Gemini Can Now Drive Your Mac From a Separate Window — and It Can Read the Verification Codes on Your Screen
TL;DR
On November 19, 2026, Google added computer use to the Gemini app for Mac. You describe a task in plain language; Gemini clicks menus, types, and fills in forms to finish multi-step work that spans local files, desktop apps, and websites.
The part that sets it apart from other desktop agents is that it doesn't take over your screen. Gemini runs inside an isolated window with a live preview in the chat, leaving your main display and cursor free so you can keep typing, designing, or taking calls.
One caveat you should know before paying: Google's blog post says the feature is available to Google AI Plus, Pro, and Ultra subscribers, while Google's own support documentation currently lists only Pro and Ultra. The two accounts don't match, so anyone upgrading to Plus specifically for this should verify first.
This article is compiled from Google's official announcement and same-day reporting by TechCrunch, Bloomberg, The Verge, AppleInsider, 9to5Google, and Android Authority.
What actually shipped
This is a feature added to the existing Mac app, not a new product. The download is still at gemini.google/mac; you turn computer use on inside the app after updating.
What it does:
- Cross-app execution. Gemini navigates menus, clicks buttons, types text, and switches between apps until the job is done. Google's own examples are concrete: batch color-grading photos and building complex 3D scenes in design software; pulling weights, dimensions, and addresses out of local PDFs to fill multi-box shipping quotes; grabbing metrics and screenshots from live dashboards across regions and assembling a Google Slides deck; analyzing clothing photos from a local folder, researching prices, logging details, and staging listings on resale platforms.
- Isolated window with a live preview. Every click and keystroke happens in a separate background window, mirrored in the chat. That is the core difference from screen-hijacking desktop agents.
- Layered permissions. Before anything runs, you choose which macOS apps Gemini may touch. You can watch each step live, pause at any moment, and approve sensitive actions before they fire.
- System permissions. According to Google's support documentation as cited by AppleInsider, first use requires granting Accessibility and Screen & System Audio Recording. You can also pick "Always allow" for an individual app so it stops asking.
Availability is narrow, and Google's footnote is explicit:
- Hardware: Apple Silicon Macs only;
- OS: macOS Sequoia 15.0 or later;
- Region and language: United States, English;
- Account: personal Google accounts, 18 and over.
Intel Macs, non-US accounts, and non-English interfaces are out for now. Google has not said when it will expand.
Why it matters
First, it targets the most annoying part of desktop agents: losing your screen.
Desktop agent demos have been impressive for a year, but the daily experience is rough — the moment the agent starts working, you can't. Your cursor is captured, windows jump around, and a fifteen-minute task means fifteen minutes of waiting. Putting execution in a background isolated window makes "human and agent working in parallel" the default behavior instead of something you solve by buying a second machine.
Second, Google is the latecomer here.
TechCrunch's report says it plainly: OpenAI, Anthropic, and Meta have offered agentic features in their desktop apps for a while, so this is Google catching up. Latecomers rarely win on capability; they win on friction. The isolated window is exactly that kind of move.
Third, it's converging with Gemini's Tasks.
The same TechCrunch piece notes that Google only recently widened access to "Tasks" inside the Gemini app — the feature that signs into websites, makes purchases, and books appointments for you. Computer use handles acting on your machine; Tasks handles acting on the web. Once those two connect, you describe an outcome and the system runs the whole chain.
The counterintuitive part: granting access reaches further than you think
This is the section worth remembering.
Google's own support documentation, quoted in full by AppleInsider, flags two things:
- Gemini can enter native Mac app sessions that are already signed in. If your email, admin console, or internal company tool is logged in on the desktop, it is potentially within reach.
- Gemini can read and act on multi-factor authentication codes visible in desktop or messaging apps. The check designed to prove "this is really you" gets weaker once an AI can see the screen.
There is also a risk shared by every computer-use product: prompt injection. The PDF or web page you ask it to read may contain instructions a human can't see but the model treats as commands, steering it toward actions you never requested.
