Originally published on The Daily Flare.
Google announced on Thursday, November 19, that Gemini’s Mac app can carry out multi-step tasks across local files, desktop applications and websites in the United States.
The feature lets users describe a task in plain language while Gemini navigates menus, types information and clicks through apps. Google says the work runs in a separate, isolated window, with a live preview shown in the chat so users can follow its progress while keeping their own screen and cursor available.
Users choose which apps Gemini can access and can pause or stop the process. Google says users can also approve actions before they happen, including key actions during a workflow. The controls allow the agent to work across different parts of a task without taking over the user’s main desktop.
Designed for repetitive workflows
Google presented the feature as a way to handle chores that involve repeated clicks or several steps across different tools. One example is extracting weights, dimensions and addresses from local PDF files, then entering those details into forms to request shipping quotes. Another is batch-processing photos, such as applying color grades in design software, or setting up complex scenes in 3D applications.
For presentation work, Gemini can gather metrics and screenshots from live dashboards covering multiple regions and use them to build a deck in Google Slides. That workflow combines information from online sources with presentation software, rather than relying only on material already stored in one document.
Google also cited preparing resale listings in batches. Gemini can examine clothing photos in a local folder, research prices, record item details and stage listings on resale platforms. The company described other possible jobs as organizing files into folders based on specified criteria, filling out forms and repeating the same operation across a batch of files.
These are examples Google provided of tasks the feature can handle, rather than a guarantee that every workflow will work in every app. The company says Gemini can interact with Mac applications and websites, while the user selects the applications it is permitted to access.
Initial availability is limited
The rollout began November 19 in English in the United States through the Gemini app for Mac. Google’s announcement says the feature is for personal Google accounts held by people aged 18 or older, and requires an Apple Silicon Mac running macOS Sequoia 15.0 or later. Users also need a Google AI Plus, Pro or Ultra subscription.
Macworld reported that the update is being distributed gradually. Its editor found that an older version of the app did not have the new computer-use capability, so attempts to automate tasks in Pixelmator Pro, QuickTime Player and Finder failed. That account illustrates that installing the Mac app alone may not mean the feature is already available on a user’s device.
Macworld noted that Mac users have been able to automate tasks with Apple’s Shortcuts and AppleScript, while Gemini offers a plain-language way to describe a workflow. TechCrunch placed Google’s launch in a market where OpenAI, Anthropic and Meta had already offered agentic features in their desktop applications. Google’s addition brings its Gemini app into that group with a tool focused on actions across the Mac, local files and the web.
Top comments (2)
The separate execution window and live preview are probably the most important parts of this announcement, not just Gemini being able to click through apps. Desktop agents need to be observable and interruptible, especially when they can access local files and submit information to external websites.
I'd be particularly cautious with workflows involving resale listings, shipping quotes, or form submissions. Extracting information and preparing a draft is one thing; actually publishing or submitting it is another. Those actions should have explicit approval checkpoints.
The real test will be reliability across different applications and unexpected UI changes. A demo with a predictable workflow is impressive, but handling errors, ambiguous instructions, and partial failures is what makes a desktop agent genuinely useful. Plain-language automation is compelling, provided users retain meaningful control.
Great, won't be using it to shave 200 milliseconds off my daily tasks ;-)