Computer use — the ability for AI to see your screen, click buttons, open apps, and navigate software on your behalf — has been technically possible for a while. Anthropic released it as an API-only beta back in late 2024. But "technically possible" and "something a normal person would actually use" are two very different things. That beta required developer skills to set up, ran rough around the edges, and stayed firmly in the territory of people who were comfortable writing code.
What Anthropic released at the end of March is different. Computer use is now built into the Claude desktop app through Cowork, and any Pro or Max subscriber can turn it on. No API. No code. No setup. You type a request into a chat window, and Claude goes to work — opening apps, clicking through menus, filling out forms, navigating browsers, creating files.
For anyone who has been following AI capabilities from a distance, waiting for the moment when these tools actually felt usable without a technical background, this is that moment.
For listeners who haven't used it, Cowork is a tab inside the Claude desktop app. Regular Claude chat answers your questions. Cowork actually does things. You can point it at a folder, give it a task, and it will execute multi-step workflows, create real files, and connect to services like Gmail and Slack.
Computer use is an extension of that. With it turned on, Claude can go beyond the apps and services it has direct integrations with and control anything on your screen. An old Windows application. A browser-based tool with no API. A legacy system that only responds to mouse clicks and keyboard inputs. If a human can use it by looking at the screen and clicking around, Claude can now attempt the same thing.
There's also a feature called Dispatch that connects Claude on your phone to Claude on your desktop. You can text Claude a task from your phone — something like exporting a pitch deck as a PDF and attaching it to a meeting invite — and come back to finished work on your computer. Combined with computer use, Claude starts to function as something close to a remote digital worker operating your machine while you're away.
The mechanics are simpler than you might expect. Claude takes a screenshot of your screen, looks at what's there, decides what action to take next — clicking a button, typing into a field, selecting from a dropdown — and then takes another screenshot to see what happened. It repeats that cycle until the task is done.
It's essentially doing what your eyes and hands do, just through a slower, more deliberate process. Each action can take five, seven, even ten seconds. You'll watch it make mistakes — clicking the wrong thing, struggling with a dropdown menu that doesn't behave the way it expected — and then reason through the problem and try again. It's a little like watching a new employee navigate unfamiliar software for the first time, which is both endearing and occasionally frustrating.
Anthropic has built a smart hierarchy into the system. Claude tries direct integrations first (things like Gmail or Slack connectors), then the browser, and only falls back to full screen control when nothing else works. It also includes prompt injection detection — protection against malicious websites that embed invisible instructions designed to hijack the AI's behavior. Anthropic has been candid that these guardrails aren't bulletproof and recommends caution with sensitive data.
One feature worth highlighting: skills. After Claude navigates an unfamiliar application, you can ask it to save notes on how it learned to use that app. The next time you ask it to do something in the same application, it draws on those notes and moves significantly faster. Think of it like writing down directions after you've driven somewhere new — the second trip goes a lot smoother.
The legacy software angle is where this gets particularly relevant.
A lot of associations are running systems that someone built 10 or 15 years ago. Maybe it's a Windows-based event registration tool. Maybe it's a membership database that only works through a desktop client. It functions fine, nobody wants to replace it right now, but getting data in and out is painful and slow because there's no API, no web interface, and no way to connect it to modern tools.
Computer use changes that equation. Claude can learn to navigate those older interfaces, and the skills feature means it gets better with each session. The first time might be clunky and slow — a lot of clicking around to figure out where things are. The second time, with saved context about the application's layout and workflow, it moves much faster.
This effectively creates a bridge between AI and older software that previously required a literal human touch to operate. For associations that have been stuck with manual data entry or clunky workarounds because their systems predate modern integrations, that's a meaningful unlock.
Even for everyday workflows — creating meeting links, sending calendar invites, moving data between systems, scheduling emails — Cowork with computer use can handle multi-step tasks that would normally eat up 15 or 20 minutes of someone's day. Multiply that across a team and a full work week, and the time savings start to compound.
If you try computer use today, your first reaction will probably be: this is cool, but it's slow. That's fair. Watching Claude take seven seconds between each click while you could have done the whole task in 90 seconds is a strange experience.
But keep two things in mind. First, AI processing speed has been on a consistent improvement curve. The models are getting faster with every release cycle, and the lag between actions will shrink noticeably in the months ahead. Second, for simpler tasks, you can switch to a smaller, faster model — something like Haiku instead of Opus — and see a real improvement in speed without sacrificing much capability. Figuring out how to navigate a website doesn't require the most powerful reasoning model available.
This is a research preview, not the finished product. The experience six months from now will look very different from what you see today.
Here's where it's worth being direct about risk.
Do not be logged into a password manager while Claude has screen control. If you use a browser plugin for 1Password or a similar tool and it's unlocked, Claude could click on it and access every stored password. From there, it could use those credentials to log into any site — including ones you'd never want an AI anywhere near.
Along the same lines, do not allow multi-factor authentication apps to be accessible in the same session. If you use a Mac and have your iPhone apps mirrored to your desktop, that could include Google Authenticator or Microsoft Authenticator. Passwords plus your second factor means access to banking, email, internal systems — everything.
Anthropic has built permission prompts and blocks around sensitive categories, but the company itself recommends caution. Take that recommendation seriously.
The best approach is to treat computer use the way you'd onboard a new team member. You wouldn't hand someone the keys to every system on their first day. You'd give them limited access, let them prove themselves on low-stakes tasks, and expand their permissions over time.
Start with a sandbox. If you have an old computer sitting around, refresh the operating system and use that. If your AMS or other core systems have a test environment, connect Claude there first — not production. Give it access to a limited set of tools and keep it away from the open web in the same session.
Begin with low-stakes tasks. Create a meeting link. Send a calendar invite. Move a file from one folder to another. Get comfortable with how Claude navigates, where it stumbles, and how the skills feature works before you hand it anything consequential.
Separate your concerns. If you're letting Claude access your AMS test environment, don't also give it access to your browser with all your saved logins. Partition what you share until your confidence in the system grows.
The moment you start using Cowork with computer use, two things will happen almost simultaneously. First, your mind will start racing with everything it could do. If it can create a Zoom link and add it to a calendar invite and email it to someone, what else could it handle? Could it process event registrations? Pull reports from your AMS? Update your website?
Second, you'll feel a pull of hesitation. Should I give it access to my Outlook? What about my file system? What if it makes a mistake somewhere that matters?
That tension is healthy. It means you're taking the capability seriously without letting excitement override good judgment. The people who will get the most out of this are the ones who sit with that tension, start small, build confidence through experience, and expand from there.
The technology to automate a significant portion of repetitive daily work is available right now, in a form that doesn't require any technical skills to use. That's new. What you do with it is up to you — but the first step is trying it.