agentscope-ai/QwenPaw

QwenPaw is a self-hosted personal AI assistant: you run it on your own machine, a container, or a cloud instance, then reach it through a web Console, a terminal UI, a beta desktop app, or messaging channels such as DingTalk, Lark, WeChat, Discord, and Telegram. Its main value is keeping one agent — with one memory and skill set — available across those entry points and for unattended tasks.

What it does

The assistant handles everyday work rather than one narrow job: scheduling recurring digests and broadcasts, reading and converting Office documents, searching the web, working in a file workspace, and editing project code. A three-layer memory keeps working context, verbatim history, and a self-evolving Markdown knowledge base powered by ReMe, so past conversations and resources become searchable notes you can edit yourself. Project overview Documentation index

Extensibility comes from Skills, a plugin marketplace, and MCP client support. Multi-agent features let you spawn sub-agents with their own memory and skills, and ACP handles cross-system orchestration. Execution is guarded by a kernel-level sandbox (Seatbelt, Bubblewrap/Landlock, AppContainer), plus tool-call rules that inspect commands before they run, file access blocks, and skill scanning before activation. Access policies can allow, deny, or request human approval per capability. Security section

Running it

There are several installation paths, each with its own prerequisites:

• pip: Python 3.11 to below 3.14, then qwenpaw init and qwenpaw app; the Console is served locally.
• Script installer: no Python pre-setup, but the script downloads uv and all dependencies, and the README warns it may fail behind restricted networks or corporate firewalls.
• Docker: official agentscope/qwenpaw images; separate volumes hold data, secrets, and backups. Connecting to a host Ollama or LM Studio requires explicit host networking configuration.
• Cloud: one-click Alibaba Cloud ECS or AgentScope Platform deployment, or a ModelScope Studio instance (which should be set non-public).
• Desktop app (Beta): Windows 10+ or macOS 14+, not fully tested across all hardware, with unoptimized startup and memory use; the macOS build is unnotarized, so Gatekeeper must be bypassed.

Building from source additionally requires building the console frontend with Node before installing the Python package, and major-version updates require rebuilding the frontend, reinstalling, restarting, and clearing browser cache. Quick start

Model choices and data

A cloud LLM provider needs an API key before chatting, and tools like web search may need their own keys. Local use through built-in QwenPaw Local (llama.cpp), Ollama, or LM Studio avoids API keys; Ollama requires a context length of at least 32k. Local-first deployment keeps data on your machine, but note that qwenpaw init collects anonymous per-version telemetry (version, install method, OS, Python, CPU architecture, GPU availability), and choosing interactive init lets you decide whether to opt in. Model setup Telemetry

The project is Apache-2.0 licensed. The README documents no benchmark results or independent security evaluation, so the security features should be read as described protections rather than verified guarantees.