Claude Can Work Your Mac While You Walk Away — Here’s What That Changes

I downloaded Claude’s desktop app months ago, mostly to avoid switching between browser tabs. It sat in my Dock doing roughly the same job as the web version, just with a nicer window and the ability to drag files into conversations. Then Anthropic flipped a switch on March 23, 2026, and the app became something genuinely different: an AI assistant that can open your applications, click through your browser, and finish tasks on your desktop while you are standing in a coffee line.

The Claude desktop app for macOS now includes three features that turn it from a chatbot window into an actual work partner: Cowork, computer use, and Dispatch. Together, they let Claude operate your Mac — navigating screens, filling out forms, compiling reports — with your permission and under your watch. The catch is that this capability is a research preview, limited to Claude Pro ($20 per month) and Claude Max ($100 per month) subscribers, and it comes with real constraints that Anthropic is upfront about.

So here’s what each feature does, how to set them up, and where the edges still show.

What Cowork, Computer Use, and Dispatch Actually Do

Cowork is the workspace inside the Claude desktop app where you hand off multi-step tasks. Think of it as the difference between asking Claude a question and asking Claude to go do something. In Cowork, you describe what you need — “pull last quarter’s sales figures from the Google Sheet, format them for the Monday presentation, and drop the file in my Downloads folder” — and Claude works through the steps.

Computer use is the engine underneath. When Claude needs to interact with something that doesn’t have a direct integration, it falls back to literally operating your screen. It moves the cursor, clicks buttons, types text, and reads what’s on the display. Anthropic describes a priority chain: Claude tries a direct connector first (Gmail, Google Drive, Slack), then uses the Chrome browser, and only resorts to screen interaction as a last resort. That ordering matters because screen interaction is the slowest and most error-prone method.

Dispatch ties your iPhone into the loop. You can assign Claude a task from your phone while commuting, and it executes on your Mac back home. One continuous conversation thread bridges both devices. The mental model is closer to texting a capable coworker than issuing commands to software.

How to Turn Everything On

The setup is quick, though Anthropic buries the toggle in a place you might not think to look.

Open the Claude desktop app and click Settings in the sidebar. Under the General section for the Desktop app, you will find a toggle labeled Computer use. Switch it on. That single toggle unlocks Claude’s ability to interact with your screen, your local files, and your installed applications.

For Dispatch, tap the Dispatch option in the left sidebar of the desktop app. You will see toggles to give Claude file access and to keep your computer awake during tasks. Both need to be on. Then open the Claude mobile app on your iPhone, and the Dispatch conversation will sync between devices.

Your Mac must stay awake and the Claude desktop app must remain open for any of this to work. If your MacBook lid closes or the display sleeps, Claude stops mid-task. I’d recommend turning off the automatic sleep timer in System Settings, then navigating to Displays and adjusting the idle sleep interval, or simply plugging in your charger and keeping the lid open while Claude works. It does, though, mean you cannot close your laptop and leave for the afternoon expecting tasks to finish — the machine genuinely needs to be running.

Where Claude Reaches and Where It Stops

The priority chain Claude follows is worth understanding because it shapes what you should and should not ask it to do.

Connectors are fast and reliable. If you’ve linked Gmail, Google Drive, Slack, or another supported service, Claude pulls data through an API — the same way any app integration works. This takes seconds and rarely fails.

Browser navigation is slower but functional. Claude opens Chrome tabs, fills forms, and reads page content. Anyone who has watched a screen recording of this in action will notice the slight delay between each click. It works, but it is not instant.

Screen interaction is where the friction lives. Claude takes screenshots of your display to understand what it sees, then sends mouse and keyboard events to interact with applications. This loop introduces real latency. A task that takes you thirty seconds to click through might take Claude two or three minutes. Complex multi-step workflows sometimes require a second attempt. Anthropic openly states that this method is “still early compared to Claude’s ability to code or interact with text.”

