Computer use
Agents that click, type, and read a screen the way you do.
So far your agent has lived in files and terminals. Computer use is the leap to agents that operate a screen the way you do: moving a cursor, clicking buttons, reading what's on the page, filling in forms. Both Claude and Codex have this now, in two flavors.
The browser. The Claude for Chrome extension (and Codex's browser tools) let an agent drive a real web page. This is the reliable, useful-today flavor, and for a builder the standout use is testing your own site like a first-time visitor:
Open my deployed site, sign up as a brand-new user, add two items, and log out. Tell me everything that felt broken, slow, or confusing along the way.The whole desktop. The more experimental flavor hands the agent your actual screen and mouse: opening apps, dragging windows, working across programs. It's genuinely rougher and slower than browser control today, and worth approaching as a preview of where things are heading rather than a daily driver.
Two honest limits. Computer use is slower than an API call, because the agent is literally looking and clicking, and it can still misfire, so you keep confirmations on for anything destructive. But the ceiling is high: anything a person can do in a browser, an agent can now attempt. Filling directory submissions, pulling competitor pricing into a table, reconciling two dashboards that never had an integration.
An API is the front door a service chose to open. Computer use is every other door, since if a human can reach it on screen, so can the agent. That's a lot of doors.
AI Clubhouse · Vibe Coding Guide