A new capability is emerging for Claude, extending its reach beyond chat and into the very core of your computer. This isn't just about responding to requests; it's about Claude actively *using* your Mac, taking control of the screen and interacting with applications as a human would.
This “Computer Use” feature builds upon recent advancements like Claude Cowork, which handles tasks on your Mac, and Dispatch, allowing control from your iPhone or iPad. Imagine delegating complex actions – not just asking for a summary, but having Claude *perform* the task for you, navigating your system with mouse clicks and keyboard inputs.
When activated, Claude assumes complete control of your screen. It can open, read, and modify files, move things around, and essentially do anything a user can. While incredibly powerful, this introduces a significant security concern: the potential for prompt injection, where malicious code could hijack Claude’s actions and compromise your data. Anthropic claims to have implemented safeguards, but the rapidly evolving nature of AI makes absolute security difficult to guarantee.
Fortunately, Claude prioritizes caution. Before accessing new applications or deleting files, it will always request your explicit permission. It first attempts tasks using built-in connectors, like drafting an email in Gmail. However, these connectors have limitations – they can’t, for example, actually *send* the email.
This is where Claude’s browser integration shines. If a task requires web access, Claude will ask to take over your Chrome browser, creating organized tab groups and automating actions like clicking “Send.” It’s a seamless experience, handling the tedious parts while you continue working elsewhere.
One compelling test involved asking Claude to summarize the top five stories from Techmeme, convert them into a Markdown file, and save it locally. While I continued writing in another application, Claude quietly performed the entire process in Chrome, delivering a well-formatted summary ready for future reference.
However, the full “Computer Use” feature can be… deliberate. Asking Claude to create a calendar appointment using the Apple Calendar app, for instance, requires granting extensive access – mouse and keyboard control, screen recording for visual understanding. All other windows are hidden during this process, minimizing the risk of unintended consequences.
Watching Claude work is a unique experience. A glowing outline indicates which elements it’s controlling, and a small window displays its progress step-by-step. While exciting to witness, the speed is a major limitation. Unlike browser automation, there’s no parallel processing; you’re simply observing Claude’s methodical, screenshot-by-screenshot analysis.
A simple calendar appointment took roughly 30 seconds, a task I could complete much faster myself. And, frustratingly, Claude scheduled it in the wrong calendar. Correcting the error required another 30 seconds of deliberate processing. It’s a promising proof of concept, but not yet a daily driver.
The real struggle emerged when I asked Claude to move recent screenshots to a different folder. Despite granting full access to the necessary folders, it stumbled, unable to navigate the Finder to locate the files. After a minute of searching, it resorted to manually entering the folder path – a feature many users likely forget exists. At that point, my patience expired.
This highlights the core issue: for tasks I routinely perform, I remain significantly faster than Claude. It’s essentially using AI to make decisions based on visual analysis, a process that is both resource-intensive and agonizingly slow. This feature may prove more valuable for tasks like analyzing large datasets, potentially within applications like Excel or coding environments.
Currently, “Computer Use” is exclusive to Claude Pro ($20/month) and Claude Max ($100+/month) subscribers on macOS, and it’s still in a Research Preview phase. To enable it, navigate to Settings > General > Computer Use. You can also manage browser automation and create a “Denied Apps” list for added security, protecting sensitive applications like password managers and banking apps.
Ultimately, my initial experience suggests this feature isn’t for me – at least, not yet. The browser automation is useful, but watching Claude grapple with Finder is simply too slow. It feels designed for those running AI agents on Mac minis, constantly processing files and automating actions. It’s a fascinating glimpse into the future, but for now, I’ll stick to controlling my own computer.
The ability to monitor Claude’s actions from another device – your phone or another Mac – offers a degree of security not found in similar systems like OpenClaw. You can intervene at any point, but the slow pace remains a significant hurdle.