There are hard limits, too. Claude will not handle stock trading, fund transfers, or investment platforms. It avoids sensitive health and financial data by design. You can also create a custom blocklist of applications Claude should never touch — a smart precaution if you keep password managers or banking apps on your Mac. In addition to the obvious value of these restrictions, I also really like that Anthropic gives you a per-app permission system rather than a blanket all-or-nothing switch.

Worth noting: Claude takes screenshots to navigate. That means anything visible on your screen during a task — personal messages, browser tabs, open documents — is potentially seen by the model. Anthropic says sensitive data like passwords and financial details are excluded from Cowork’s memory, but the screenshots themselves are processed during the session. Close anything private before handing your screen over.

The Scenarios Where This Clicks

The pitch makes more sense when you think about specific moments rather than abstract productivity.

You are heading into a meeting and need a summary of three long PDF reports dropped into a shared folder. You open Claude on your iPhone, describe what you need, and walk into the conference room. By the time you sit down, a formatted summary is waiting on your desktop.

You need to fill out a repetitive internal form that requires pulling data from one spreadsheet and entering it into a browser-based tool your company uses. Claude can navigate both — the spreadsheet via a connector, the browser tool by interacting with the page directly.

You want to reorganize a project folder, rename files to match a convention, and compile a status document from scattered notes. That kind of tedious file-management work is exactly where handing your screen to Claude saves real time.

In the worst case of something going wrong mid-task, Claude stops and tells you what happened. It does not silently push through errors, which is a design choice that trades speed for trust.

The Honest Gaps Right Now

This is a research preview, and it feels like one in specific places.

There is only one Dispatch conversation thread. You cannot run parallel tasks — say, researching a topic while simultaneously organizing files. Everything queues into a single stream. For an AI that can control your computer, the inability to manage two things at once feels like an artificial ceiling that will almost certainly be raised later.

Complex chained instructions can be fragile. Asking Claude to “download that attachment, open it in Preview, crop the image, and paste it into the Keynote slide” might work on the first try or might need correction at step three. Screen-based interaction simply does not have the reliability of direct API calls yet.

And the macOS-only restriction locks out a significant chunk of users. If you work across a Mac and a Windows machine, Claude’s computer use features only cover half your setup.

One more practical detail: computer use is not available on Claude’s Team or Enterprise plans. This is strictly a Pro and Max subscriber feature during the preview, which means workplace rollouts are not on the table yet.

How Claude Compares to What Was Already There

Mac users are not strangers to AI desktop tools. Apple Intelligence handles text rewrites and notification summaries natively in macOS Tahoe. Spotlight’s integration with Siri provides basic query answering. Third-party options like Raycast AI or the open-source OpenClaw add varying levels of AI capability to the Mac experience. If you’re exploring what these AI assistants can do for your daily workflow, our breakdown of Apple Intelligence Writing Tools on Mac covers the built-in side of the equation.

What separates Claude’s computer use from all of these is scope. Apple Intelligence operates within Apple’s own apps and system-level text fields. Raycast AI runs commands and queries. Claude navigates across arbitrary applications on your screen, fills in forms in non-Apple tools, and bridges your phone and your desktop into a single task queue. No other Mac AI tool does all three.

The trade-off is subscription cost and the current research-preview roughness. Apple Intelligence is free with your Mac. Claude Pro runs $20 per month, and the Max tier that gives you heavier usage limits is $100 per month. For a more comprehensive look at the Mac productivity landscape, our guide to macOS Tahoe terminal commands explores what the operating system can do without any third-party tool at all.

Setting Up Safely

Before you hand Claude your screen, take five minutes to tighten the environment.

Close browser tabs you do not want Claude to see. Quit applications with sensitive data — banking, health portals, password managers. Enable the app blocklist in Claude’s settings to prevent access to specific apps permanently. Keep your Mac plugged in so a low-battery sleep does not interrupt a task halfway through. And review what Claude accomplished after each session rather than assuming everything went perfectly.

Anthropic’s own documentation warns that safeguards are part of how Claude is trained and instructed, but they are not absolute. Treating computer use as a capable but imperfect assistant — rather than an autonomous agent you can forget about — is the right mental model for where this technology sits in March 2026.