Google has brought computer use to the Gemini app for Mac, letting the AI assistant complete multi-step tasks across your local files, desktop apps and the web on your behalf. The Gemini computer use Mac rollout is limited to English in the United States for now, and it works inside an isolated background window while your own screen and cursor stay free. The important question is not just what it can do, but what you can see, pause and approve while it works, because this is an agent acting on your computer, not a chatbot answering in a window.
Gemini computer use Mac: what it is
Google announced the feature on 19 November 2026, in a short blog post by Michael Friedman. Computer use is a type of agentic capability: instead of only answering questions or generating text, Gemini can take actions directly on your Mac, navigating native Mac apps and browsing websites in the background to finish a task you described.
Google frames it as help with repetitive work. In the announcement, Friedman writes that computer use helps you get multi-step tasks done across your local files, desktop apps and the web, and that it saves time on repetitive work that normally takes dozens of clicks. You describe what you need, and Gemini handles the clicking, typing, menus and forms.
That distinction matters. A chatbot waits for your next prompt. Computer use is meant to carry a task through several steps that span tools which do not connect to each other, for example pulling information from a local file, using it in a desktop app, and then finishing the job on a website.
What it can do: multi-step tasks across files, apps and the web
Google suggested four kinds of task to start with. Together they illustrate the feature’s overall scope — multi-step work across local files, desktop apps and the web — without implying that every suggested task crosses all three in order.
- Speed up creative workflows: batch color-grade photos in your favourite design software, or set up complex 3D scenes.
- Skip manual form entry: extract weights, dimensions and addresses from local PDFs to fill out multi-box shipping quotes.
- Turn live data into ready-to-present decks: pull metrics and screenshots from live dashboards across multiple regions and build a presentation in Google Slides.
- Batch-list items for sale: analyse clothing photos from a local folder, research pricing, log item details and stage listings on resale platforms.
The Verge, reporting the same announcement, summarised the suggested uses more briefly as speeding up workflows, filling in forms, making presentations and listing items for sale. The common thread is repetition at scale: the same task repeated across a batch of files, the same form filled many times, or a set of listings prepared one after another.
What computer use does not promise, at least in the announcement, is open-ended autonomy. The tasks Google describes are ones you define and can watch. Gemini is not presented as deciding your goals for you; it is presented as doing the clicking, typing and navigating once you have described the outcome you want. The research for this article found no verified example of a complete end-to-end task beyond the four cases Google itself suggested, so readers should treat those four as Google’s suggested use cases, not as independently validated demonstrations, and anything more ambitious as unproven until Google or independent testing shows it working. The evidence supplied for this article contains no independent end-to-end validation.
How it works and what you can see while it works

Gemini does its work inside an isolated background window with a live preview right in your chat, according to Google. Your main screen and cursor remain completely free, so you can keep typing, designing or taking calls without interruption while Gemini works in the background. The Verge reports the same arrangement: users can follow along as Gemini works because it has its own background window that shows a live preview in the chat.
To start a task, you add @Computer use to your prompt, according to 9to5Google. That tells Gemini the request is a computer-use task rather than an ordinary chat request. In the app, under Customize and then Connectors, you will see new Desktop apps, such as Blender and Autodesk Fusion, joining the existing Git and Obsidian integrations. Those entries are what the reporting documents; the complete scope of apps computer use supports is not established in the evidence supplied for this article.
The live preview is the part of the design that matters most for trust. Because you watch every step live, in Google’s words, you can see what Gemini clicked, typed or opened as it happens, rather than receiving only a finished result. If the agent goes down the wrong path, you can see it going wrong while there is still time to intervene, which is the practical difference between a background window with a preview and an agent that works invisibly and reports back at the end.
Controls, permissions and safety

Google puts the controls in one short sentence in its announcement: you choose which apps Gemini can access, watch every step live, pause at any time, and approve key actions before they happen. TechCrunch reports the same three controls in plainer terms: users can pause, stop and approve Gemini actions before they happen. 9to5Google adds two specifics: you can stop Gemini at any time with one click, and Google will pause for logins and payments, with confirmation needed for sensitive actions.
That last point deserves emphasis, because it is the main safety boundary Google has described. An agent that can click and type can also, in principle, sign you into things, spend money or change files. Pausing for logins and payments and requiring confirmation for sensitive actions is meant to keep a human decision in front of the highest-stakes steps. It does not remove the need for care: you still need to read the confirmation, check what Gemini is about to do, and be ready to stop it if the preview shows something you did not ask for.
