Your agent uses the apps on your Mac while you keep typing
Native apps through the accessibility tree, on windows behind yours, with every action read back and verified.
Computer-use agents usually work from screenshots and pixel coordinates, and they need the screen to themselves: they move your mouse, steal focus and type into whatever window is in front. Codecast works from the other end. The agent reads one native window as a list of named elements, acts on an element by name, and reports exactly what changed, all in a window behind the one you are using. Preview, Slack, System Settings or a save dialog are in reach without the agent taking your screen. Inside a web page the agent uses your browser instead.
Like the browser, this is something you ask for rather than a control you operate: "sign the lease in Preview with my saved signature, don't send it". The first time an agent needs it on a Mac, a Let the agent use your Mac's apps card appears in the thread with Open System Settings for each of the two grants it needs.
41 button Sign.What you see on your Mac
Mostly, nothing changes. No window comes forward on its own. Presses, field values and menu actions reach the app through macOS accessibility in the background, and typing goes into the target app's own queue, never the app you are typing in.
What you do see is an orange agent pointer that glides to each action and pulses on the press, drawn over the screen without taking focus. Your black pointer stays where you left it.
A few controls answer only a real mouse (a drag, say), which macOS drops on background windows. Only then does the agent bring the window forward, and your pointer returns to where it was.
A window, read as text
The agent gets a window back as lines: a number, a role, a name and the extra actions an element offers (41 button Sign, Secondary Actions: show menu), or just the matches for "Sign". A number from before a scroll or navigation fails cleanly instead of clicking whatever sits there now.
Verified versus attempted
Delivered is not the same as taken, so every action reports a verdict and the lines of the window that changed. A value the agent set can be read back, so it reads completed. macOS cannot confirm that a press did what a button promises, so a click reads attempted, and the diff under it is the evidence: in Preview, pressing Sign added a popover with a saved signature and a Create Signature button, along with the numbers to use next. An action the app ignored says No change, and the agent knows not to assume it worked.
These rows land in the conversation like any other step, with screenshots, so you can review what it did without having watched.
Built for a machine with your life on it
Some limits are enforced where they cannot be talked around. Password managers (1Password, Bitwarden, Dashlane, LastPass, NordPass, Proton Pass) are refused outright, under any name, and no setting turns that off. Password, passcode and one-time-code fields always read as [redacted], so their values never reach the agent or the transcript. And every agent carries the same standing rule: reading is the agent's to do; sending, submitting, buying, deleting or changing settings waits for you to ask.
Setup is two macOS grants, Accessibility and Screen Recording, given once to a small signed helper app rather than to your terminal. Settings → Desktop in the desktop app shows both beside the microphone and camera and opens the pane for whichever is missing. Linux under X11 works the same way; Windows does not yet.



