Google has added a new capability to the Gemini app for macOS: with Gemini Computer Use, the assistant can operate programs on a Mac by itself as of November 19, 2026. Gemini clicks, types, scrolls, navigates menus and fills in forms across local files, desktop applications and websites. The work runs in a separate virtual window with a live preview in the chat, so the user’s screen and cursor stay free. At launch, however, the feature is limited to English in the United States and to paying personal subscribers.
Gemini Computer Use: what Google announced
According to the announcement on Google’s blog, the feature is aimed at repetitive tasks that normally take “dozens of clicks”. Users describe in natural language what needs doing, and Gemini works through the steps in menus, input fields and dialogs. Google gives four example scenarios: batch colour-grading photos in design software or setting up complex 3D scenes; extracting weights, dimensions and addresses from local PDFs to fill out shipping quotes; pulling metrics and screenshots from live dashboards into a Google Slides presentation; and analysing clothing photos from a local folder, researching prices and staging listings on resale platforms.
Technically, Gemini moves the work onto a virtual screen. A live preview appears in the chat and can be expanded to full screen; clicking “Stop task” ends the job at any time. According to the Gemini Apps release notes, the assistant pauses for logins and payments and may ask for confirmation before sensitive actions.
Requirements: Apple Silicon, US English, a subscription
The barriers to entry are high for now. According to Google’s help page, you need:
- a Mac with an Apple Silicon chip running macOS Sequoia 15.0 or later,
- a personal Google Account and a minimum age of 18,
- a Google AI Plus, Google AI Pro or Gemini AI Ultra subscription,
- the app set to English with a US location.
Work and school accounts are excluded for the time being. For better results, Google recommends Gemini Hydrogen 4 or newer models. The feature is switched on in the Mac app’s settings under “Computer use”; macOS requires the Accessibility and Screen & System Audio Recording permissions. Within a task, users invoke it with @Computer use. There is no date yet for other countries or languages.
Permissions and backups in detail
The Computer Use help page describes the permission model in more detail than the blog post. Gemini can only control apps that the user has granted access to. The first time it wants to use an app, it asks whether to deny, allow once or always allow. “Always allow” adds the app to the “Allowed apps” list, after which Gemini no longer asks. There is also a “Blocked apps” list for programs Gemini should never open.
Before Gemini changes files, it tries to create temporary backup copies; the “Use on-demand file backups” option must be enabled for that. Google itself adds caveats: this does not always work, and content in some apps cannot be backed up depending on their settings and permissions. Backups are kept for 24 hours or until the next prompt, whichever comes first. Anyone planning to rely on these copies as a safety net should keep that in mind.
Google spells out the risks itself
The most notable part of the help page is its section on risks. Google lists, unusually openly, what can go wrong when an AI agent runs on your own computer:
- Gemini can share information with desktop apps, not just with websites.
- The assistant has access to logged-in accounts and active sessions in native Mac apps, not only in the browser.
- Some actions cannot be undone despite the backup attempt, because the app in question does not allow it.
- Untrusted content on screen in one app can manipulate Gemini into jumping across application boundaries and acting in other allowed programs – a textbook prompt injection.
- Granting access to code or text editors and developer tools can enable shell execution or broader file system changes.
- Gemini can read or interact with multi-factor authentication prompts and codes visible in desktop or messaging apps.
- Automating certain third-party apps may conflict with their terms of service, and according to Google the user is responsible for that choice.
This list is the most important part of the announcement. It makes clear that the security model rests mainly on which programs users allow and on how closely they pay attention. An “always allowed” terminal emulator or IDE effectively hands far-reaching control of the system to an agent that can be manipulated through injected screen content.
What it means for developers and admins
For developers who use Gemini on a Mac, a few pragmatic rules follow. Editors, terminals and Git clients should start out on the block list, or at most be used in “allow once” mode. Password managers and apps that display one-time codes should stay blocked as well. Anyone combining tasks with web content – “copy the data from this page into that spreadsheet”, for example – should expect that crafted web pages or documents may contain instructions the agent mistakes for a request.
For corporate admins, the situation is manageable for now because Workspace accounts are excluded. That may change once Google widens the rollout. At that point the question becomes how allow lists can be managed centrally; the help page says nothing about this yet. Combining the feature with the app’s connectors, through which Gemini already reaches services such as Gmail or Calendar, also widens the attack surface.
Privacy remains a separate issue. How long Google keeps Gemini conversations changed only a few days ago; we covered the new Gemini chat history retention period in detail. Whether and to what extent screen content from Computer Use tasks ends up in those histories is not explained in detail on the help page.
Context: agents are moving onto the desktop
Computer use agents are not new; several vendors have shown browser and desktop variants in recent months. What is new is that Google builds the feature directly into its consumer Mac app and uses a virtual screen that lets users keep working in parallel. That solves a practical problem of earlier approaches, in which the agent took over the user’s mouse and keyboard.
How reliably Gemini Computer Use works day to day cannot be judged from the announcement: Google gives neither success rates nor benchmarks for the feature. According to the help page, quality also depends on the chosen model; our article on Gemini 4 Neon gives an overview of the current model family. Until independent hands-on reports arrive, the feature is mainly a tool for small, easily supervised routine tasks – not something to leave unattended with access to your development environment, mail client and password vault.
Conclusion
Gemini Computer Use brings a real desktop agent to Google’s Mac app: it operates programs in its own virtual window, asks for permission per app and tries to create backups. Availability is limited to US subscribers with Apple Silicon Macs. The most valuable part right now is Google’s own risk list: it shows that prompt injection via screen content, access to logged-in sessions and permitted developer tools are the real levers. Anyone testing the feature should start with a short allow list and consistently block sensitive programs.




