QwenPaw

• Running a personal assistant inside DingTalk, Feishu, or WeCom that answers questions and manages todos
• Posting a scheduled news digest or report to several chat channels every morning
• Triage of a Gmail or QQ Mail inbox, with replies to unknown recipients held for approval
• Running an assistant fully offline on a laptop with a downloaded QwenPaw-Flash model
• Giving a small team separate assistants on one shared server with QwenPaw Hub
• Converting and summarizing PDF, Word, Excel, and PowerPoint files from a chat message

QwenPaw covers much of the same ground as OpenClaw and Hermes Agent: a self-hosted assistant that lives in chat apps, keeps Markdown memory, runs scheduled jobs, and extends through skills and MCP. Choose it over those two for specific reasons. If your team works in DingTalk, Feishu, WeCom, WeChat, or QQ, it treats those as primary channels. If you want an agent that runs on a laptop with no API key, its own QwenPaw-Flash models and built-in local provider get you there from the Console. If you are wary of letting an agent run shell commands, its sandbox, guards, and approval policies are part of the core rather than optional add-ons. If you live in WhatsApp or Signal, need a signed and supported app, or want to expose it to untrusted users, it is not the right fit yet.

About QwenPaw

QwenPaw (short for Qwen Personal Agent Workstation) is a Python application built on the AgentScope framework. You install it with pip, a one-line script, Docker, or a beta desktop app for Windows and macOS, then configure models and channels in a local web Console at port 8088; a full-screen terminal UI and a REST API drive the same agent. One instance can answer in many chat apps at once. DingTalk is the recommended channel, and the docs also cover Feishu (Lark), WeCom, personal WeChat, QQ, Yuanbao, XiaoYi, Discord, Telegram, Slack, iMessage (macOS), Mattermost, Matrix, MQTT, Microsoft Bot Service, and phone calls through Twilio or SIP. Each agent has its own workspace, memory, skills, and chat history, and agents can spawn sub-agents and collaborate. Memory is a ReMe-powered, self-evolving knowledge base of readable Markdown files, with every turn persisted and older turns indexed for recall rather than summarized away. Beyond chat, it runs cron jobs and heartbeat digests, reads and writes PDF and Office files, browses the web, connects MCP servers, manages IMAP/SMTP mailboxes with approval rules, and (in beta, on the desktop app only) operates approved desktop applications. Three things set it apart from other self-hosted assistants. First, it ships its own QwenPaw-Flash models (2B, 4B, and 9B, with 4-bit and 8-bit builds) tuned for its agent tasks, plus a built-in llama.cpp provider that downloads the runtime and a model from the Console, so after setup it can run fully offline with no API key. Second, security is layered and built in: a tool guard checks tool calls before they run, a file guard blocks sensitive paths such as ~/.ssh, shell commands can run in a kernel-level sandbox (Seatbelt, Bubblewrap/Landlock, or AppContainer), and a skill scanner checks skills before activation. Third, version 2.2 added QwenPaw Hub, a self-hosted multi-user server that gives each team member a separate QwenPaw. The tradeoffs: there is no WhatsApp or Signal channel, the managed mail workflow supports only nine personal mail domains (mostly Chinese providers plus Gmail), the built-in local runtime and desktop app are still in testing, Hub is meant only for teams whose members trust one another, Console login is off until you enable it, and anonymous telemetry is on by default when you initialize with defaults.