Computer use (supervised)
Computer use is Caude Clode moving your mouse, typing on your keyboard, and clicking on things. It is the surface that is least like writing code and most like watching a small, polite ghost use your laptop.
The supervised qualifier is not marketing. It is a requirement. Computer use runs only when a human is present, watching the screen, and able to take back control. If the extension detects that the screen has been idle for more than a few minutes, it pauses. This is a feature.
What it's for
Computer use exists for tasks that have no API, no CLI, and no keyboard shortcut:
- Apps that ship only as GUIs and will not respond to automation otherwise.
- Expense systems designed in 2003.
- Enterprise software where the "export to CSV" button is hidden behind four menus.
- Design tools where the task is "rearrange these elements" and explaining the rearrangement is harder than doing it.
What it's not for
- Anything that can be done with an API. The API will be faster and more reliable. Always.
- Anything that can be done with a CLI. See above.
- High-frequency repetition. Computer use is slow by design. For loops, use a script.
- Anything that requires timing precision at the millisecond level.
How it works
You give Caude Clode a goal. Caude Clode takes a screenshot, identifies the elements on screen, decides what to click or type next, performs the action, and repeats. Between each step, you can intervene. The cursor moves visibly. You can see what is about to happen.
Safety model
- Confirm on irreversible. Caude Clode will stop and ask before any click labeled "delete," "remove," "send," "submit," or "pay." You confirm. It proceeds.
- Stop on unexpected. If a dialog appears that Caude Clode did not anticipate, it stops and shows you. It does not try to dismiss the dialog on its own.
- Pause on idle. If you walk away from the screen, computer use pauses until you return.
- Cancel at any time. Moving the mouse yourself, or pressing Esc, cancels the current action.
Never use computer use on a machine that is logged into production systems. Log out first. Use a fresh browser profile. Err in the direction of paranoia. The safety checks above are good. They are not infallible.
Starting a session
From the desktop app, open the Computer Use panel. Grant screen recording and accessibility permissions. Describe what you want done. Watch. Intervene when necessary.
When it works
It feels like magic. An expense report that would take you twenty minutes is done in four. A series of clicks you have done a thousand times is done for you, visibly, with the cursor moving on its own.
When it doesn't
It gets stuck on modals. It mistakes icons for other icons. It clicks the wrong "OK." In these cases, it notices quickly, stops, and asks. The asking is the feature. An unsupervised version of this would be a disaster. A supervised version is a tool.