Skip to content
Sign up

An SI agent that works on your computer, as far as you allow

The SI agent in Minda Super Intelligence, the AI agent app for your computer, looks at your screen, opens sidebar programs and, once you turn on computer control, clicks and types in other Windows apps. You pick the level: Off, Look only, Ask each step (the default) or Full control. On Free, acting steps share the 10 daily agent actions.

Coming soon for Windows

Plan
Every plan; acting steps are agent actions (Free: 10 a day)
Works on
Windows for clicks and typing; files and programs on macOS and Linux
Your data
A screen look goes to the model you picked, after your OK each time
Limit
Its connectors need uv or Node.js installed; Minda installs neither

How it works

  1. Turn on computer control

    Press Turn on in Settings, Integrations, This computer. A dialog lists what the agent could do and the risks; Minda then adds Windows control and Desktop Commander.

  2. Pick a Computer level

    Choose Off, Look only, Ask each step or Full control from the chat box's Modes menu or in Settings. Full control asks you to confirm before it turns on.

  3. Watch, pause or undo

    A purple border and a bar at the bottom of the screen show the agent at work. Esc or Pause holds it, Stop ends it, and Undo restores changed files.

What can an AI agent do on my computer?

Minda's SI agent can look at your screen or one app window, open programs and documents, find files and, with computer control on, click, type and move files in other apps.

Looking and opening are built in. The agent looks only after your OK each time, and opens the programs on your sidebar and the documents in folders you share with it, each in its own app. A model that cannot see pictures gets the text Windows reads in the window instead, read on your computer.

Clicking and typing come from two connectors you turn on yourself: Windows control, which drives any Windows app, and Desktop Commander, which reads and writes files and starts programs on Windows, macOS and Linux.

How do the control levels work?

The Computer level decides which steps run and which wait for you, and the agent reads it again before every step.

Computer control5
  • Off: no computer tool is offered, file search included.

  • Look only: only the tools a connector marks read-only, such as a screenshot or a folder list; anything unmarked counts as acting.

  • Ask each step, the default: looks run, and each acting step waits for Allow or Deny, even in Auto mode.

  • Full control: acting steps run without asking.

  • Lowering the level in the middle of a task holds the very next step.

How do I stop or pause the agent on my screen?

A bar at the bottom of the screen the agent works on has Pause, Continue and Stop, and Esc pauses it between steps.

The bar floats over every app, never takes the focus from the app the agent types into, and shows only while the agent controls the screen, is paused or waits for your OK. When a step waits, the bar shows Allow and Deny with every word of what the step will do.

A purple border and a purple pointer mark the screen while the agent clicks and types, and the border is left out of the agent's own screenshots. After a pause, the agent looks at the screen again before its next step, since you may have changed something.

Can I undo what the AI changed on my computer?

Yes for files: before a step writes, moves or deletes a file, Minda keeps a copy, and Undo under that step puts it back.

Undo moves what is there now to the Recycle Bin or Trash and restores the copy: a moved file goes back, a new file goes to the Trash, a deleted file returns. Copies are emptied at the start of the next session, and a file over 200 MB or a folder of more than 2,000 files is not kept.

Clicks, typing and programs a step ran cannot be undone, so Ask each step is the safer level for work you cannot easily redo.

What it does not do

Where this version stops, in plain words, so you know before you rely on it.

  • Clicking and typing in other programs works on Windows only; on macOS and Linux the agent works with files and programs.

  • Screen looks need Screen Recording permission on macOS and return nothing on Linux with Wayland; reading text from a screen picture is Windows only.

  • Windows control needs uv and Desktop Commander needs Node.js; Minda installs neither for you.

  • Undo covers files, not clicks, typing or programs a step started.

Frequently asked questions

Is computer control on by default?

No. It stays off until you press Turn on in Settings or during setup, and a dialog first lists what it lets the agent do and the risks, with Cancel in focus. Turn off in the same place takes the reach away again. An organization can block it with the McpServersEnabled policy.

Can a web page take over the agent on my computer?

A page can try to steer it, which is why Ask each step is the default. Page text reaches the model fenced as data, not instructions, and every acting step on your computer waits for your Allow unless you chose Full control.

Which model does the agent use on my computer?

The one you pick in the chat: Ollama on your computer, your own OpenAI, Claude, Gemini, DeepSeek, OpenRouter or Qwen key, or Minda's hosted models on credits. With Keep SI on this computer turned on, only a local model answers.

Does computer control work on a Mac or Linux?

Partly. Clicking and typing in other apps works on Windows only. On macOS and Linux the SI agent works with files and starts programs through Desktop Commander, and it looks at the screen on macOS once you grant Screen Recording; Linux with Wayland gives it nothing to see.

Can I give my computer a task from my phone?

Yes. Pair your phone once, type a task in Minda on your phone, and your computer does it while you are away, with the answer back on the phone. Each task is signed on the phone and checked on the computer before it runs, and the phone can Allow, Deny or Stop it.

Minda Super Intelligence is coming soon for Windows.

Last updated