Gemini computer use comes to Mac — paid plans only

Share
Gemini computer use comes to Mac — paid plans only

Google's Mac assistant can now take the mouse, and it landed the same week Apple began showing developers the security settings built to police software that does exactly this.

Google started rolling out computer use in the Gemini app for macOS on Thursday, letting the assistant work through multi-step jobs across local files, desktop apps and websites while the user keeps typing. It is the biggest change to the Mac desktop-agent market this month, and it comes with a condition list long enough to decide whether you can use it at all: an Apple Silicon Mac, macOS Sequoia 15.0 or later, English only, United States only, a personal Google account, a user aged 18 or over — and one of the three paid Google AI plans.

The paid-plan requirement is the part worth pausing on. Computer use is not a free-tier feature, and Google puts it behind AI Plus, Pro and Ultra alike, so the cheapest entry point is AI Plus rather than the top tier. That is a different commercial posture from a free consumer rollout: the company is charging for agency, not for chat.

What it actually does

The pitch is repetition, not intelligence. Google's own examples are chores: batch colour-grading photos in a design app, pulling weights, dimensions and addresses out of local PDFs to fill in multi-box shipping quotes, assembling a slide deck from metrics and screenshots spread across regional dashboards, and staging a folder of clothing photos as resale listings complete with researched prices.

Mechanically, Gemini works inside an isolated background window with a live preview inside the chat, so your screen and cursor stay free — you can keep typing while it clicks through menus behind you. You choose which apps it may touch, you can pause or stop it mid-run, and you approve key actions before they execute. That combination — scoped apps, visible steps, an approval gate — is now the standard safety story for desktop agents, and Google is following it rather than inventing it.

The fine print matters more than the demo. Personal accounts only means no Workspace tenants at launch, which rules out the audience that would get the most value from batch form-filling and dashboard-to-deck assembly. English and the United States only means this is a beachhead release, not a product. And "Apple Silicon, Sequoia or later" quietly excludes a large installed base of older Intel Macs that still run the Gemini app.

Google is late, and the framing shows it

Google's competitors did not wait. OpenAI, Anthropic and Meta have been shipping agentic features in their desktop apps for some time, and TechCrunch is blunt that Google is a little late to the game. The same report notes the adjacent moves that make Thursday's launch look like catching up rather than leading: Google only this week widened access to Gemini's Tasks, the in-app feature that can sign into websites, make purchases and book appointments, and its revamped "CC" household agent, announced in September, is still admitting early testers.

Being late is not fatal here, because the market is still unformed. Nobody has shipped a desktop agent that ordinary people trust with a credit card and a password manager, and every vendor is currently optimizing for a demo that survives contact with a real desktop. Google's advantage is distribution — a Gemini app already installed on Macs, tied to accounts people already have — and its disadvantage is that it is arriving after the first wave of trust was spent on rivals.

The Apple collision

The timing is not a coincidence, or if it is, it is an awkward one. Apple is preparing new restrictions for apps that request "full disk access," explicitly citing the risks of AI agents that can operate a computer, and a preview of the new settings shipped with this week's macOS 27.3 beta, The Verge reports. The 9to5Mac write-up of that beta is the clearest public look at what is coming: an operating system that will start asking harder questions of any process that can read the whole disk.

That puts Google in the position of shipping a desktop agent into a house whose locks are being changed. Gemini's per-app permission model may map cleanly onto Apple's new controls, or it may not; Google's announcement does not address the question, and Apple's rules are still in beta. What is clear is that the Mac is about to become the most regulated surface for computer-use agents, which is either a moat for a company that ships its own OS-integrated agent or a tax on everyone else. Google is currently everyone else.

The trade-off nobody has priced

An agent with permission to click through your apps is a new attack surface, and the oldest attack on it is prompt injection: text planted in a document, a web page or an email that tells the model to do something you did not ask for. Per-app scoping and approval gates reduce the blast radius, but they do not remove it, and the chores Google is selling — reading local PDFs, scraping dashboards, listing items for sale — are exactly the tasks that require an agent to read untrusted content.

Google's announcement does not say how long action logs are kept, whether screenshots from the isolated window are stored, or what happens when an agent is pointed at an app that holds credentials. Those are not gotchas; they are the questions a buyer should be able to answer before letting software type on their behalf. Until Google publishes them, the honest summary is that the safety design is visible at the level of controls and silent at the level of records.