Then add "Always allow," which persists per app. One convenient click and the grant stays.
So the useful question isn't "is this feature safe." It's "how much room am I giving it, and for how long."
How to use it: five steps to a setup you can live with
Step 1 — Confirm you actually qualify. Apple Silicon Mac, macOS Sequoia 15.0 or later, a US personal account, and a subscription tier that really covers the feature. Pro or Ultra is the safest read right now; the Plus situation is contradictory, so verify before upgrading.
Step 2 — Grant only the system permissions it needs. Under System Settings > Privacy & Security, give Gemini Accessibility and Screen Recording. Don't hand over Full Disk Access along the way — that's a different level of trust and this feature doesn't need it.
Step 3 — Start with low-risk apps. Don't tick your email, company console, and payment apps on day one. Run something with no consequences first — Finder plus a folder of images — and watch how it moves in the preview window.
Step 4 — Be stingy with "Always allow." For anything that sends, deletes, spends, or changes credentials, keep per-action approval. Reserve standing grants for read-only tools like image viewers and document readers.
Step 5 — Keep an eye on the preview during long runs. The isolated window means you can work in parallel, not that you can walk away. The pause button exists so you can stop it the moment it drifts.
Who it fits, and who should wait
A good fit right now:
- You have work that means clicking the same things dozens of times: batch photo edits, repetitive form filling, folder cleanup.
- You need data scattered across dashboards and local files assembled into one deliverable — Google's own example is regional dashboards into a Slides deck.
- You work on a Mac and are willing to spend ten minutes configuring permissions properly.
Not a good fit yet:
- You're on an Intel Mac or below macOS 15.0. It simply won't install.
- You're outside the US or don't use an English interface. Google has committed to US English only and given no expansion timeline.
- You want it inside core company systems. It can see signed-in sessions and MFA codes; that risk level doesn't belong in production before an internal review.
- You want a fully autonomous agent you can ignore. This product assumes you stay in the loop by design.
- You need dependable access from mainland China. This is a consumer subscription product; availability depends on service status in your region.
Pitfalls and cost
Pitfall 1 — Treating the blog post as final. The announcement lists AI Plus, Pro, and Ultra; the support documentation currently lists only Pro and Ultra for controlling desktop apps. Before you pay, open your own account and check whether the feature actually appears.
Pitfall 2 — Assuming permission is a one-time thing. Access is granted per app, and "Always allow" persists. Review those rules the day you set it up, not after something goes wrong.
Pitfall 3 — Expecting a smarter chatbot. This is an execution layer, not a knowledge layer. The more specific the task — which folder, which fields, what output format — the higher the success rate.
Cost (US list pricing, captured November 19, 2026):
- Google AI Plus: $4.99/month;
- Google AI Pro: $19.99/month;
- Google AI Ultra: from $99.99/month (with a $199.99/month higher tier).
If you're subscribing for this one feature, Plus is the cheapest door — and it happens to be the tier with contradictory documentation. The safer route is a month of Pro, two weeks of real use, then a decision on Ultra.
What Google hasn't said: whether computer-use calls have a separate quota, how long operation logs from the isolated window are retained, and when Workspace accounts get access. None of those numbers appear in the announcement or the pricing page.
One next step
Don't upgrade yet. Open the Gemini app on your Mac and see whether the computer-use entry point is there. If it is, give it exactly one job — renaming images in a folder by date — and watch the whole run in the preview window. Your own read on where its boundaries sit will beat any review.
Sources
- https://blog.google/innovation-and-ai/products/gemini-app/computer-use-mac/
- https://gemini.google/mac
- https://gemini.google/subscriptions/
- https://techcrunch.com/2026/11/19/google-gives-gemini-the-ability-to-use-your-mac-for-you/
- https://appleinsider.com/articles/26/11/19/gemini-can-now-control-mac-apps-but-you-still-shouldnt-let-it