Claude’s New Feature Allows You to Automate Almost Every Process on Your Mac, but It Poses a Huge Security Risk.
Claude is eager to take control of your computer. Anthropic has released a new feature called “Computer Usage,” which can now control your Mac’s keyboard and mouse to perform tasks on your behalf. This feature is finally available in preview for Claude Pro and Max subscribers, after first being teased back in 2024 .
The Computer Usage feature builds on recent additions like Claude Cowork and Dispatch. Claude Cowork is a tool that lets you run tasks on your Mac (albeit in the background, using built-in features and macOS integration), while Dispatch is a new remote access feature that lets you control Claude Cowork from your iPhone or iPad (a limited version of OpenClaw from Claude), as long as your Mac is connected to the internet.
How the new Computer Usage feature works in the Claude app
As the name suggests, when the “Computer Usage” feature is activated, Claude takes over the entire screen. “Computer Usage” works in both Claude Cowork and Claude Code, but I only tested it in Cowork (as I’m not yet a programmer ). It can move the cursor, use keyboard input, move and delete files, and basically do whatever it wants. It can open files, read them, and perform actions based on the file’s contents. This, of course, poses a huge security risk, as it creates the possibility of malicious code injection , where a well-hidden line of code could hijack the AI, compromising your privacy.
Anthropic claims to have developed protection mechanisms against instant malware intrusion , but since its AI models are so quick to react and prone to hallucinations, it’s hard to take Anthropic at its word. Fortunately, Claude always asks for permission before accessing new apps and deleting files.
The good news is that Claude treats computer use as a last resort. When you assign a task in Claude Cowork or Claude Code, the system will first attempt to complete it using its MCP connectors . For example, you ask Claude Cowork to write an email to your boss. The system first uses the Gmail connector to compose a draft of the email. But MCP connectors are limited; they can’t do everything. For example, the Gmail integration can’t actually send the email for you. You’ll have to click a button.
This is where Claude’s browser integration comes in handy. Claude will ask you if it can manage your Chrome tabs. If you have Claude for Chrome installed and enabled, it will automatically create the appropriate tab groups, open the desired website, manage the tabs (an indicator will light up around the tabs managed by Claude), and click the “Submit” button for you. Once complete, it will also be able to close these tabs.
Claude’s browser-based functionality proved to be the most suitable option for me, and it will likely be suitable for other computer-intensive users as well. In another test, I asked Claude to visit Techmeme, generate a summary of the top five news stories, convert it to a Markdown file, and save it locally. This all happened in the background in Chrome while I continued writing in Obsidian. The program did a good job of compiling a summary of all the posts, and I was able to read the results and save the file for later use.
Working on a computer can be painfully slow.
The “Use Computer” feature, as I mentioned before, is the final straw. Let’s say you ask to create a calendar event using Apple’s own Calendar app. There’s no MCP server (though there is for Google Calendar). So instead, Claude will ask you if he can simply take control of the computer.
But first, you’ll need to grant Claude mouse and keyboard access, as well as screen recording access so he can take screenshots as the program runs (this is the only way Claude actually knows what to do). Then he’ll ask you for full access to a specific application. The nice thing is that all other applications and windows are hidden during this process, significantly reducing the chances of Claude getting out of control and ruining, say, your Word documents.
Then Claude will get to work. You’ll see the familiar glowing edges and a small window showing all the steps Claude is attempting.
Given what you’re asking Claude to do, watching it can be either fascinating or excruciating. Unlike Claude’s browser automation feature, you don’t have to do anything. Claude takes up the entire screen. This is the biggest limitation of this feature, and perhaps it will only be fixed if Claude creates a virtual environment to perform local tasks.
When Claude takes control of the computer, he truly takes control. You’re locked in, defenseless. When I asked him to create a new calendar entry for me, it was quite captivating to watch it happen in real time. It only took about 30 seconds. Of course, I could have completed such a simple task faster myself. In fact, Claude added the task to a random calendar, while I have a separate calendar for Formula 1 (any self-respecting Formula 1 fan would do that). When I asked Claude to reschedule an event, he complied, but it took another 30 seconds. Overall, it’s a decent prototype, but I don’t plan on using it in my everyday life.
When I tried to use Claude Cowork to move recent screenshots to another folder, things went awry. Even after granting full access to the Downloads folder and Finder, the program still couldn’t find the path to the Desktop folder to retrieve my recent screenshots. After a minute or two, Finder’s “Go to Folder” feature opened, allowing me to manually enter the path to the Desktop folder (at least, some people use this underrated but useful feature). That’s when my patience ran out, and I abandoned the task.
And here’s my main problem with Claude’s “Use Computer” feature. For tasks I regularly perform on my Mac—like moving files, resizing images, converting documents, adding calendar events, and writing research papers—I still complete them much faster than with Claude, which essentially uses an AI model to make decisions by analyzing screenshot after screenshot. Not only is this resource-intensive, but it’s also incredibly slow. Of course, if I were a corporate employee needing to analyze data from multiple files throughout the day, things might be different. But in that case, I’d probably use Claude inside Excel or a programming app to directly interact with the data, rather than letting Claude “play the role” on my Mac.
How to enable computer usage in Claude
This feature is only available for Claude on macOS and only for paid subscribers of Claude Pro ($20 per month) and Claude Max ($100 per month and up). As mentioned above, it’s in preliminary testing, so it’s still a beta version. That’s why it’s not enabled by default for everyone.
To enable this feature, go to Settings > General > Computer Usage . To enable browser-based automation features, enable “Browser Usage” (this will allow Claude to open and navigate to any website in Chrome without your prior consent). Here, you can also add apps to the blocked apps list to prevent Claude from ever accessing them. Good places to start are apps like 1Password and banking apps. In this section, you can also grant access to accessibility features and screen recording.
The Computer Usage section is for a Mac mini, which I don’t have.
After trying this feature a couple of times, I realized it wasn’t for me. At least not until I got a Mac mini. The browser feature can still be useful, as it can run in the background and perform tasks. But sitting and waiting while Claude tries to figure out where to click in Finder is beyond my patience.
So this feature is largely aimed at people who want to run AI agents on their Mac mini using OpenClaw, running 24/7, processing files and automating actions ( which is a bad idea from a security standpoint ).
You can control Claude’s actions from your phone or, even better, from your other Mac, the one you use for work. It’s better security-wise than OpenClaw, because you can see exactly what Claude is doing and can stop or take control at any time. But, oh my, is it slow.