Google put computer use into the Gemini app for Mac on November 19, 2026. Gemini can now navigate local files, desktop apps and websites, and it does the clicking, typing and form entry in an isolated background window, according to Google's announcement by Gemini group product manager Michael Friedman. It is rolling out in English in the U.S. to paying subscribers on Apple Silicon Macs.
That makes the Gemini Mac app a direct rival to the desktop agents we already cover, including Claude's background computer use and Meta Muse for Mac. This post explains what Google actually said, what it left out, and how to decide whether to try it.
TL;DR: Gemini computer use on Mac
| Question | Answer |
|---|---|
| What is it? | A mode of the Gemini app for Mac that clicks, types and fills forms across local files, desktop apps and the web |
| When did it launch? | November 19, 2026, as a rollout |
| Who gets it? | Personal Google accounts aged 18+ on Google AI Plus, Pro or Ultra |
| What hardware? | Apple Silicon Mac with macOS Sequoia 15.0 or later |
| Where? | English, U.S. only for now |
| Does it block my screen? | No. It uses an isolated background window with a live preview |
| What controls do I have? | Pick which apps Gemini can reach, watch live, pause, approve key actions |
| Where do I get it? | gemini.google/mac |
What did Google announce for Gemini on Mac?
Google describes the feature as a way to automate repetitive daily tasks. You describe a multi-step workflow in plain language. Gemini then navigates menus, types information and clicks through apps to finish the work.
The key design choice is the window. Google says Gemini "works inside an isolated background window with a live preview right in your chat." Your main screen and cursor stay free, so you can keep typing, designing or taking calls while the agent works. Earlier computer-use products often took over the mouse, so this is the part of the launch that matters most day to day.
Google lists four starter examples in its post:
- Creative workflows: batch color-grade photos in your design software, or set up complex 3D scenes.
- Form entry: pull weights, dimensions and addresses from local PDFs and fill multi-box shipping quotes.
- Decks from live data: pull metrics and screenshots from dashboards across several regions and build a presentation in Google Slides.
- Resale listings: analyze clothing photos from a local folder, research prices, log item details and stage listings on resale platforms.
The hero demo in the post shows a prompt asking Gemini to take a child's butterfly sketch, model it and rig it in Blender.
Who can use Gemini computer use on Mac?
Access is narrow at launch. Google's footnote says the feature requires an Apple Silicon Mac running macOS Sequoia 15.0 or later. It is available for personal Google accounts aged 18 and over, on Google AI Plus, Pro and Ultra. The rollout is in English in the U.S.
Google points readers to gemini.google/mac to download the app. Treat Google's own post as the authority on the terms, because plan rules can change.
What the announcement does not say:
- No price change. It names three existing plans and does not add a new tier.
- No usage cap for computer-use tasks. Google gives no limits per day or per task.
- No support for Intel Macs, Windows or work and school accounts. The footnote says personal accounts.
- No benchmark numbers or error rates for the underlying model.
If you want the plan details, check Google's plan pages before you subscribe for this feature alone.
How do the safety controls work?
A door left ajar, a picture of the approval prompt that gates Gemini computer use on Mac
Google lists four controls, all in the user's hands:
- App permissions. You choose which apps Gemini can access.
- Live preview. You watch each step in the chat as it happens.
- Pause. You can stop the agent at any time.
- Approvals. Gemini asks before key actions.
This matches the pattern that every major desktop agent now follows, and it is the right pattern. The weak point is the phrase "key actions." Google's post does not list which actions need approval. A reader should assume that deleting files, sending messages, paying for things and submitting forms should all be gated, and should verify that on a harmless test before trusting the agent with real work.
The need for care is not theoretical. Reuters reported this week, as relayed by IBTimes, that Meta's Muse agent could read Signal messages through device permissions that users had granted at setup, which we covered in our Muse and Signal report. The lesson there applies to Gemini too: an agent that can see your screen can read whatever is on it. Apple's own permission model is the backstop, and we explain it in our guide to Full Disk Access and AI agents.
How does it compare with Claude and Muse on Mac?
All three now offer an agent that works on a Mac in the background. The table compares only what each company has published. Where a cell says "not stated," the vendor has not published the figure.
| Gemini (Google) | Claude (Anthropic) | Muse (Meta) | |
|---|---|---|---|
| Surface | Gemini app for Mac | Cowork and Claude Code on macOS | Muse for Mac |
| Background operation | Yes, isolated window | Yes, per our September coverage | See our Muse for Mac post |
| Approvals | Key actions | Per Anthropic's controls | Per Meta's controls |
| Access | AI Plus, Pro, Ultra; U.S. English | See Anthropic's docs | See Meta's docs |
| Published benchmark for this feature | Not stated | See Anthropic's launch post | Not stated |
For details on the other two, read our Claude background computer use post and our Muse for Mac desktop launch post. We do not rank them here, because Google has not published a measured result for the Mac feature and a ranking without numbers would be a guess.
What should you try first?
Start with a task that is boring, repeatable and easy to check. Google's own examples are good templates. A practical order:
- Form fill from a PDF. Give Gemini one PDF and a quote form with three fields. Check every value it enters.
- A read-only dashboard task. Ask it to collect numbers from a dashboard into a document. Nothing is sent or deleted.
- A batch job on copies. For photo or listing work, point it at a folder of copies, never the originals.
- Only then a task with side effects. Posting a listing or sending a deck should wait until you have seen how the approval prompts behave.
A prompt template that works across agents:
Goal: fill the shipping quote form in my browser using the PDF in ~/Desktop/quote-test.
Rules: use only the files and apps I allowed. Stop and ask before submitting anything.
Output: tell me each value you entered and where it came from.
Keep an eye on how long each run takes compared with doing it yourself. The point of a background window is that your attention is free, not that the agent is faster.
Why this matters beyond Google
Computer use is turning into a standard feature of the assistant app, not a research demo. Google has been moving Gemini in this direction for months. Our Gemini managed agents with self-hosted sandboxes post covers the developer side, and our Gemini 4 Neon launch post covers the model line. The Mac app is the consumer side of the same push, and the Howie scheduling agent acquisition points the same way.
Three points for different readers:
- Builders: Treat the Mac app as a test bed for agent UX. The background window and live preview are patterns your own agents may need. Read our agent skills guide and loop engineering guide for how to structure repeatable tasks.
- Founders: Desktop agents compete for the same workflows. If your product is a thin layer over "click through this app," the assistant vendors now ship that layer for a monthly plan.
- Marketers and operators: Google's deck-from-dashboards and resale-listing examples are the realistic wins. They are repetitive, multi-app and easy to review.
The practical skill in all three cases is writing tasks that an agent can finish and a human can check.
What is still open?
- Which actions need approval. Google says "key actions" and gives no list.
- Reliability. No error rate, task success rate or benchmark was published with the feature.
- Limits and pricing. The plans exist already, but Google does not say whether heavy computer-use sessions count against any cap.
- Data handling. The post says you choose app access. It does not describe what is stored or sent to Google from the live preview. Check Google's privacy terms before using it with client data.
- Expansion. Other countries, languages, Intel Macs and Windows are not mentioned.
When Google answers these, the answers will appear in the Gemini app help pages and on the Google blog post linked above.
Related reading
- Claude can now use your computer in the background
- Meta Muse for Mac desktop launch
- Meta Muse and Signal messages: what Reuters found
- Full Disk Access and AI agents on macOS
- Gemini managed agents with self-hosted sandboxes
- Gemini 4 Neon: launch, price and benchmarks
- Google Labs acquires the Howie scheduling agent
Details reflect Google's announcement of November 19, 2026. Availability and plan terms can change.
