A background workspace for computer actions

According to Google, Gemini can click, type and navigate menus in an isolated background window, with a live preview in chat. Users choose accessible apps, can pause execution and approve key actions. The launch initially covers English in the United States for adult personal accounts on Google AI Plus, Pro or Ultra.

The distinction matters: this announcement concerns a product capability, not downloadable model weights or a new LoRA training recipe. It does not establish that arbitrary desktop workflows will complete correctly.

The desktop app and this feature have different availability

Google’s desktop page lists Apple Silicon and macOS Sequoia 15.0 or later for Mac. It describes shortcuts for opening mini or full chat, screen-context sharing, dictation and Tasks for organizing local files. It also advertises Windows downloads, while warning that native capabilities and availability vary.

Consequently, a download button or supported country for the general desktop app is not confirmation that this particular computer-use rollout is enabled. For readers in Ukraine, the US-only launch scope is the relevant limit. This report does not claim access outside it or an offline execution mode.

App access and data controls deserve separate attention

The Gemini Privacy Hub says information shared with Gemini can include files, screen context and browser page content. It also lists generated task steps, tool calls and interaction information among collected data. Its settings section distinguishes managing history from managing Connected Apps access.

The same notice explains that disabling generative-model training on chat content does not stop processing needed to provide responses and protect the service, including some human review. These are the service’s stated policies, not an independent audit of this new Mac feature. A separate workspace should not be read as a promise that all processing remains on the device.

What a useful practical trial would measure

Our assessment: evaluate the finished artifact, not just the sequence of successful clicks. A shipping-form trial could compare every entered value with a small set of test PDFs, record omitted fields and stop before sending a real order. A slide-building trial could check that each number and screenshot belongs to the intended region and reporting period.

Measure elapsed time including corrections and approvals, alongside the manual baseline. Keep a record of interrupted tasks and incorrect outputs rather than excluding them from the result. These are proposed evaluation criteria; we have not run hands-on tests and Google’s announcement provides no comparable end-to-end success-rate benchmark.