QwenPaw: Pick it and plan

This is the first of eight pages for readers who chose QwenPaw instead of Agent Zero. It covers who QwenPaw suits, what to have ready, which models it can use and what they cost, and the catches to know before you install anything on the next page.

QwenPaw is a free agent from AgentScope, a project run by Alibaba. Its license is Apache-2.0, an open-source license, and before April 2026 it was called CoPaw.

QwenPaw runs on your real computer, not in a container the way Agent Zero does. It has a few safety walls of its own, but all of them have gaps, so the safety section on the next page matters more than anything else in these eight pages.

Pick this if

• You want a model that runs on your own computer at no cost per use. QwenPaw has its own small Qwen models built in, which you download with a click and use without a key.
• You want to reach your agent from chat apps. QwenPaw connects to many, especially DingTalk, Lark (also called Feishu), WeChat, QQ, and iMessage, and Discord and Telegram are on the list too.
• You want a double-click desktop app on Windows or Mac. The app is marked Beta, which means it is still being tested.

QwenPaw was not one of the six agents this guide scored. The research looked at it but did not give it points, so this page cannot tell you how it compares. The notes are in the landscape chapter.

The project is young: its GitHub page was created in February 2026, and the newest release, v2.3.0, came out on 11 November 2026.

Before you start

The QwenPaw program is free, and the model is a separate choice. Read Pick your model and budget if you have not already.

The built-in local models are free too, but a small computer makes a weak agent. The research on local models says a computer with 8 GB or 16 GB of memory can chat but is weak at multi-step work, where a later step needs the result of an earlier one. About 24 GB is the first size with real room, and the details are in the local models chapter.

QwenPaw says its own small models, called QwenPaw-Flash, were trained for agent tasks. That is QwenPaw's claim, and the research did not test those models.

If your computer is small, a free Google Gemini key is a stronger start, and the next page shows both. Either way, use your everyday account on the computer, not an administrator account.

The models it can use and what they cost

QwenPaw itself costs nothing, so the model is the only thing that can cost money. It works with three kinds of model, and you can mix them.

• The built-in local models. QwenPaw Local runs a model on your own computer with llama.cpp, with no key and no bill. The FAQ lists QwenPaw-Flash in 2B, 4B, and 9B sizes, each with smaller 4-bit and 8-bit versions for computers with less graphics memory, and the research notes also list 35B-A3B and 27B sizes.
• Ollama or LM Studio. Both are free and both appear in the same Models list. The FAQ says you must set the model's context length to 32K or higher in those apps first, or QwenPaw may forget earlier instructions and fail at multi-step tool work.
• Cloud models with a key. The README names DashScope (Qwen), OpenAI, Anthropic, Google Gemini, DeepSeek, Kimi, and OpenRouter "and more," and its feature table says "14+ cloud providers." The config docs also list a Custom provider, where you type in your own Base URL.

A local model costs only electricity, and a free Google Gemini key costs nothing, although Google may use free-tier prompts to improve its products. A paid key is pay-as-you-go, so the bill grows with every job, and the full price table is in Pick your model and budget.

QwenPaw also uses your model when you are not typing. Its long-term memory runs a job called Auto-Memory after every five of your messages, and another called Auto-Dream every night at 23:00, and both use the agent's own model by default. On a free tier, those jobs eat into your daily allowance, and on a paid key they add to the bill, so Files, the browser, and memory shows where to turn memory off.

The Token Usage page under Settings shows how many tokens you used, by day and by model. It is a record, not a limit, because QwenPaw has no dollar limit of its own.

You can also pick a different model for one conversation with the control at the top right of the Chat page, which leaves the agent's default unchanged.

The catches

• AUTO is the default. Until you switch to STRICT, Tool Guard only stops commands its rules flag.
• The sandbox switch starts off. In v2.3.0, Enable Sandbox Execution starts off, so commands run with no sandbox until you turn it on, and on Windows without administrator rights you get a weaker version that limits only writing.
• The whole disk is readable. The sandbox stops writing outside the agent's folder, but it does not stop reading.
• The internet is open. Nothing in the default setup limits where the agent can connect.
• No sandbox means no wall. If your system has no sandbox, the mode is none. On Windows, the full sandbox needs administrator rights, which this guide does not ask for.
• Runs with nobody watching can keep going. A scheduled run may log a danger and continue, so block the command tool before you schedule.
• Chat apps start open to anyone. Switch them to your own user ID first.
• No dollar limit inside QwenPaw. Any limit is the one you set at your provider.
• Memory uses your model in the background. Auto-Memory and Auto-Dream run on the agent's own model, even when you are not chatting.
• Small local models are weak agents. On 8 GB or 16 GB, expect chat, not long multi-step jobs. QwenPaw Local is still being tested, and Intel Macs may not run it.
• The app is Beta and unsigned. Windows warns that the installer is not signed, and Apple has not notarized the Mac app.
• Not scored here. QwenPaw was not one of the six agents this guide compared, so there is no score to lean on.
• Some things this guide could not confirm: whether the app runs on Windows ARM PCs; the exact card labels for the command and browser tools on the Tools page (the tool docs name them execute_shell_command and browser); whether every chat app's card shows the allow list settings; where phone-sent approval requests appear; whether schedules run while QwenPaw is closed; and which Gemini models are free today.

Sources

Checked 14 November 2026.

• QwenPaw research notes.
• Landscape chapter.
• Local models by hardware.
• QwenPaw README, for desktop Beta, local models, cloud providers, "14+ cloud providers".
• QwenPaw v2.3.0 release.
• QwenPaw FAQ, for QwenPaw-Flash sizes and versions, 32K context length, per-conversation model choice.
• QwenPaw models, for QwenPaw Local, Ollama, LM Studio.
• QwenPaw config and working folder, for built-in provider list, Custom provider.
• QwenPaw long-term memory, for Auto-Memory every five turns, Auto-Dream at 23:00, both on the agent's model.
• QwenPaw Console, for Token Usage page, per-conversation model control.
• QwenPaw v2.3.0 config source and Console text, for sandbox switch starts off.
• QwenPaw security, for Tool Guard default, sandbox defaults, Windows administrator note.

Next

Continue to Install and first job.