QwenPaw

Quick Verdict

QwenPaw is for people who want an AI assistant installed on their own machine and kept long-term. The name stands for Qwen Personal Agent Workstation. It is maintained by the open-source AgentScope team under Apache-2.0, and pip install qwenpaw gets you a personal agent with a web console that you can then connect to DingTalk, Lark, WeChat, Telegram, and Discord — one instance, one memory, the same assistant in whichever window you reach for. That is the key difference from platforms like Coze and Dify, which exist to help you ship bots for other people: QwenPaw’s default user is you.

The deciding question is whether you actually need a persistent personal assistant. For occasional questions, a browser chat tool is less work. If you want something that remembers preferences from six months ago, pushes you a digest on schedule, reads and writes your local documents, and stays reachable across several messaging apps, this is currently among the most complete open-source options. The trade-off is that deployment, model access, and security boundaries are yours to own.

Best For

• Privacy-minded users: Want data on their own machine with no third-party hosting
• Teams in Chinese workplaces: DingTalk, Lark, and WeChat are first-class channels, with DashScope/Qwen direct access and a documented Alibaba Cloud ECS path
• People who care about memory: Need an assistant that retains preferences, project context, and past conclusions
• Agent builders and tinkerers: Want to write Skills, plugins, or sub-agent orchestration on the AgentScope stack
• Not ideal for: users who refuse to touch a command line; teams needing multi-tenancy, SSO, and audit compliance; or light users who only ask occasional questions and would not recoup the maintenance cost

Key Features

• Three-layer memory: Live working context, full verbatim history, and a self-evolving personal knowledge base powered by ReMe. Conversations and resources continuously become readable, editable, searchable, linked Markdown rather than an opaque vector store
• Scroll Context keeps history: Every turn is persisted, and turns evicted from the window are indexed for on-demand recall. The project states explicitly that nothing is summarized away, which suits a long-lived assistant better than compression
• Local models, no API key: A built-in llama.cpp-based QwenPaw Local runtime plus purpose-trained QwenPaw-Flash 2B / 4B / 9B agent models (Q4 and Q8 quantizations on both ModelScope and Hugging Face), alongside Ollama, LM Studio, and 14+ cloud providers
• Kernel-level security stack: Sandbox, Tool Guard, File Guard, Skill Scanner, and Access Policy block dangerous commands before they run. Each agent’s workspace splits into Resources (transparent on disk), Governance (allow/deny/ask/sandbox), and Sandbox
• Multi-agent and protocol layer: Spawn sub-agents with their own memory and skills at runtime; the Drivers layer is a protocol-neutral MCP / A2A / ACP connector with encrypted credentials and a per-call policy gate
• Reachable anywhere: DingTalk, Lark, WeChat, Discord, Telegram, iMessage, and QQ from one instance, plus a web console, terminal UI, and desktop app (beta)
• Extensible ecosystem: Skills for scheduling, documents, browser, and news; the Oh-My-Paw plugin market; the PawApp mini-app platform; user-editable Agent Modes; and Loop Engineering templates such as Coding Mode and Mission Mode
• Unified file workspace: Navigation, preview, editing, diffs, upload, and download across project and agent files, with workspace checkpoints, Browser-use, and Computer-use added in v2.1.0

Use Cases

• Scheduled jobs and push: Generate daily digests, weekly reports, or data summaries and broadcast them to several messaging channels
• Personal knowledge Q&A: Turn documents, meeting notes, and subscriptions into ReMe memory and query it directly
• Document processing: Read, write, and convert PDF, Word, Excel, and PowerPoint files
• Code and development: Read, edit, review, and test project code in the unified file workspace, with Codex and Qoder agent integration since v2.1.0
• Information gathering: Search the web, follow feeds, summarize videos, and file results into the knowledge base
• Multi-channel ops: Route alerts, summaries, or generated content per channel or broadcast everywhere at once

Pricing

QwenPaw is Apache-2.0 open source and free, with no paid tiers. Deployment options are generous: pip, one-line script, Docker, Alibaba Cloud ECS, ModelScope, a beta desktop app, and the official AgentScope Platform, which offers free hosted QwenPaw deployment, plugin sharing, and a Skill marketplace if you would rather not run a machine yourself.

