Claude-style computer use on your Mac in 2026 (step by step)

A computer-use agent looks at the pixels on your screen, decides what to click, and then clicks it. That is the whole trick. Where a chatbot writes you a paragraph about how to rename 200 files, a computer-use agent opens Finder, selects the folder, runs the rename sheet, and watches the count tick up. The work happens in your real apps with your real account, not in a research demo of a fake browser.

On a Mac, that means three things have to be true. The agent needs Screen Recording permission so it can see what is in front of you. It needs Accessibility permission so macOS will let it move the cursor and send keystrokes. And it needs a model good enough to read a noisy Retina screenshot, find the small button it wants, and not click the wrong thing. Whisply ships all three in one app, with the model already wired up, so you do not pay a separate API bill or stitch SDKs together.

The result is the kind of help that used to require a junior assistant. You describe the outcome, the agent does the clicking, you keep working in another window. When something looks off, you take the cursor back. There is nothing to deploy and nothing to babysit.

Most computer-use demos online run inside a virtual Ubuntu desktop in a browser tab. They are great for screenshots and useless for the thing you actually want to do, which is move money in your banking app, reschedule a calendar invite in your real Calendar, or clean up your Downloads folder. The moment you try to bridge a cloud sandbox back to your laptop, you end up wiring tunnels, copying cookies, and explaining to the agent that no, the Chrome inside its container does not have your logged-in tabs.

The other common path is a Python script that drives the official Anthropic computer-use sample on your machine. It works, briefly, until macOS asks for Accessibility permission for a Terminal-launched process, until the screenshot resolution does not match what the model expects, until the script crashes mid-task and leaves a half-renamed folder behind. You spend an evening on plumbing and you still have no UI to pause the agent when it heads somewhere you did not intend.

Whisply is the version that treats macOS as the target, not as a footnote. The agent is a real Mac app with a menu bar icon, a hotkey, a pause button, and a transcript you can scroll. Permissions are requested the way macOS expects. Screenshots come from the system, not from a screen-capture hack that breaks on the next OS update.

Start with the chores that are too small to script and too annoying to do by hand. Sort the screenshots on your desktop into a folder by month. Open the last three PDFs in Downloads, pull the totals into a Numbers sheet, and save it to a specific iCloud folder. Reply to every unread message in a Slack channel with a one-line acknowledgment and a thumbs up. The agent will narrate what it is doing in the overlay, so you can watch it work the first few times and then stop watching.

Move up to multi-app workflows next. Take the invoice PDF that just landed in Mail, drop the line items into a Google Sheet in Chrome, then go back to Mail and reply with a confirmation. Open a research paper in Preview, summarize it into Notes, and start a draft in Mail to your advisor with the summary pasted in. These are exactly the tasks that fall through the cracks between single-purpose tools, and they are where a computer-use agent earns its keep.

The third tier is the one people did not expect to like. Hand it the boring stretch in the middle of a project. Renaming exported assets to a style guide. Filling a form a hundred times with rows from a CSV. Walking through a settings checklist on a new install. The agent does not get bored. You get the hour back.

A good computer-use agent has a brake pedal and an off switch. In Whisply, moving the trackpad or pressing Esc stops the run instantly. The overlay shows you what the agent is about to do before it does it on slower steps, so you can intervene before it clicks the wrong Send button. Nothing about the design assumes you want to walk away and trust it blindly on the first try.

Scope matters too. Tell the agent which apps it should touch and which files it should not. Hand it a specific folder rather than your whole home directory. For anything that costs money or sends an email to a real human, ask it to draft and pause before committing. Treat it like a sharp intern on day one, then loosen the leash as you watch it get the same job right three times in a row.

Because Whisply runs locally on your Mac, the agent is not piping every screenshot to a third-party log server. The model sees what it needs to see in the moment and does not keep a permanent archive of your screen. That matters when the task involves a banking app, a medical record, or a private channel where the rest of the room never agreed to be part of an AI training run.

A chat model is a thinker. A computer-use agent is a doer. The trick is knowing which one to call. If you need a draft, an explanation, or a plan, the chat panel in Whisply gives you that without touching the cursor. If you need the plan executed across three apps, you switch on Computer Use and let the agent take the wheel. Same hotkey, same app, different mode.

The two modes feed each other in practice. You ask the chat side to write the spreadsheet formula you want, then ask the agent to paste it into the right cell in Numbers and drag it down the column. You ask the chat side to summarize a meeting, then ask the agent to file the summary into Notion under the correct project. The split keeps each part honest. Thinking stays cheap and fast. Doing happens in your real apps, with your eyes on it.