New on GitHub
Sources 33Updates always the freshest
0 Newsletter
Animated illustration for: Gemini can now click through your Mac apps while you keep workingGeminiform.pdf4MacEN · US2026-11-19gemini.google/mac
AI News · Nov 21, 2026 · 3 min

Gemini can now click through your Mac apps while you keep working

You describe a task and Gemini works through files, apps and websites on your Mac. I show what it is good for, how to stay in control and what to watch out for.

In short
  • On 19 November Google added computer use to the Gemini app for Mac.
  • Gemini clicks and fills in forms in a separate background window while you keep working.
  • You pick allowed apps and approve key steps; at launch it is English-only in the US.

On 19 November Google added a feature to the Gemini app for Mac: you describe a task and Gemini works through local files, apps and websites on its own, clicking, typing, choosing from menus and filling in forms. It does this in a separate background window, so you can keep working in the meantime. At launch it only works in English in the US, but the direction is clear: the assistant stops just giving advice and starts doing the work.

How it works in practice

The feature is called computer use: AI operates the computer the way a person does, looking at the screen and using the mouse and keyboard. Michael Friedman from the Gemini team writes that the agent works in an isolated window in the background, with a live preview in the chat. Your main screen and cursor stay free, so you can type, design or take a call.

Google lists four safeguards:

  • you choose which apps Gemini can access,
  • you watch every step live,
  • you can pause at any moment,
  • you approve key actions before they happen.

Four jobs Google uses as examples

The examples in the announcement show well who this is for. A photographer can hand over batch colour grading in a design app. An online shop can have it pull weights, dimensions and addresses from local PDFs and type them into a shipping quote form. Someone who builds reports can collect metrics and screenshots from dashboards for several regions and turn them into a Google Slides deck. And a second-hand clothing seller can give Gemini a folder of photos: the agent looks at them, researches prices, logs the details and stages listings on resale platforms.

The common thread is work that takes dozens of clicks but little thinking. In my view that is exactly where an AI agent pays back fastest, because the result is easy to check.

Where to draw the line

Approving steps is a good idea, but it is easy to start clicking “yes” without reading. A few rules worth applying to any agent with access to a computer:

  • give access only to the apps the task needs, not the whole system,
  • always stop before sending, paying or publishing, even when full automation is tempting,
  • start with a task on a copy of the data and compare the result with what you would have done.

There is a second thread from the same week. On 20 November Signal wrote that it cannot reliably block screenshots on macOS, and screenshots are how most agents read apps. Google does not claim Gemini reads Signal, but the general point stands: an agent sees whatever the apps you let it into show. Choose them deliberately and close windows with data the agent does not need.

Availability outside the US

For now the feature is reaching English-language users in the US, and the app is downloaded from gemini.google/mac. Google has not given a date for other countries or languages. If you work on a Mac outside the US, this is a good moment to list two or three repetitive tasks from your week and decide which could safely go to an agent once the feature arrives. I write more about splitting work between people and AI in the Human in the loop hub.

Sources

Human in the loop

Related