Some work isn't in a web page at all. It is in Preview, Slack, Mail, System Settings, an installer, a file picker, or the desktop build of the app you are making. With Computer use on, your agents work in those apps the way you would: they read the window, find the button by its name, click, type, and tell you what happened.
They do it in the background. The agent works in a window behind the one you are using, an orange pointer shows you where it acts, and your own pointer and keyboard stay yours. You can keep writing an email while an agent fills in a form in another app.
Asked to work out a sum in Calculator and check Bluetooth and the sound output in System Settings. Each step sits above the reply with a picture of the window the agent read, and it changed nothing it wasn't asked to.
Turn it on
Computer use runs on a Mac with codecast connected (Getting started covers connecting a computer).
Open Agent features from your account menu and pick the Mac your agents run on.
Switch on Computer, under Hands on the machine.
The card asks for two macOS permissions. Click Open System Settings and turn on codecast computer under Accessibility, then under Screen Recording.
The card turns green and reads Accessibility and Screen Recording granted.
The Computer detail in Agent features. The switch at the top right turns it on for the selected Mac.
Accessibility lets the agent read a window and click, type and scroll in it. Screen Recording lets it take a picture of the window it is reading. Without Screen Recording it can still work, it just can't show you or itself what the window looks like.
The permissions go to a small codecast helper app, not to your terminal, so granting them doesn't hand the same access to everything else you run there. You grant them once; they survive codecast updates. If an agent needs the permissions before you've granted them, the same steps appear as a card under its step in the conversation, with a Tell the agent to continue button once they're on.
Ask for it in plain words
Name the app and say what you want:
"Sign the lease in Preview with my saved signature. Don't save or send it."
"Pick the export folder in the save dialog."
"Check that the desktop build shows the new settings pane, and send me a screenshot."
"Is Bluetooth on, and which microphone is selected?"
"Open the exported PDF in Preview and tell me how many pages it has."
For anything inside a web page, the agent uses your Chrome instead (Browser). Computer use covers what a web page can't reach: the address bar, a file picker, a permission prompt, or an app that isn't a website.
What you see on your Mac
Your pointer stays in the window you are using. The agent's orange pointer works in the window behind it.
The orange pointer. Whenever the agent clicks or types somewhere, an orange pointer glides there and pulses on the press. It never takes your focus; it is there so you can see what the agent is doing.
Windows stay where they are. The agent doesn't bring apps to the front. Typing and most clicks reach a window in the background.
It asks before taking your screen. A few controls only respond to a real mouse click on a window in front. The agent first looks for another way (a keyboard shortcut, a menu item), and if there is none, it asks you before bringing the window forward.
What you see in codecast
Each step the agent takes lands in the conversation above its reply, with a thumbnail of the window it read. Click a thumbnail to see it full size. The reply tells you what the agent did and what changed.
The agent also knows the difference between it worked and I pressed the button. When it can read the result back (a field now holds the text, a checkbox is now ticked) it says so. When an app gives no way to confirm, it looks at the window again before claiming success, and tells you if the app ignored it.
What it will and won't do
Reading and checking are the agent's to do. Anything that leaves a mark waits until you ask for it.
In an app that holds private content, such as Mail or Messages, the agent reads only what you asked it to read.
Two refusals that hold no matter what the agent is asked: 1Password, Bitwarden, Dashlane, LastPass, NordPass and Proton Pass can't be read or driven at all, and a password or one time code field reaches the agent only as [redacted].
When something is off
What you notice
What to do
The card asks for permissions you already granted
Click Open System Settings, switch codecast computer off and on again in that list, then come back
The agent says it can't take pictures of windows
Turn on Screen Recording for codecast computer. Everything else keeps working meanwhile
The agent asks to bring a window forward
Say yes if you can spare the screen for a moment, or tell it to find another way
The agent can't find a window
Open the app, or the document, and ask again
Computer use is built for macOS. On a Linux machine running a desktop, the same feature drives apps through Linux's accessibility layer.