Google on Nov. 19 launched a “computer use” feature for its Gemini app on Mac, letting the AI agent operate programs, files and websites in a separate background window while users keep working on something else, the company said in a blog post.
Google said the feature can batch color-correct photos and 3D scenes, pull data from PDFs into shipping forms, assemble presentations from data spread across several dashboards, and list items in bulk on resale marketplaces.
The tool needs an Apple Silicon Mac and at least version 15 of macOS Sequoia, and only works for Google account holders 18 and older who subscribe to one of Google's paid AI tiers, Plus, Pro or Ultra. For now it is limited to English-language use inside the United States; the Gemini Mac app can be downloaded at gemini.google/mac.
Google said users stay in control: they pick which programs the agent is allowed to touch, follow its actions step by step as they happen, halt a task mid-way whenever they want, and sign off on sensitive moves in advance.
Google's own documentation for the feature warns that text hidden inside documents could push the agent into taking actions a user never asked for, a technique known as a prompt-injection attack. The agent can reach active app sessions and sign-in codes and may carry out steps that cannot later be undone.
A backup copy of a file can be turned on for up to a day or until the agent starts its next task, but that option does not exist for some apps.
Similar computer-control tools already exist from OpenAI's Codex, Anthropic's Claude and Moonshot AI's Kimi Work. Apple, meanwhile, is tightening its own permission system in the macOS 27.3 beta, where users must confirm sensitive access all over again.


