Gemini can now click and type in your Mac apps for you, if you let it
Google's Gemini app for Mac can now run multi-step chores in your apps and on websites. It starts with paying US subscribers, and you decide which apps it may touch.
- On November 19, Google gave the Gemini app for Mac "computer use": it can click, type and fill in forms across your apps, files and websites1,3.
- It works in a separate background window you can watch live, so your own screen and mouse stay free1,3.
- You pick which apps it can use, can pause or stop it any time, and approve key actions first1,2.
- For now: US, English, adults with a paid Google AI Plus, Pro or Ultra plan, on an Apple Silicon Mac with macOS Sequoia or later1.
Should you care?
3/5 Read this before you switch it on
How hot right now
HotAnnounced yesterday; Google's post plus steady coverage from tech and Apple sites
Fire meter checked Nov 20, 2026
Hype check
We'll check back on Feb 18, 2027: did it live up to the hype?
Who is this for?
- ✓US Mac owners who pay for Google AI Plus, Pro or UltraFor youYou can try it now, as long as your Mac has Apple Silicon and runs macOS Sequoia 15.0 or later1.
- ~People drowning in repetitive computer choresMaybeGoogle pitches it for batches of files, form filling and building slides from dashboards1, but test it on copies first.
- ~Anyone who keeps banking, health or work apps on their MacMaybeThink carefully about which apps you allow; Google's own help pages warn it can make mistakes5.
- ✕Gemini users outside the USSkip itIt is rolling out in English in the US only, and Google hasn't said when other countries get it1,2.
- ✕Windows PC and free Gemini usersSkip itIt's only in the Mac app and only for paid plans1,2.
Gemini can now do the clicking for you on a Mac
Google's Gemini app for Mac can now use your apps and websites on its own to finish chores you describe. You choose which apps it can touch and can stop it any time.
Until now, an AI chatbot could tell you how to do something on your computer. Google's Gemini app for Mac can now do it for you. Google calls this "computer useJargon busterWhen an AI looks at a screen and clicks and types in apps itself, like a person would.", and it arrived on November 191,2.
You describe a job in plain words, and Gemini does the clicking, typing, menus and form filling, across your files, your Mac apps and websites1,3. Google's examples include pulling addresses and package sizes out of PDF files to fill in shipping quotes, and turning numbers from online dashboards into a slideshow in Google Slides1.
Think of it like a helper sitting at a second computer next to yours. Gemini works in its own separate window in the background, and a small live preview in the chat shows you what it's doing. Meanwhile your own screen and mouse are free, so you can keep working1,3.
Can it do things without asking?
Not everything. You decide which apps Gemini is allowed to use, you can watch each step, pause or stop it, and you approve the important actions before they happen1,2. According to Google's help pages, Gemini is designed to ask first before steps like paying for something, permanently deleting files or sharing sensitive information4,5. Google hasn't published a list of apps that are always off limits: the choice is yours.
Google's own advice is worth repeating: Gemini can make mistakes, so don't give it files you can't afford to lose, and don't type passwords or card numbers into the task5.
Who gets it?
For now, only adults in the US using Gemini in English, with a personal Google account and a paid Google AI Plus, Pro or Ultra plan. The Mac needs an Apple SiliconJargon busterApple's own M-series chips (M1 and later) used in Macs since 2020. chip (M1 or newer) and macOS SequoiaJargon busterThe version of the Mac's operating system numbered 15. 15.0 or later1.
If that isn't you, there's nothing to do. If it is, start small: allow one app, on a copy of your files, and watch what it does.
What should you do?
Google catches up on desktop agents, with the brakes on
OpenAI, Anthropic and Meta already had agents in their desktop apps. Google's twist is a background window, app-by-app permission and step-by-step approval.
Late to the party
TechCrunch points out that OpenAI, Anthropic and Meta have offered agent features in their desktop apps for a while, so Google is arriving late2. It's part of a busy month for AI agents. Days before, Google widened the rollout of Gemini's Tasks, which can sign into sites, book appointments and buy things online2, and on November 13 OpenAI launched ChatGPT Wallet so its agents can pay at online stores (see our ChatGPT Wallet story).
On the Mac, computer useJargon busterWhen an AI looks at a screen and clicks and types in apps itself, like a person would. builds on the app's existing Task mode, previously called Gemini Spark. That could already work with folders you connect and with Google apps like Gmail and Calendar; now it can also drive desktop apps and websites3,5.
What's actually new
The background window is the clever part. An agent that drives your real screen has to borrow your mouse while it works, so you sit and watch. Google says Gemini's agent runs separately, with a preview in the chat, leaving your screen and cursor alone1,3. That makes it more like handing a chore to an assistant than lending them your keyboard.
The pitch is boring jobs done in bulk: color-grading a folder of photos, setting up 3D scenes in design software, or photographing clothes and staging resale listings with researched prices1.
The permission model
Gemini's agent permissions use the same choices across its apps: Don't allow, Allow once, Allow for this chat (Google apps only) or Always allow, which also covers future task threads4. Google says built-in safeguards may still stop to check with you on some actions even when something is set to Always allow4. For files, Gemini only sees folders you add, asks before permanently deleting or sharing files, and keeps temporary backups that vanish after 24 hours or when you start a new task5.
Our take: the risk isn't Gemini going rogue, it's Gemini being tricked or simply wrong inside an app where you're already signed in. App-by-app permission is the right design, but it only helps if you keep the list short. "Always allow" on your email or messaging app is a lot of trust to hand out.
What to watch
Google hasn't said when computer use reaches other countries or platforms2. Also watch whether it gets cheaper plans or moves beyond its "key actions" approvals into running unattended.
