QwenPaw
A self-hosted personal AI assistant that runs locally or in the cloud, reachable from many chat platforms at once, with layered memory, a plugin and skill marketplace, and sandboxed tool execution.
QwenPaw keeps a three-layer memory of live working context, full conversation history, and a self-evolving personal knowledge base, so conversations and resources become searchable, editable Markdown memory over time. It can run fully locally with its own built-in runtime and small QwenPaw-Flash models, or connect to Ollama, LM Studio, or more than a dozen cloud model providers, and it is reachable from DingTalk, Lark, WeChat, Discord, Telegram, iMessage and QQ, plus a console, terminal UI and desktop app.
It runs dangerous commands through a kernel-level sandbox, a tool guard, a file guard and a skill scanner before allowing them, and can spawn independent sub-agents, each with its own memory and skills, coordinated through an agent communication protocol. Built-in skills cover scheduling, document reading and conversion (PDF, Word, Excel, PowerPoint), web search and video summarization, and a unified file workspace supports editing and diffing project and agent files; a plugin architecture with a marketplace extends these further, and MCP integration connects external tools.
When self-hosted, its data stays on the user's own machine. It is written in TypeScript and licensed under Apache-2.0.