Cost only appears on the model side. Cloud APIs (DashScope/Qwen, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, OpenRouter, and more) require a key before chatting works at all. Local models via QwenPaw Local, Ollama, or LM Studio need no key, and combined with the small QwenPaw-Flash models the running cost can be zero. Some tools need extra keys — web search wants TAVILY_API_KEY, set in the console’s environment variables. Release cadence is fast: 2.0.0, a ground-up rewrite on AgentScope 2.0, shipped 2026-07-10, and v2.1.0 landed 2026-08-13, so check that any tutorial you follow still matches the current architecture.

Pros

• Apache-2.0 open source with fully local data and models, no third-party hosting required
• The most serious memory design in this category: nothing summarized away, and a readable, editable Markdown knowledge base
• Security is designed in rather than bolted on — sandbox, tool gate, skill scanning, and access policy form a coherent stack
• Strong China-side path: Qwen and DeepSeek direct access, ModelScope weights, Alibaba Cloud deployment, DingTalk/Lark/WeChat channels
• Enough deployment options that setup friction is hard to justify as an excuse, plus a free hosted platform for trials
• Self-trained QwenPaw-Flash models make running an agent on modest hardware realistic

Cons

• Fast iteration: 2.0 was a full rewrite, so older tutorials and plugins break easily
• A resident service means real operations work — ports, autostart, backups, upgrades are yours
• Powerful capabilities carry matching risk; a misconfigured IM integration can leak data or take unwanted actions
• Built for individuals, not teams — no multi-tenancy, SSO, or audit tooling
• A default-flag install auto-accepts anonymous telemetry, so privacy-sensitive users must intervene
• The desktop app is still beta and less polished than commercial assistants

Alternatives

FAQ

Is QwenPaw free?

The project is Apache-2.0 and free, and the official AgentScope Platform offers free hosted deployment. The only spend is cloud model APIs; running QwenPaw Local, Ollama, or LM Studio keeps cost at zero.

Do I need an API key to use it?

For cloud models, yes — it will not chat without a valid key. Local models need no key at all: the built-in QwenPaw Local runtime and QwenPaw-Flash 2B/4B/9B models can be downloaded straight from the web UI.

Is deployment practical for users in China?

Yes. Models can connect directly to DashScope/Qwen, DeepSeek, and Kimi, weights are mirrored on ModelScope, the docs include an Alibaba Cloud ECS section, and DingTalk, Lark, and WeChat channels are built in, so no overseas service is required.

How is this fundamentally different from Coze or Dify?

Those platforms help you build an application for other people. QwenPaw is about keeping an assistant of your own: single-user by default, local-first, memory that accumulates, and one identity across messaging channels. For multi-tenant product delivery, choose the other two.

Is connecting it to a WeChat or DingTalk group safe?

Only if you configure it that way. It ships a kernel-level sandbox, Tool Guard, File Guard, Skill Scanner, and Access Policy, but an agent with file and command execution exposed in a group chat is inherently risky — tighten policy and restrict tools before connecting.

Will upgrading break my config and memory?

2.0.0 was a full rewrite on AgentScope 2.0, so major-version upgrades should follow the release notes rather than an in-place overwrite. Memory lives on disk as Markdown, so backing up the workspace directory first is the safe move.

Bottom Line

QwenPaw handles the “personal AI assistant” problem more completely than most open-source projects: memory that is not thrown away, security as a system, full channel coverage, local models as a fallback, and a smooth path for users in China. The barrier is not technical difficulty but willingness to maintain a resident service. If you are willing, it becomes an assistant that genuinely remembers you and is always within reach; if not, a desktop client plus a browser chat tool is less trouble. Try it in the order that costs least: run it on AgentScope Platform to learn the concepts, then deploy locally, connect one channel, set one scheduled task, feed it a batch of documents, and check after two weeks whether it actually saved you time.