Gemini computer use on Mac is a supervised, isolated-window feature for eligible US English subscribers; keep money, identity, deletion, and sharing under your own approval.
The 30-second verdict: delegate chores, keep logins and final clicks
Let Gemini do the navigation and transcription, keep every login, payment, send, and delete behind your own approval. It runs in a separate virtual screen with a live preview in chat, pauses for logins and payments, and asks before sensitive actions, while Google's help page warns Gemini "can make mistakes and do unexpected things" (Google Support).
Who can turn it on: eligibility and cost in one check
These five checks cover the main eligibility requirements in Google's linked announcement and help page (Google Blog, Google Support).
| Requirement | What Google states |
|---|---|
| Region and language | United States, English only at launch |
| Account | Personal Google Account, age 18+; work and school accounts excluded |
| Plan | Google AI Plus, Pro, or Ultra (entry from $4.99/month per MacRumors) |
| Hardware and OS | Apple Silicon Mac running macOS Sequoia 15.0 or later |
| App | Gemini app for Mac (gemini.google/mac); support page suggests Hydrogen 4 or newer for better-quality responses |
The rollout began November 19, 2026; Intel Macs, free plans, Workspace accounts, and non-US users are excluded. The linked announcement gives no wider expansion date.
Turn on @Computer use without over-granting
In a Task thread, select Task and add @Computer use; persistent app lists, macOS permissions, and in-chat approvals are separate controls (Google Support).
- Open
Gemini → Settings → Computer use(macOS shortcut⌘ + ,per Google's settings path). - Turn on Allow Gemini to work with your Mac, then grant macOS Accessibility and Screen & System Audio Recording when prompted.
- Optionally turn on Use on-demand file backups before file tasks; treat it as best-effort, not a restore guarantee.
- Name source files, target app, and output format in the Task prompt.
- Safety tip (author recommendation, not a Google rule): at first contact with each desktop app, choose Allow once before ever choosing Always allow.
Use the live preview and Stop task, starting with disposable copies in a non-signed-in folder.
Connectors affect which integrations appear, while per-app Allow decisions still gate native control: Desktop entries include Blender, Autodesk Fusion, Git, and Obsidian (9to5Google).
Where it stops: the cross-app boundaries that matter
Gemini navigates multiple apps, transfers information between them, and handles click, type, scroll, menus, and forms. Google's four launch patterns (Google Blog):
- Batch photo edits and 3D-scene setup in creative apps
- PDF weights and addresses into multi-field shipping quotes
- Regional dashboard metrics plus screenshots into Slides
- Clothing-folder review into staged resale listings
Important boundaries, per Google's help page (Google Support):
- Approvals interrupt autonomy. Pauses for logins and payments; staging a listing is not publishing it.
- Sessions travel with the clicks. Allowed signed-in apps expose that session, including visible MFA content.
- No guaranteed rollback. Backups are best-effort, kept 24 hours or until the next prompt, with some app content excluded and some actions irreversible.
- Check third-party rules. Confirm the target app allows automation before delegating.
"Every one of those menus and forms sits behind your logins. Handing the clicking to an agent means handing it the session too. Who's watching the permissions page?" — @weiwan_ok, X
Permissions, approvals, and the 24-hour backup limit
Chat-level choices apply to one step only; only Always allow persists into Allowed apps. Blocked apps is the inverse persistent list (Google Support).
| Decision | Scope | When to use it |
|---|---|---|
| Don't allow | This step only | Unknown app, unexpected request, any login or payment prompt |
| Allow once | This step only | First run with a needed app |
| Always allow | Persistent until removed | Only for a repeatedly needed, low-risk app |
| Blocked apps | Persistent block | Mail, messages, password managers, finance, anything with irreversible power |
Three documented risks: untrusted on-screen content can steer Gemini into other allowed apps; code editors or developer tools can enable shell execution and broad filesystem changes; backups "are kept for 24 hours" and do not cover every app. Remove temporary permissions after the task and verify backups via View backups in the task chat.
Mac app or Computer Use API? A developer rule
The Mac app is supervised chores on your own machine; the API is a screenshot loop your code executes wherever you host it (Google AI Studio docs).
| Dimension | Gemini Mac app (@Computer use) | Computer Use API |
|---|---|---|
| Who executes clicks | Gemini in an isolated Mac window | Your executor (Playwright, Browserbase) runs the returned function_call |
| Coordinate model | Hidden behind the window | Normalized 1000x1000 x/y, scaled by your code to the viewport |
| Cost shape | Subscription (Plus/Pro/Ultra) | Per-token loop: each screenshot plus history plus action output (pricing) |
| Safety owner | Google prompts plus your Allowed/Blocked lists | You: turn caps, confirmations, sandbox, injection detection, audit logs |
| Best for | Personal cross-app chores you watch | Browser automation, form-fill at scale, UI testing, product research |
Use the API for headless or repeated product workflows; use the Mac app when a person watches the preview and approves the final action. Details are in Gemini Computer Use API GA Guide: Pricing, Safety.
FAQ
Why is the Computer Use option missing from my Gemini Mac app?
Recheck the five rows above, then rollout status and the Task plus @Computer use entry point.
Can Gemini see passwords, sessions, or MFA codes?
Per Google's warning, treat any allowed signed-in app as exposed: as a conservative operating rule, never persist-allow password managers, mail, or finance apps.