Google has given the Gemini app for Mac computer use: since 19 November, Gemini can click, type, scroll and work through menus in desktop apps, local files and websites on a user’s behalf, Google announced. It is now the cheapest way to get an assistant that operates a whole Mac, because the feature is included from the $4.99 Google AI Plus plan. The catch is reach: US English, personal accounts and Apple Silicon only, while Anthropic and OpenAI already ship comparable agents for Windows and company plans.
Who can use it, and how it runs
The rollout covers personal Google accounts of users aged 18 or older with a Google AI Plus, Pro or Ultra plan, on an Apple Silicon Mac with macOS Sequoia 15.0 or newer. Work and school accounts are excluded for now. On Google’s US plan page, those tiers cost $4.99, $19.99 and from $99.99 a month, and for this feature they differ in how many Gemini tasks each plan allows.
- Setup: users switch on “Computer use” in the Gemini settings and grant two macOS permissions, Accessibility and Screen & System Audio Recording.
- Starting a job: in the Task mode, a prompt that mentions @Computer use hands the work to the agent.
- Where it works: Gemini moves the app it controls into a virtual screen and streams a live preview into the chat, so the user’s own cursor and windows stay free.
- Per-app consent: the first time Gemini wants an app, the user can deny it, allow it once or always allow it. A separate block list keeps apps off limits.
- Backups: an optional setting makes Gemini copy files before it edits them. Google keeps those copies for 24 hours or until the next prompt, and says some apps cannot be backed up at all.
Google’s examples are deliberately mundane: color-grading a batch of photos, reading weights and addresses from PDFs into a shipping form, building a slide deck from dashboard numbers, or staging resale listings from a folder of clothing photos.
The risks Google lists on its own help page
The launch post mostly stresses control. The Gemini help page is more candid. We counted seven separate warnings there, and four of them describe real attack or damage paths rather than generic “AI can make mistakes” language:
- Text shown in one app can manipulate Gemini into acting in another allowed app, a cross-app form of prompt injection.
- Gemini works inside logged-in accounts and active sessions of native Mac apps, not only in Chrome.
- Allowing code editors or developer tools can open the door to shell commands and wide file system changes.
- Gemini can read multi-factor authentication codes that are visible in messaging or desktop apps.
The other three: Gemini can pass information to desktop apps, some actions cannot be undone even with backups, and automating certain third-party apps may break their terms of service, which Google makes the user’s responsibility.
Gemini, Claude and ChatGPT side by side
Google is the last of the three big labs to let its assistant drive desktop apps. Comparing the three vendors’ own documentation shows where Gemini is ahead and where it lags:
- Entry price: Gemini from $4.99 a month. Anthropic’s pricing page puts Claude Pro, the cheapest plan with computer use, at $20 a month or $17 with annual billing.
- Platforms and accounts: Gemini runs on the Mac only. According to Anthropic’s help center, Claude’s computer use works on macOS and Windows and is in beta for Team and Enterprise plans once an admin turns it on. OpenAI offers it in the ChatGPT desktop app on macOS and Windows, inside Work and Codex.
- Hard limits: OpenAI states that ChatGPT cannot automate terminal apps or itself, cannot authenticate as an administrator and cannot approve security prompts. Google’s help page names no comparable block. It only warns that developer tools can lead to shell execution and leaves the choice to the user.
- Approach: Claude uses a connector first, then a browser, and clicks through the screen only as a last resort, which Anthropic calls slower and more error-prone. Gemini’s Mac feature starts from the screen.
- Working in the background: all three can now operate Mac apps without taking over the user’s cursor, which was the main practical objection to earlier screen agents.
Computer use itself is not new at Google either. Developers have had it through the API since a Gemini 3.5 Flash preview in June; the Mac app now brings it to consumers five months later. Perplexity has offered a background Mac agent with access to local files since April.
The Apple factor
Timing matters here. Apple’s macOS 27.3 beta reworks the permissions AI agents rely on: Accessibility becomes its own setting, Full Disk Access can be limited to single folders, and a menu bar icon shows when one app is controlling another. Gemini’s computer use depends on exactly those rights, so Mac users will soon see Apple’s stronger warnings at the moment they hand them to Google.
Apple is also working on its own answer. It reportedly wants to turn Siri into a cross-app agent by 2027. Until then the company sets the rules for third-party agents on its platform without offering one itself.
What to set up before the first task
- Block first: put messaging apps, password managers, authenticator apps and terminals on the block list before running anything, since Google itself names MFA codes and shell access as risks.
- No “Always allow” for code: grant code editors and developer tools once per task at most.
- Backups on, but short-lived: switch on file backups, and check results right away, because the copies disappear with the next prompt.
- For companies: Gemini’s version only works with personal accounts, so it sits outside Workspace admin controls. IT teams should check whether staff use it on company Macs with private Google logins.
i10x Perspective
Google is competing on price and reach into households, not on safety engineering. A $4.99 entry point puts a screen agent in front of far more casual users than Claude or ChatGPT at $20, and casual users are exactly the people least likely to maintain a careful block list. OpenAI has chosen to hard-block terminals; Google has chosen to warn. We think Google will have to follow OpenAI’s route once the first public prompt injection case on a consumer Mac appears.
The more important deadline is Apple’s. macOS 27.3 makes every agent ask for narrower rights, and a Siri agent in 2027 would compete with all three on their own turf. Gemini’s window to become the default Mac agent is roughly one year.
Still open
Google has not said when computer use will reach other languages, countries such as Germany, or Workspace accounts. It has also not published success rates for multi-step tasks or the number of tasks each plan allows. The help page only recommends Gemini Hydrogen 4 or newer models for better results, without saying which model drives the clicks by default. How the agent behaves once macOS 27.3 lets users restrict it to specific folders is another open question we will add here once it ships.






