Computer MCP

Move the mouse pointer without clicking (to trigger hover states).

Open an app's OWN door in the background - a Spotify track, a WhatsApp chat, or the app itself - without activating it (if the app pulls itself to the front anyway, the front is handed back and the reply says so). This is the way to reach an app whose window is covered, or that will not act from behind: it uses the app's own URL, not its window. It carries navigation ONLY, never a send: it opens a chat with no pre-filled text and starts a track - it cannot send a message. Sending is a separate, ordinary action on the open window, subject to the same rules as any other write (this tool does not…

Put text on the clipboard, press Cmd+V, then put your own clipboard back. Use it instead of computer_type for anything long or awkward - a paragraph, a URL, an emoji, text in a script the keyboard layout cannot produce - because it lands in one step rather than character by character. There is deliberately NO tool that READS the clipboard: a person copies a password out of their password manager, and one read would hand it to the model past every other guard. To put your previous clipboard back this does read it, in memory only, for the few milliseconds the paste takes; that value is never ret…

List the actions that were refused because they needed a human and there was no way to ask without taking over the screen. In background mode a dialog is itself an interruption, so a write that would need one is refused rather than raised - and the refusal only reaches the person if they happen to read the right conversation. This is where it becomes visible: what was asked for, when, and why it stopped. It is a list, not a button. Nothing here can be approved from here - a consent granted without a human is exactly what the gate exists to prevent. To allow the work, the person changes mode; t…

Check which macOS permissions are granted (Accessibility, Screen Recording). Call this first if anything fails. It also lists the rules in your MCP client (Claude Code settings) that let an agent click, type or send on this Mac without going through this server - osascript, an open shell, browser automation - because this server's consent never sees those.

Press an element through its own accessibility action instead of simulating a click on a coordinate. Works while the window is BEHIND another one and never moves the human's mouse pointer - so it is the tool to reach for when the agent should not take over the screen. The exception is anything that opens a menu (a pop-up, a menu button, a combo box): pressing it behind would draw the menu over the person's screen and hold their mouse and keyboard, so it is refused until you ask for the screen with computer_request_screen. Two or more matches is a refusal, not a guess: narrow the search, or pas…

Ask an application to quit, the same way Cmd+Q does - so it still gets to ask you about unsaved work. It never kills a process. Quitting can lose work, so it asks for consent every time, in every mode.

Record the screen to a video file, for a person to watch afterwards - a tutorial, a demo, proof of what was done. action "start" asks the person every time (it is the screen, for minutes; in read-only mode it is refused), then records one display until "stop" or maxSeconds. Only password managers are left out of the recording, including ones opened while it runs; everything else is recorded as it looks - password fields are NOT blacked out the way they are in screenshots. The answer is the file path: this server never reads the file back and never sends it to you. "status" says whether a recor…

Ask the person to lend you the screen for a while - for the few steps that cannot be done from behind: moving the pointer, dragging, bringing a window forward, switching desktop. Before calling this, tell the person in the chat exactly what to do: press Allow (Touch ID) in the Computer MCP box in the top-right corner of the screen they are using (the same question is also under the orange menu bar icon), then confirm with Touch ID - this call then waits (several minutes) for them to do that, so do not call it silently. They approve it with Touch ID; you get it for at most 15 minutes, only you,…

Screenshot ONE display, or one app. Password fields and password-manager windows are blacked out BEFORE the image is written, so they never reach the model. On a machine with several displays this captures one of them; the answer says which, how many there are, and where that screen sits on the desktop. A window you cannot find is usually on another display - call computer_displays and pass displayId. Set redact=false only if you know the screen holds no secrets.

Scroll by pixel deltas. With app, the scroll lands in that app's window at the point x, y (in points, e.g. the centre computer_find gives you) - also when the window is behind another one, and without moving the pointer; leave x and y out to scroll the middle of the app's front window. Without app it scrolls at the current pointer position. A scroll is not confirmed by the app - check the effect with computer_find or a screenshot.

Write text straight into a text field without moving the pointer or bringing anything forward. On a pop-up button it chooses the item with that name instead: the menu opens on the person's screen for a moment and the item is pressed - only after 2 s without input from the person (otherwise person-active with idle_seconds). Nothing is chosen for a missing item (no-such-item, with the items there are), two items with the same name (ambiguous), a greyed-out item (disabled), a menu that did not show (menu-not-seen), or a slow app (app-busy). If the person starts using the keyboard or mouse while t…

Switch to the desktop (Space) to the left or right, the way Control+Arrow does. This is how you reach a full-screen app: macOS gives every full-screen window its own Space, so a window you cannot find is often simply on another one. It moves what the person is looking at, so it asks for consent every time, in every mode - the same rule as quitting an app. If the system shortcut for switching Spaces is turned off on this machine it says so and sends nothing, rather than pressing a key that does nothing. And it tells you whether the desktop actually changed, measured by which windows are on scre…

Type text into the focused field as Unicode, so it works on any keyboard layout. The text is never written to the audit log in clear text.

Wait until an element appears, instead of taking screenshots in a loop. Give it the same search as computer_find plus a timeout in seconds; it returns as soon as something matches, or says plainly that nothing appeared. Twenty polls cost one call here and twenty images the other way - and an image is the most expensive thing you can put in a context window.

Move, resize, close or minimise a window. Coordinates are global points, the same space computer_click uses, so a negative x is a screen to the left - this is how you put a window on another display. Closing a window can lose unsaved work, so close asks for consent every time, in every mode.

List open windows with titles and screen frames.