There is a reasonable counter-argument: an agent that shows every step in a live preview and waits for approval is more auditable than a person clicking blindly, and the alternative — running an unscoped script — is worse. True. But the comparison that matters is not agent versus human, it is agent versus the vendor's own documentation, and right now the documentation stops where the interesting part starts.

What to watch

Three signals will tell you whether this is a real product or a demo with a subscription attached. First, whether Apple's macOS 27.3 rules land in a form that treats third-party computer-use agents as first-class or as a special case. Second, whether Google extends computer use to Workspace accounts and to Windows, which is where the paying enterprise demand sits. Third, whether the approval flow survives real use — users who approve every step will turn it off, and users who stop reading approvals have built themselves a very expensive liability.

For now, the honest read is that Google has closed a gap it should have closed a year ago, on one platform, in one country, for paying customers. That is a product launch. Whether it is a product is a question for the next three months.

Would you hand a desktop agent your apps for the price of AI Plus? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Amazon's World Product Graph: every product, by October 2027

Amazon's World Product Graph: every product, by October 2027

Amazon wants to be the answer to "what should I buy," not just the place where you buy it. An internal document reviewed by Reuters describes a catalogue of nearly every product sold anywhere on earth — including the ones Amazon doesn't stock, scraped from the retailers that do. Amazon is building a project it calls the World Product Graph: a plan to hold data on nearly every product sold anywhere, partly by crawling rival retailers' websites.

ChatGPT Wallet launches: agents pay, OpenAI takes no cut

ChatGPT Wallet launches: agents pay, OpenAI takes no cut

OpenAI's second run at shopping inside ChatGPT is not a checkout button. It is a card wallet where the buyer is software, the human still signs every purchase, and — for now — OpenAI takes none of the money. OpenAI launched ChatGPT Wallet on Friday for US ChatGPT Pro subscribers aged 18 and over: add a card once, then let a "dots" agent find, fill and pay for an order, with every purchase approved by Face ID or fingerprint, per OpenAI's announcement.

DoorDash opens a waitlist to order food by text with an AI agent

DoorDash opens a waitlist to order food by text with an AI agent

DoorDash is betting the next app you order from won't be an app. On Tuesday it opened a US waitlist for ordering through an Apple Messages thread, plus a developer waitlist for a connector that lets AI assistants place corporate orders — the first time a major delivery platform has said out loud that agents, not screens, are where it expects to be found. DoorDash opened a US waitlist for an AI agent that takes orders inside Apple Messages — 20,000 users are already in, and it is iOS-only for…

Microsoft quietly kills the 'Copilot Plus PC' brand

Microsoft quietly kills the 'Copilot Plus PC' brand

Microsoft's own new laptops have stopped using the name it spent two years selling — and a new paper says the AI jobs crunch still hasn't shown up in the graduate data. Two stories for this evening. Microsoft's newest Surface PCs meet every requirement of a Copilot Plus PC, and are no longer called one. "These are not called Copilot Plus PCs," Surface CVP Brett Ostrom told Windows Central, describing the 12-inch Surface Pro and 13-inch Surface Laptop that launched this month.

Microsoft folds coding and always-on agents into one Copilot app

Microsoft folds coding and always-on agents into one Copilot app

Microsoft finally shipped the consolidated Copilot it promised in June — and the surprise is that an app sold to knowledge workers now writes and hosts internal apps. Two other stories worth your time: ICLR's second anonymity incident in two cycles, and Berlin's AI street cameras, which are not judging anything yet. Microsoft unveiled the Copilot "super app" it has been teasing since June: one app with three tabs — Home, Code and Autopilot — that merges Copilot Chat with Cowork and rebrands…

ByteDance's agent phone goes on sale Sep 16 — and apps get a veto

ByteDance's agent phone goes on sale Sep 16 — and apps get a veto

ByteDance stopped demonstrating its phone agent and started shipping it. The consumer version of Doubao Phone Assistant is now official, the first handset carrying it goes on sale this week, and the part that matters beyond one device is a consent protocol that lets third-party apps refuse the agent outright. ByteDance released the consumer version of Doubao Phone Assistant, with the first device shipping it — ZTE's Nubia NaviX Ultra — open for pre-order and on sale September 16.