There is also a system-level gate before any of this works. AppleInsider reports that before Gemini can operate Mac apps, users must enable computer use and grant macOS Accessibility and Screen and System Audio Recording permissions if they are not already enabled. Those are powerful permissions, because they let software observe and control what happens on screen. Granting them is a deliberate decision, and it should be treated as one: enable computer use only on a Mac you control, for an account you trust, and review which apps you have allowed Gemini to access. The evidence supplied for this article does not establish how to revoke access, or what happens to an active task if a permission is revoked, so do not assume the permission steps above are a complete account of access control.
Two things remain unresolved in the reporting, and a careful reader should not assume either one. First, the subscription tiers: TechCrunch and 9to5Google report that computer use requires a Google AI Plus, Pro or Ultra subscription, while AppleInsider reports that Google support documentation currently specifies only Pro and Ultra for controlling desktop apps. Check the current support page before paying for a plan on the strength of this feature. Second, the announcement as reported does not settle what data Gemini sees, retains or sends to Google while working in the background window. The announcement and the research supplied for this article do not establish what data Gemini sees, retains or sends to Google while working, so treat the permission and approval controls above as the documented safeguards, and the data-handling picture as an open question.
Where Gemini fits among desktop agents
Google is not first here. TechCrunch notes that the company is a little late to the game with this launch, as other leading AI model makers like OpenAI, Anthropic and Meta have already been offering agentic features in their desktop applications for some time. The article does not compare rival products feature by feature, and neither should this one on the basis of the announcement alone: the verifiable point is timing. Desktop agents that act on a computer are already an established category, and Gemini computer use is Google bringing its Mac app into that category rather than inventing it.
The launch also sits next to a second, related rollout from the same week. TechCrunch reports that Google only this week announced the broader rollout of Gemini Tasks, an in-app feature that allows the AI to work more as an agent, helping users sign into websites, make online purchases or book appointments. How Tasks relates to computer use is not established in the evidence supplied for this article.
What can be said on the evidence is narrower: both announcements describe Gemini acting beyond answering inside its own window, and the relationship between Tasks and computer use remains an explicit open question. The controls described above — app choice, live preview, pause and approval of key actions — are the documented safeguards for computer use.
Availability and whether to try it
Computer use is rolling out in English in the United States in the Gemini app for Mac, according to Google. TechCrunch adds that Google did not say when it would expand to other markets or platforms, so readers outside the US, or on other platforms, should not expect a date yet. Multiple reports of the footnote to Google’s announcement also describe an Apple Silicon Mac running macOS Sequoia 15.0 or later and a personal Google account for users aged 18 and over; verify those requirements against Google’s current support page, since the reporting on subscription tiers already shows that announcement and support wording can differ.
Who should try it first? The clearest fit is a Mac user who already pays for a Google AI subscription, does repetitive multi-step work of the kind Google describes, and is willing to supervise an agent while it learns the shape of their workflows. Batch file work, repeated form filling and presentation building from live data are Google’s suggested use cases; the evidence supplied for this article contains no independent end-to-end validation of them.
Who should wait? Anyone uncomfortable granting Accessibility and screen-recording permissions to an AI app, anyone whose tasks involve sensitive accounts, payments or irreplaceable files, and anyone outside the US or on a subscription tier whose eligibility is in doubt should wait for clearer documentation and independent testing. The same applies if you cannot watch the live preview while Gemini works: watching the preview is a prudent recommendation alongside the documented controls — choosing which apps Gemini can access and approving key actions — rather than a complete safety model on its own.
If you do try it, start small and supervised. Confirm your subscription tier against the current support page, enable computer use, grant only the permissions asked for, and choose a narrow set of apps for Gemini to access before you widen it. Begin with a low-stakes, repeatable task, such as filling in a form from a document you can check afterwards, one of Google’s suggested uses, treated as a supervised trial rather than a demonstration that the feature works end to end, and watch the live preview from start to finish. Approve or stop each key action deliberately. If that first supervised run does what you asked, you have a grounded reason to trust the second one a little further. If it does not, the preview can help you notice unwanted actions, although it does not guarantee you will pinpoint exactly where it went wrong.












