One-line takeaway
The point is not that an AI can move a cursor. The point is that it can only touch the apps you explicitly hand it, in a window you can watch and stop.
Background
On November 19, 2026, Google brought Computer Use to the macOS version of the Gemini app. A user describes a task in plain language, and Gemini switches between desktop applications to click, type text, fill forms and navigate menus on its own.
The timing matters. The release arrived shortly before macOS 27.3 is expected to ship. According to Google and multiple reports, the new OS version will add risk disclosures for this class of agent, spelling out what an AI can see and what it can execute. The capability comes first; the system-level guardrails follow.
How it runs
| Item | Detail |
|---|---|
| Interaction | Natural language request, cross-app execution |
| Execution surface | Separate background window |
| Screen and cursor | Not occupied, user keeps working |
| Visibility | Live action preview inside the chat interface |
| Interruption | User can pause or stop at any time |
| Sensitive steps | Confirmation requested before execution |
| Permission scope | User selects which apps Gemini may access |
| Hardware | Apple Silicon Mac |
| OS | macOS Sequoia 15.0 or later |
| Account | 18+, eligible personal Google account |
| Plan | Google AI Plus, Pro or Ultra (US first) |
| Price | From USD 4.99 per month |
Background execution is the design decision that shapes everything else. Most automation tools take over the foreground: the cursor moves, windows steal focus, and the user simply waits. Gemini keeps the work in a separate window, leaves the main screen and pointer alone, and shows the action stream in the chat pane instead.
Four intended use cases
Google's examples are workflows rather than demos.
First, batch editing photos: apply a uniform edit across a folder of images instead of repeating the same action dozens of times.
Second, extracting information from a local PDF and entering it into an online form. The hard part here is crossing media: the data lives in a file, the destination lives in a browser, and a human has to shuttle between them. An agent can do it in one pass.
Third, building a presentation from live data, pulling figures and screenshots from several sources into slides.
Fourth, listing second-hand items: analyse photos of clothing in a local folder, organise the attributes, then fill in the listing form.
All four share the same shape: fixed steps, multiple apps, individually trivial actions, high total count. That is precisely where human time is wasted and where the cost of an AI mistake stays bounded. The reverse case, continuous judgement with irreversible consequences such as bulk deletion or changing system settings, does not appear in the official examples.
How permissions are structured
This is the part worth remembering. Gemini does not ask for control of the whole machine. It asks the user to choose which applications it may access. That is an app-level allowlist, distinct from a system-level grant: once a system grant is issued, an agent can act anywhere; an app-level list confines it to the programs the user ticked.
Two further constraints sit on top. Sensitive actions require confirmation before they run, and the live preview lets the user see what is happening and intervene when something looks wrong.
With macOS 27.3, a third layer arrives from the operating system: a description of the risk of restricted data access, an explicit statement of the agent's permission scope, and a visible indicator that one app is currently controlling another. Apps already installed and authorised will not need to be re-approved after 27.3 ships.
Reading it alongside the same day's news
Google was not alone in this space that week. According to multiple reports, Apple is developing its own computer-use agent capable of operating apps, browsing the web and writing code, still in an early stage, and is preparing additional permission warnings for competing agents. Meta's Muse drew criticism after private conversations surfaced in the wrong sessions: employees reported seeing other people's parenting messages, personal reminders and bank balances. Meta said the version being tested was not yet public.
Three stories, one question: once an agent holds a computer's controls, who defines the boundary. Google's answer is a user-managed allowlist plus live preview plus confirmation on sensitive steps. Apple's direction is a system-level warning. Meta's problem was session isolation, a different failure mode entirely.
Boundaries
- The release date, execution model, permission design, hardware and OS requirements, plan conditions and pricing come from Google's official materials and reporting by ITHome, PCOnline, ZOL and Google's Gemini help pages; the sources agree on the essentials;
- The macOS 27.3 risk disclosures and transparency indicator are pre-release descriptions; actual behaviour depends on the shipped system;
- Google has not disclosed the model version behind the feature, success rates, or rollback behaviour; this article does not speculate;
- Apple's in-house agent is press-reported and not officially confirmed; it is recorded here only as an industry development;
- The Meta Muse privacy reports come from employees and press, with Meta stating the tested build was not public;
- This article does not rate the feature's security or constitute usage advice.
Sources
- ITHome: Google Gemini can now operate a Mac directly
- PCOnline: Gemini can drive a Mac, with tasks running in the background
- ZOL: Gemini for Mac launches computer control with fine-grained permissions
- Google: Gemini for macOS product page
- Google Help: Using the Gemini app on Mac
- The Neuron: Everything that happened in AI today, November 20, 2026
