Mage-VL-Cowork Computer Use

Mage-VL-Cowork Computer Use

Lets you give a task to the Mage-VL-Cowork model and watch it click, type, scroll and use a terminal on a disposable Linux desktop. It suits developers studying computer use agents.

Made by nielsr on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: Mage-VL-Cowork Computer Use is a Gradio app that shows a 4B parameter agent working a Linux desktop. You give the microsoft/Mage-VL-Cowork model a task, and it clicks, types, scrolls and uses the terminal. Each run starts its own throwaway desktop in a Hugging Face Sandbox.

Who it suits: Developers and researchers who want to watch a computer use agent at work and read how the pieces connect.

Industry | Who uses it | How they use it | Example
AI engineering | Agent builders | Watch a small model drive a desktop | A short desktop task
Research | ML researchers | Study how screenshots and action summaries form its memory | Reading agent.py and prompt.py
Software | Developers | Read how a desktop sandbox is wired to a model | Opening desktop.py
Education | Instructors | Demonstrate computer use agents live | A classroom demo

If you only want a chat model, you can skip it, since this app needs a sandbox and a token to run a copy.

Where to use it:

• A web browser, on the Space's page at Hugging Face
• Python 3.12 on your own Windows, macOS or Linux computer, using a downloaded copy

Use case | Who | What they get back
Watching an agent | Developers | A live view of clicks, typing and scrolling
Terminal tasks | Researchers | An agent that also runs commands in a bash terminal
Code study | Learners | Separate files for the agent, desktop, model and prompt

How to set it up:

• Python 3.12
• Python packages from requirements.txt, with Gradio 6.31.0
• HF_TOKEN set as a secret, with permission to run Jobs, because sandboxes are billed to its namespace
• Optional settings: SANDBOX_NAMESPACE to bill an organisation, DESKTOP_IMAGE to pick the desktop image, SANDBOX_FLAVOR for sandbox hardware, which defaults to cpu-upgrade

1. Install the Hugging Face hub package.
2. Download the Space with the hf command.
3. Move into the downloaded folder.
4. Install Gradio 6.31.0, then the requirements.
5. Set the token secret, then start app.py.

pip install -U huggingface_hub
hf download nielsr/mage-vl-cowork-computer-use --repo-type space --local-dir mage-vl-cowork-computer-use
cd mage-vl-cowork-computer-use
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: Online, open the Space page, give the agent a task and watch it work on the desktop. On your own computer, start the app with the command below once the token is set.

python app.py

Tool | What it does
app.py | The main Gradio file that starts the app
agent.py | Agent logic, as the file name suggests
desktop.py | Code for the sandbox desktop
model.py and prompt.py | Model loading and the system prompt files

Example: You type a task for the agent and start the run. A fresh Linux desktop opens in a sandbox and you watch the model click, type and scroll, and use the terminal, until it finishes or stops.

For newcomers and experts: Open the Space page and give the agent a short, harmless task first. You can follow each click on the desktop, which makes it easy to see what the model is doing.

The model is served through the Transformers library on its own L4 graphics card, and its system prompt mixes two modes, computer_use and bash. Its working memory holds the three most recent screenshots, while older steps shrink into one line Action summaries. The desktop is Xvfb, XFCE and Firefox, driven with xdotool. If sandboxes fail with an image not found error, the source says to factory reboot the desktop image Space.

You give the agent a task
A disposable Linux desktop boots in a sandbox
The model reads screenshots and picks actions
xdotool carries them out on the desktop

Good to know:

• Each run boots its own disposable desktop, so nothing lasts between runs
• Model: runs on a dedicated L4 GPU through Transformers
• License: apache-2
• Cost to run: HF_TOKEN is needed, sandboxes are billed to the token owner or the chosen organisation, and the model uses a dedicated L4 GPU
• Data it reaches: a Linux desktop with Firefox and a terminal, plus the task you type
• Not for: the source names no limits

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/nielsr/mage-vl-cowork-computer-use
• Files: https://huggingface.co/spaces/nielsr/mage-vl-cowork-computer-use/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.