rather than being locked to one chat surface.

3. Feature Categories

🧠 Memory & Context

Three-layer memory architecture built on the ReMe personal knowledge framework.

Active working context for the current conversation
Complete, searchable conversation history
A self-evolving personal knowledge base stored as linked Markdown
Scroll-based context management with persistent turn storage
Purpose: give the assistant continuity across sessions instead of amnesia after every restart.

🔒 Security & Sandboxing

A layered defense system around every action the agent can take.

Kernel-level sandboxing (Seatbelt on macOS, Bubblewrap on Linux, AppContainer on Windows)
Tool Guard with pre-execution command inspection (ShellEvasionGuardian)
File Guard to block access to sensitive directories
Skill Scanner that screens third-party skills for threats before activation
Purpose: let the agent safely run commands, edit files, and install skills without becoming an attack surface.

💬 Multi-Channel Connectivity

One agent, many front doors.

DingTalk, Lark (Feishu), WeChat, Discord, Telegram, iMessage, QQ
Purpose: let users talk to the same assistant from whichever app they already live in, without duplicating setup.

🤖 Model & Deployment Options

Run anywhere, with any brain.

Local QwenPaw-Flash models (2B / 4B / 9B) with no API key needed
Ollama and LM Studio integration for local inference
14+ cloud LLM providers (OpenAI, Anthropic, DeepSeek, Kimi, and more)
Docker, Alibaba Cloud ECS, AgentScope Platform, ModelScope Studio, and a beta desktop app
Purpose: match the deployment model to the user's privacy, budget, and hardware constraints.

🛠️ Productivity & Development Tools

Built-in capabilities beyond chat.

Coding Mode: a three-panel web IDE with file navigation, diff preview, and structural code search
Document processing for PDF, Word, Excel, and PowerPoint files
Web search and video summarization
Scheduled task automation and workflow composition
Purpose: turn the assistant into a working tool for coding, research, and day-to-day office tasks, not just conversation.

4. Key Highlights

Local-first, API-key-free operation — QwenPaw-Flash models let users get started without signing up for any cloud provider.
Structured, evolving memory — the ReMe-powered memory system makes past conversations genuinely reusable rather than lost after the session ends.
Defense-in-depth security — kernel sandboxing plus Tool Guard, File Guard, and Skill Scanner reduce the risk of an agent with shell access going wrong.
True omnichannel reach — seven messaging platforms are supported out of the box from a single running instance.
Multi-agent spawning — the assistant can spin up sub-agents with independent memory for parallel or specialized tasks.
Multiple deployment paths — pip, install script, Docker, cloud platforms, and a desktop app cover almost every user comfort level, from CLI-first developers to non-technical users.

5. Use Cases by Role

General developers — Use Coding Mode as a lightweight, chat-driven IDE companion for code search, diffs, and quick automation scripts.

DevOps/SRE — Deploy via Docker or Alibaba Cloud ECS, and use scheduled task automation to handle recurring operational chores; sandboxing limits the blast radius of agent-executed commands.

Security engineers — Review and extend the Tool Guard, File Guard, and Skill Scanner modules, or evaluate QwenPaw's sandboxing model (Seatbelt/Bubblewrap/AppContainer) for agent-safety research.

Data/research scientists — Use document processing and web search skills to summarize papers, spreadsheets, and reports, with persistent memory keeping research context across sessions.

Project managers — Reach the assistant through existing team chat tools (DingTalk, Lark, WeChat, Discord, Telegram) to delegate reminders, document summaries, and scheduled follow-ups without adopting a new app.

6. Getting Started

Find what you need — browse the documentation site for guides on models, channels, skills, security, and the API/CLI:

https://qwenpaw.agentscope.io

Install and run — the quickest path is via pip (Python 3.11–3.13):

pip install qwenpaw
qwenpaw init
qwenpaw start

Docker, an automated install script, and a beta desktop app are also available for users who prefer not to manage Python environments directly.

Contribute — set up the dev environment and run the required checks before opening a pull request:

pip install -e ".[dev,test,full]"
pre-commit install
pre-commit run --all-files
pytest

Follow Conventional Commits (e.g. feat(models): add custom provider for Azure OpenAI) and check GitHub Discussions for areas currently seeking contributors, such as new channels, models, and skills.

7. Project Structure

QwenPaw/
├── src/qwenpaw/     # Core Python package (agent runtime, memory, security)
├── console/         # Web frontend (Node.js) for configuration and chat UI
├── plugins/         # Plugin system for extending channels, skills, and providers
├── deploy/          # Docker and deployment configurations
├── tests/           # Unit test suite
├── e2e/             # End-to-end tests
└── website/         # Documentation site source

Key entry points: src/qwenpaw/ holds the agent logic, model integrations, and security guards; console/ is the local web UI reachable at http://127.0.0.1:8088/ after startup; agents/skills/ (referenced in contributing docs) is where new skills are added, each with a SKILL.md describing when the agent should trigger it.

8. Related Ecosystem

Upstream dependencies: Qwen model family (for QwenPaw-Flash local models), the Model Context Protocol (MCP) for tool integration, and the ReMe framework for personal memory.
Deployment platforms: Alibaba Cloud ECS, AgentScope Platform, ModelScope Studio, and Docker Hub / Alibaba Cloud Container Registry for images.
Complementary tools: Ollama and LM Studio for local model serving; OpenAI-, Anthropic-, DeepSeek-, and Kimi-compatible APIs for cloud inference.

9. License

QwenPaw is released under the Apache License 2.0.

✅ Use, modify, and distribute the software for personal or commercial purposes
✅ Create and distribute derivative works, including proprietary products
✅ Use it in production without paying license fees
❌ Use the QwenPaw name or trademarks to imply official endorsement without permission
❌ Hold contributors liable for damages arising from use of the software
ℹ️ Retain the original copyright notice and license text in redistributed copies
ℹ️ Any contributed changes are accepted under the same Apache 2.0 terms

10. FAQ

Q: Do I need an API key to try QwenPaw?
A: No. QwenPaw-Flash local models (2B/4B/9B) run without any cloud API key; cloud providers are optional for higher-capability responses.

Q: Can I connect QwenPaw to the messaging app my team already uses?
A: Yes. Built-in channel support covers DingTalk, Lark, WeChat, Discord, Telegram, iMessage, and QQ from a single running instance.

Q: Is it safe to let the agent run shell commands and edit files?
A: QwenPaw applies kernel-level sandboxing plus Tool Guard, File Guard, and Skill Scanner to constrain what an agent action can access before it executes.

Q: What's the fastest way to install it?
A: pip install qwenpaw followed by qwenpaw init and qwenpaw start (Python 3.11–3.13); Docker and a guided install script are alternatives for those who prefer not to manage Python directly.

Q: How can I add a new skill or messaging channel?
A: Skills live under agents/skills/ with a SKILL.md manifest; new channels subclass BaseChannel and register in registry.py. See the contributing guide for details.

11. Quick Links

Repository: https://github.com/agentscope-ai/QwenPaw
Official docs / homepage: https://qwenpaw.agentscope.io
Contributing guide: https://github.com/agentscope-ai/QwenPaw/blob/main/CONTRIBUTING.md
Community: GitHub Discussions on the repository, plus Discord and X (Twitter) links listed in the README

12. Summary

QwenPaw packages persistent memory, multi-channel messaging, and built-in security into a single open-source personal AI assistant that can be run entirely offline or scaled to the cloud. It is best suited for developers who want a hackable, self-hostable agent platform, and for teams who want to bring one assistant into the chat tools they already use — all under a permissive Apache 2.0 license.