Ask Gemini to compare three public product pages while you head out for lunch, and it may be able to finish even after you shut your laptop. Ask it to open a folder of unexported photos and work through them in a desktop editor, and the same closed laptop is now the thing standing between you and a result. Both jobs can start with a sentence in Gemini. They do not run in the same place, and they do not fail for the same reasons.
On November 19, Google introduced computer use in the Gemini app for Mac. With permission, Gemini can click, type, scroll and navigate desktop apps. Google moves the work to a virtual screen with a live preview, rather than using the cursor on your main display. That makes it possible, for example, to keep writing while the agent works through a batch of image exports. It does not make the files, apps or active sessions on your Mac available to a separate cloud browser.
The useful question is not whether the agent can imitate a sequence of clicks. It is where the necessary files, sessions and tools already live, and what proof you will need when the work comes back. Answer those two questions before giving the agent a job, and the difference between the desktop and the cloud stops being a settings-menu detail.
The first is native Mac computer use. This is the new feature in the November announcement and the one to look for when a task depends on a local app. Google’s examples include batch color-grading photos in design software and turning local PDF details into shipping quotes. In Google’s Mac help documentation, the feature requires Accessibility and Screen & System Audio Recording permissions to operate. The first time Gemini needs an app, it asks whether to use it. The virtual screen is a way to keep the agent’s activity out of your way; it is not a reason to assume the agent cannot reach signed-in accounts inside the apps you allow.
The second is local Chrome. Some jobs need websites rather than desktop apps, but they depend on the browsing session you already have. That can make local Chrome convenient if, say, a dashboard is already open under the right account. Google says local browsing requires your device and Chrome to stay awake and running. If the device becomes unavailable, Gemini might switch to its remote browser. Whether the job can finish there depends on what it needed from the original session. A task that began inside an authenticated dashboard may stop when it meets the same site from a different browser.
The third is Gemini’s remote browser. The websites are still real, but the browser runs separately from the browser on your Mac. Google says a remote browser task can continue after you close your device. With a Google AI subscription, Google’s browser settings let you choose Remote browser only. A site can prompt for sign-in, at which point Gemini may ask to use Sign in with Google or saved credentials, or pause for you to take over. The remote browser can retain sign-in data between sessions, but Google removes saved data after inactivity and lets you clear it yourself.
These modes have different availability. For native Mac control, Google’s launch notes specify Apple silicon and macOS Sequoia 15 or later, an adult personal Google account with an eligible Google AI Plus, Pro or Ultra plan, and a rollout in English in the United States. That is narrower than availability for Gemini Tasks in general. If you can open the Gemini website or already have the desktop app, that tells you very little about whether the new Mac control setting will be there.
Consider a weekly check of a small publication’s public pages. The task is to open its five newest stories, verify that each author link goes to the team page, and note any links that reach a missing article. A suitable brief would name the starting page and the five pages in scope, require one result per story and ask for the final destination URL. Nothing needs to be edited. This is a good candidate for the remote browser because it depends on public facts, not on the publishing account or any file saved on the original machine.
Now change the brief: open that same publication’s editor, inspect pages that have not been published, and correct the ones with bad links. The new task has different prerequisites even though the pages look similar. It needs the site's editing account, a way to distinguish a draft from a live page, and an explicit rule about what may be changed. A report that misses one broken link is inconvenient. A fix that publishes somebody else’s unfinished draft has a different cost. For a first run, I would keep the review and the edit separate, even where the same person normally does both.
Local apps change the economics again. Say a product photographer needs twelve 1,600-pixel-wide JPEGs from a folder of finished images. If the desktop editor has a dependable batch export preset, that is likely the simpler answer. Its behavior is predictable, and the output can be checked without debating what the model meant by “done.” An agent makes a stronger case when the input is variable: one source is too small, another is a duplicate, and three require decisions about which rendition to export. Even then, let the agent identify the exceptions and have the preset handle the uniform work.
