Artificial intelligence November 20, 2026

Google adds computer use to Gemini for macOS, with live previews

Google added computer use to Gemini for macOS on November 19, 2026. Give it a task and it can work across native Mac apps, local files, and websites in the background. A live preview in chat shows what it’s doing while you keep using your screen and ...

Google adds computer use to Gemini for macOS, with live previews

Gemini can now work across Mac apps while you keep using your desktop

Google added computer use to Gemini for macOS on November 19, 2026. Give it a task and it can work across native Mac apps, local files, and websites in the background. A live preview in chat shows what it’s doing while you keep using your screen and cursor. You can pause or stop a run, and Google says you can approve actions before they happen.

At launch, the feature is available in English in the U.S. It requires the Gemini Mac app and a Google AI Plus, Pro, or Ultra subscription. Google hasn’t said when it will expand to other markets or platforms.

Google is catching up. OpenAI, Anthropic, and Meta already offer agent features in their desktop apps, TechCrunch reported on November 19, 2026. The practical draw for Mac users is that they can hand off a job that spans several apps and get on with something else. How much time that saves depends on how often they have to check Gemini’s work.

The work between apps

Google suggests using Gemini to build a slideshow from local files or an online dashboard. That’s a better test than filling out a single form. Take a routine project update: open last week’s deck, pull the latest numbers from a dashboard, pick the right chart, and drop it into a new slide. It’s easy enough by hand, and building and maintaining an integration may cost more than the clicks it saves.

A desktop agent lets you try handing off that work without building connectors for every app. Start with jobs that are awkward to script, easy to inspect, and cheap to undo. A draft deck qualifies. So does sorting copies of files by a rule you can check afterward. Google also suggests repetitive work across batches of files, where a successful run could save a lot of clicking.

If the job runs every morning, the calculation changes. With a supported dashboard API, a scheduled job can request a specific metric and date range, validate the response, and log what it used. That takes engineering time up front, but the log gives you somewhere to look when a number is wrong. A slide produced by an agent can look fine and still show last week’s figures. Price in the time it takes to catch that.

The Mac release comes alongside Google’s wider rollout of Gemini Tasks, reported November 17, 2026, which helps users sign in to sites, make purchases, and book appointments. Google also announced its household organization agent CC on September 18, 2026 and has opened it to early testers. Neither product tells you which transactions Gemini for Mac supports or when it will ask for approval. The Mac release adds the ability to work across desktop apps and the web.

Risks and limitations

The live preview helps you spot a run going wrong, provided you’re still watching it. If you’ve moved on to another job, approval prompts matter most before actions such as submitting a form or changing a batch of files. TechCrunch describes pause, stop, and approval controls but doesn’t say which actions require approval or what happens to changes already made when you stop a run. Don’t count on the stop button to undo anything.

The reporting also leaves out how Gemini interacts with macOS apps, what permissions it needs, and which app or webpage data stays on the Mac versus going to a remote service. The fact that a run happens in the background doesn’t tell you where the processing happens. On a managed Mac with access to internal dashboards or customer documents, settle those questions in an access and data-handling review before rolling it out.

Text in a webpage or document can also look like an instruction to the agent. Gemini needs to treat that text as data and stick to the actions the user authorized. The reporting doesn’t explain how it enforces that boundary. That uncertainty is enough to keep sensitive files and consequential actions out of early trials. It doesn’t establish a flaw in the release.

There’s little basis for estimating throughput. The report offers examples but gives no completion rates, timings, or details on how the agent handles partial failure. If a file job stops halfway through, you need to know which files changed and whether it’s safe to run the job again. You can check a small batch by hand. For a larger one, a script that logs each change may be easier to operate, even if it takes longer to write.

Start with a job you can check

Pick a task someone already does by hand and can verify in a minute or two. Ask Gemini to sort a folder of copied files into a separate destination using a clear rule, or have it build a draft deck from a dashboard for a known reporting period. Check what’s missing as well as what’s wrong. A missing chart is easier to overlook than one with obviously bad numbers.

Count the time you spend giving instructions, approving actions, checking the output, and fixing errors. Compare it with how long the job normally takes. Keeping control of the desktop matters because you can use the time the agent spends running. If checking the result takes as long as doing the job, keep the manual workflow or write the script. Start with work that’s much faster to check than to produce.

Keep going from here

Useful next reads and implementation paths

If this topic connects to a real workflow, these links give you the service path, a proof point, and related articles worth reading next.

Relevant service
AI agents development

Design agentic workflows with tools, guardrails, approvals, and rollout controls.

Related proof
AI support triage automation

How AI-assisted routing cut manual support triage time by 47%.

Related article
OpenAI acquires Sky to bring AI actions directly into macOS

OpenAI has acquired Software Applications, the startup behind Sky, an unreleased AI interface for macOS that can sit above the desktop, read what’s on screen, and take actions across apps. That pushes OpenAI past the chat window and into the OS. If C...

Related article
Perplexity brings its Personal Computer agent to all Mac users

Perplexity has made Personal Computer available to all Mac users through its desktop app. The pitch is straightforward: give an AI agent access to local files, native Mac apps, web tools, and a large set of connectors so it can handle multi-step ...

Related article
Simular launches a macOS agent that can operate your computer directly

Simular has released a 1.0 macOS agent that can operate a computer directly, and says Windows support is coming through Microsoft’s Windows 365 for Agents program. It also raised a $21.5 million Series A led by Felicis, with NVentures and South Park ...