By the numbers
How we got here
How a computer-use agent drives a Mac, and what Google hasn't said
Computer-use agents work in a loop: look at the screen, pick one action, do it, look again. On macOS that normally needs two privacy permissions the user grants by hand.
The loop
Google hasn't said which model powers the Mac feature, but it has documented how its computer-use tool works since the Gemini 2.5 Computer Use model went into developer preview on October 7, 20256:
| Step | What happens |
|---|---|
| Input | The user's request, a screenshot of the current screen and a history of recent actions6 |
| Output | Usually one function callJargon busterA structured instruction an AI model sends to an app, such as "click here", which the app then carries out. for a UI action, such as a click or typing; for some actions, like purchases, a request for the user to confirm6 |
| Execute | The client code, not the model, performs the action6 |
| Feedback | A new screenshot (and, in a browser, the current URL) goes back to the model, and the loop repeats6 |
| Stop | Task done, an error, or a safety response or user decision ends it6 |
The model only ever sees pixels and returns intents; the app around it does the clicking. That split is where safety hooks live. Google described an out-of-model, per-step safety service that checks each proposed action before it runs, plus system instructions that can force refusals or confirmations for high-stakes steps6.
Notably, Google said in 2025 that the model was mainly tuned for web browsers and was "not yet optimized for desktop OS-level control"6. That changed on paper on June 24, 2026, when Google made computer useJargon busterWhen an AI looks at a screen and clicks and types in apps itself, like a person would. a built-in tool of Gemini 3.5 Flash and said it covers desktop as well as browser and mobile environments9. Native Mac apps, with their menus, palettes and dialogs, are the harder case, and the Mac announcement doesn't say whether 3.5 Flash is what's driving them.
What macOS demands
On a Mac, an outside app can't simply control others. Apps that drive the Mac through accessibility features must be granted access by the user in System Settings > Privacy & Security > Accessibility, and that switch can be turned off at any time7. Seeing the screen needs a separate grant under Screen & System Audio Recording8. These are system-level gates, independent of Gemini's own per-app choices, and the place to cut it off completely.
Open questions
- How the background window works. Google says the agent works in a walled-off window you can follow from the chat1, but not how it renders or captures apps there without touching your session.
- Prompt injection. Every screenshot-driven agent reads untrusted text, from web pages to emails, that could contain instructions aimed at it. Google's 2025 model was trained against injections and scams6, and Google says 3.5 Flash got targeted adversarial training against them9; the Mac feature's specific defenses aren't documented.
- Data handling. Google hasn't said in the announcement whether screenshots from these sessions are stored or how long. Its earlier change to Gemini chat retention (see our story on that) is the closest guide, but it doesn't address screenshots.
- No blocklist published. Google lists no apps that are always off limits; control rests on what you allow1,4.
The facts
Hover or tap anything underlined to see what it means.
- What
- Computer useJargon busterWhen an AI looks at a screen and clicks and types in apps itself, like a person would. in the Gemini app for Mac1
- Maker
- Google1
- Where
- US, in English1
- Plans
- Google AI Plus, Pro or Ultra1
- Account
- Personal Google account, 18+1
- Mac
- Apple SiliconJargon busterApple's own M-series chips (M1 and later) used in Macs since 2020., macOS SequoiaJargon busterThe version of the Mac's operating system numbered 15. 15.0 or later1
- Runs in
- Separate background window, live preview in chat1
- Controls
- Choose apps, pause or stop, approve key actions1
- Windows PC
- No date for other platforms2
Jargon buster
- Computer use
- When an AI looks at a screen and clicks and types in apps itself, like a person would.
- AI agent
- An AI that carries out tasks step by step on its own, instead of only answering questions.
- Apple Silicon
- Apple's own M-series chips (M1 and later) used in Macs since 2020.
- macOS Sequoia
- The version of the Mac's operating system numbered 15.
- Prompt injection
- Hidden instructions in a web page, email or file that try to trick an AI into doing something you didn't ask for.
- Function call
- A structured instruction an AI model sends to an app, such as "click here", which the app then carries out.
- Accessibility permission
- A Mac setting that lets an app control other apps, originally meant for assistive tools.
- Screen recording permission
- A Mac setting that lets an app see or record what's on your screen.
Sources9 pages · Google, TechCrunch, 9to5Google and moreShowHide
The small numbers in the text point to these pages. Everything is written in our own words. How we work · Spotted a mistake? Tell us
- 1Let Gemini tackle repetitive chores on your Mac · Google
- 2Google gives Gemini the ability to use your Mac for you · TechCrunch
- 3Gemini app for Mac adds computer use to control desktop apps · 9to5Google
- 4Manage browser & app permissions for tasks in Gemini · Google (Gemini Apps Help)
- 5Assign tasks in the Gemini app on Mac · Google (Gemini Apps Help)
- 6Introducing the Gemini 2.5 Computer Use model · Google
- 7Allow accessibility apps to access your Mac · Apple Support
- 8Control access to screen and system audio recording on Mac · Apple Support
- 9Introducing computer use in Gemini 3.5 Flash · Google
Read next
AI · ★ Saved
Gemini will now hold on to your "off the record" chats for about two months, not 72 hours
Google will keep Gemini chats for about two months even when you've turned history off or used a Temporary Chat. In return, you get a separate switch to stop Google training its AI on what you type.
AI · ★ Saved
Google's Gemini 4 Carbon is out, but mostly for paying subscribers and developers
Google released Gemini 4 Carbon, its most powerful AI so far. Unless you pay for Google's AI plans or write software, your Gemini app and Search won't change.
AI · ★ Saved
ChatGPT Wallet lets OpenAI's AI agents buy things for you, with a Face ID check each time
OpenAI gave its AI helpers a way to pay for things with your credit card. You still say yes to every purchase with your face or fingerprint, and only a small group of paying US users can try it.