Splitting the job at the right point is often more useful than insisting that one agent complete it end to end. An agent can shortlist candidates or collect evidence while you are away; the local machine can handle a narrowly specified conversion when it is available. The handoff becomes a list of files or URLs rather than a vague instruction to pick up wherever the last browser stopped. A short handoff like that is easy to inspect and can often be reused with a different tool later.
A poor delegation brief asks Gemini to tidy up files or check a website. Either instruction leaves it to invent what counts as a problem. A better brief gives it one folder or a fixed starting URL, the allowed operations, the deliverable and a stop condition. For a batch export, tell it to copy originals into a work folder, export into a new destination and report which files did not meet the size requirement, without overwriting anything. For the website check, specify that it should report missing content and failed links, not edit pages or follow permission prompts into another account.
Give the finished job an artifact you can check quickly. If the expected output is an image set, look for the named files, the dimensions and the count, then spot-check a few images. If the task is a small web audit, give each row a source URL and an observed destination. Ask the agent to separate “not verified” from “failed.” A sign-in prompt is not evidence that a link is broken, and a screenshot of a page title is not evidence that its author link goes to the right place.
The reporting step matters most when the task gets interrupted. Suppose an agent can verify four of five public pages, but the last redirects to a preview login. It should return four results and one specific blocker rather than continue through an unexpected account prompt. That gives you enough to decide whether to fix the link, verify the redirect yourself or defer it. A completion badge alone cannot tell you whether the work is worth accepting. The value is in a result that survives inspection without asking you to watch the whole run.
A browser that can use an existing login saves steps, but those steps were also a barrier between reading a page and acting on it. The practical default is to use the narrowest access that can finish the task. Public-page comparisons do not need a logged-in browser. A file export does not need email or messaging access. For unfamiliar desktop apps, one-time approval lets you learn what the task actually needs before deciding whether to allow repeated access. Google says it is designed to ask about sensitive actions, but also warns that untrusted content in one allowed app can steer Gemini to act in another. The agent may see verification codes displayed in permitted desktop or messaging apps. Both are reasons not to grant access to apps that the job does not need.
A similar distinction applies to backups. Google describes its on-demand Mac backups as optional and best-effort. It keeps them for up to 24 hours or until your next prompt, and some apps may not be backed up at all. That is useful for recovery, but it is not a substitute for starting with copies when you are automating a destructive or hard-to-reverse operation. It is safer to ask the agent to produce a preview folder you can compare against the originals than to let it modify your only copy and test its restore mechanism afterward.
Finally, remote does not mean hands-off. It means the browser can carry on without that Mac being awake. A login, an interactive check or a confirmation step can still require a person. If the job has a deadline, plan for that pause rather than assuming the clock stops when the agent needs you. A workflow that reliably arrives at a human approval point is useful; a workflow that silently waits at an unknown prompt is often harder to manage than the manual job.
Pick a small, reversible job you already perform, not one you could only justify by having an agent. Use five records or pages where the correct answer is easy to verify, and include a known exception. Run the job once by hand and once with Gemini. Count the minutes you spend preparing the inputs, monitoring the run, checking the output and correcting it. Record elapsed time separately. A remote report that takes a while but needs only two minutes of your attention can still be a win. A quick result that sends you back to inspect every step is not.
Check whether the agent really needs the machine. If a public browser job finishes in the cloud with evidence you can read, do not keep a Mac awake just to provide a window. If the job depends on a mounted drive, a desktop-only app or an existing local session, do not expect a remote browser to recreate that environment on demand. Keep the native job small enough that it can stop safely and resume from an obvious point, and keep a simpler script or preset for any part that can be made predictable.
Gemini’s Mac rollout gives people a new way to reach work that used to be trapped behind menus and local files. The remote browser solves a different problem: continuing web work when your laptop is no longer there to do it. The most useful setup will probably use both selectively. Decide which environment has the inputs, give it only the access the job needs and ask for a result you can verify. Then closing the lid becomes a deliberate choice instead of an experiment.