CoPaw

CoPaw (the name QwenPaw also appears in the repository) is an open-source personal AI agent workstation, advocating "work for you, grow together with you". It can be deployed by users on their own machines or in the cloud, supports extending capabilities through skills and plugins, and connects to multiple chat channels.

Main Uses

CoPaw is aimed at personal assistant scenarios, serving as a long-running, extensible agent workbench. Users can use it to complete automation and scheduled tasks (such as news summaries, report generation, multi-channel broadcasting), code and development work (reading, editing, reviewing, and testing code in a unified file workspace), document processing (reading, writing, and converting PDF, Word, Excel, PowerPoint), information gathering (web search, subscription tracking, video summaries, personal knowledge base retrieval), and multi-channel operations (syncing or pushing alerts, summaries, or AI-generated content to DingTalk, Feishu, Discord, Telegram, etc., either across channels or per channel). Users can also combine built-in capabilities, plugins, and scheduled tasks into custom workflows.

Main Capabilities

According to official materials, CoPaw 2.0 is a rewritten version based on AgentScope 2.0, bringing the Agent OS architecture, Loop Engineering, Scroll Context, the ReMe v0.4 self-evolving personal knowledge base, and a built-in terminal UI. Among these, Agent OS provides three pillars around each agent: resources (transparently visible on disk), governance (allow/deny/ask/sandbox), and sandboxing (macOS / Linux / Windows); Drivers provide a protocol-neutral MCP / A2A / ACP connection layer, supporting encrypted credentials and a per-call policy gate; Loop Engineering provides advanced agent loop templates (such as Coding Mode, Mission Mode, etc.) and composable approval gates; Scroll Context persists every round of conversation, indexes eliminated rounds, and supports on-demand recall; ReMe v0.4 continuously transforms conversations and resources into readable, editable, searchable, and linkable Markdown memory; Terminal UI provides full-screen terminal chat, sharing the same agent, memory, and sessions with Console and various channels.

Version records show: v2.2.1 adds the Creator 1.2 blueprint workbench, per-agent model routing, and unified environment management; v2.2.0 adds self-hosted multi-user QwenPaw Hub, QwenPaw Mail, native QwenPaw Data, unified model routing, Creator 1.1, unified marketplace and Console, skills, channels, and reliability improvements; v2.1.0 adds QwenPaw OS Shell, unified Files workspace, QwenPaw Creator, Codex/Qoder Agent integration, Browser-use, Computer-use, workspace checkpoints, and long-conversation continuity; v2.0.1 adds the PawApp mini-program platform, user-editable Agent Modes, the Oh-My-Paw plugin, ReMe memory enhancement, and desktop UX improvements.

How to Use

There are seven installation methods:

1. Pip installation: requires Python >= 3.11 and < 3.14, then open in the browser http://127.0.0.1:8088/ to configure the model; to chat in DingTalk, Feishu, WeChat, etc., refer to the Channel setup documentation.
2. Script installation: no need to configure Python yourself; one command installs everything; the script automatically downloads uv (the Python package manager), creates a virtual environment, and installs CoPaw and its dependencies (including Node.js and frontend resources). The official note says it may not be applicable in restricted network environments or behind corporate firewalls.
3. Docker: images are located on Docker Hub (agentscope/qwenpaw), with tags latest (stable) and pre (PyPI pre-release); an Alibaba Cloud Container Registry ACR address is also provided: agentscope-registry.ap-southeast-1.cr.aliyuncs.com/agentscope/qwenpaw (same tags). After installation, open http://127.0.0.1:8088/ Use Console. Configuration, memory, and skills are stored in the qwenpaw-data volume; model provider settings and API keys are in the qwenpaw-secrets volume; backup archives are in the qwenpaw-backups volume. Passing in API keys (such as DASHSCOPE_API_KEY) can be done via -e VAR=value or --env-file .env. If connecting to services such as Ollama on the host machine, you can use an explicit host binding (change the Base URL to http://host.docker.internal:<port>, Ollama is usually 11434, LM Studio is http://host.docker.internal:1234/v1），或在 On Linux, use host network mode (no -p port mapping is needed, but container ports will be exposed on the host and may conflict).
4. Alibaba Cloud ECS one-click deployment: open the corresponding deployment link and follow the prompts.
5. AgentScope Platform: provides one-click cloud CoPaw deployment, plugin sharing, and a skill marketplace; the materials state it is free and online 24/7.
6. ModelScope: ModelScope Studio also supports cloud deployment, and the official guidance says you should set the Studio to non-public to prevent others from controlling your CoPaw.
7. Desktop app (Beta): the Tauri build can be obtained from the official download page. For Windows it is QwenPaw-Tauri-<version>-Windows-setup.exe, and for macOS it is QwenPaw-Tauri-<version>-macOS.zip (Apple Silicon recommended). The official description says zero configuration, cross-platform (Windows 10+, macOS 14+), and a visual interface; the first launch may take 10–60 seconds because it needs to initialize the Python environment and load dependencies. If macOS has not notarized it, it may prompt that the developer cannot be verified; you can open it by right-clicking, allow it in System Settings, or run xattr -cr "/Applications/QwenPaw Desktop.app" (the official description says this method is not recommended for most users).

After installation, you can configure models in Console → Settings → Models, add skills (PDF, Office, browser, news, etc.), set up channels (DingTalk, Feishu, Discord, etc.), and configure Cron scheduled tasks and automation. End users can run qwenpaw to open the full-screen chat TUI, which shares the same agent, memory, skills, MCP tools, and sessions with Console and IM channels, and supports streaming replies, slash commands (/help, /resume, /theme, as well as the agent's built-in /model, /clear, etc.), pasting files/long text as attachments, and inline tool permission prompts.

Applicable scenarios

Suitable for users who want to control the deployment of their personal AI assistant themselves, including local or cloud self-hosting, calling it anytime through chat apps, scheduled automated push, code and document processing, personal knowledge base accumulation, and scenarios requiring unified multi-channel operations.

Limitations andNotes

The materials mention: script installation may not be suitable for restricted networks or corporate firewalls; localhost in Docker refers to the container itself rather than the host machine; Windows LTSC or environments with strict security policies may be in Constrained Language Mode, causing scripts to fail to automatically write to Path, requiring manual configuration of environment variables; the desktop app is still Beta and has known limitations such as incomplete compatibility testing, performance still needing optimization, and some features being unstable or missing. Pip installation requires Python >= 3.11 and < 3.14. The materials do not provide pricing or fee information.