Background computer use on the Mac: give every AI agent its own cursor
What background computer use is, how Claude, Codex and other tools handle your screen, and how any MCP agent can operate Mac apps in background windows with its own cursor, keyboard and clipboard.
Short answer: Background computer use means an AI agent operates apps on your Mac in windows behind yours, with a cursor, keyboard and clipboard of its own. Foreground computer use borrows your pointer and keyboard and brings its windows to the front, so you wait. Claude and Codex now offer a background mode in their own desktop apps. ShadowCursor gives every MCP agent, in any client, its own cursor, keyboard and clipboard, and several agents can work at once.
What computer use means
Computer use is the ability of an AI model to operate a computer the way a person does: it looks at a screenshot, decides where to click, moves the pointer, types and reads the result. It is how an agent reaches everything that has no API: a web form behind a login, a spreadsheet in a desktop app, a dialog in a native Mac app, a PDF that needs a signature.
The model only decides what to do. The part that actually moves a pointer and presses keys is a separate tool on your Mac. That tool decides whether the agent works with you or instead of you.
Why computer use often makes you stop
The simplest way to build computer use is to drive the one pointer and the one keyboard focus your Mac has. Many tools still do. That has three consequences:
- The agent moves your cursor. Every click it makes is a jump of the pointer you are holding.
- Your keys land in its window. If you type while it works, your text goes wherever the agent has focused, and its text can end up in your document.
- Its windows come to the front. To click something, the tool raises the window or hides everything else, which covers what you were doing.
So the practical rule with these tools is: start the task, take your hands off the keyboard and wait. And it is one agent at a time, because two agents cannot share one pointer.
How the built-in options handle your screen
Claude and Codex both ship computer use of their own, and on the Mac both have added a background mode. What they do differs by app and plan (as documented in October 2026):
| Tool | While it works | Agents at once | Where it runs |
|---|---|---|---|
Claude Code CLI (computer-use server) | Hides your other apps until the turn ends | One session holds a lock | macOS, Pro or Max plan |
| Claude Desktop and Cowork | Background mode on macOS, otherwise hides your windows | Not documented | macOS and Windows, Pro or Max plan |
| Codex in the ChatGPT desktop app | Background on macOS, foreground on Windows | Not documented | ChatGPT desktop app |
| ShadowCursor | Background, with its own cursor, keyboard and clipboard | Several, each with its own cursor | Any MCP client on macOS 15+ |
If you only ever use one of these apps and one agent at a time, the built-in mode may be all you need. Background computer use becomes a separate tool when you want it in the client you already work in (the Claude Code or Codex CLI, Cursor, VS Code), on a plan the built-in feature does not cover (Claude Team, Enterprise, an API key or Bedrock), or for several agents at once. Sources: Claude Code docs, Claude Desktop docs, ChatGPT computer use docs.
What background computer use needs
For an agent to work next to you rather than instead of you, the computer use tool has to give it a full set of its own:
| Part | Foreground computer use | Background computer use |
|---|---|---|
| Cursor | Your pointer | Its own cursor, drawn in its color with its name |
| Keyboard | Your keyboard focus | Keys that arrive only in the agent's window |
| Clipboard | The system clipboard | A clipboard per agent |
| Windows | Brought to the front | Stay where they are, even when covered |
| Agents at once | One | Several, each with its own cursor |
The hard part is the windows. macOS stops drawing windows that are completely covered, and browsers pause pages that are not visible. A background tool has to keep those windows alive and capture screenshots of exactly the window the agent works in, not of your screen.
How ShadowCursor does it
ShadowCursor is a Mac app and MCP server built for exactly this. When Claude Code, Codex, Cursor or another MCP client connects, the agent gets:
- Its own cursor. It glides to its target on screen, in the agent's color and with a name tag, so you always see who is doing what.
- Its own keyboard. What the agent types goes to one app and one window. Nothing else receives it.
- Its own clipboard. Copy and paste work for the agent without touching your clipboard or another agent's.
- Its own windows. The agent works in its windows in the background, even when your windows cover them, and sees only those windows.
Input moves like a person: the pointer travels in curves and typing has a human rhythm at about 200 words per minute. That matters on the web, where hover menus, sliders, date pickers and signature fields expect real pointer movement.
Setup in three steps
- Download and open ShadowCursor. Allow Accessibility and Screen Recording when your Mac asks. Those two permissions are what any computer use tool on a Mac needs.
- Connect your AI. ShadowCursor finds Claude, Codex, Cursor and VS Code on your Mac and connects them with one click. There is no configuration file to edit.
- Ask for the task. For example: "Add the new invoices from my Downloads folder to the spreadsheet." The agent opens the files and the sheet in its own windows and works while you keep going.
What agents can do in the background
Anything with a window. Some typical tasks:
- Spreadsheets: fill a sheet from invoices, emails or a web page, in a window you are not even looking at.
- Websites: search, read, compare and fill forms in Safari, in a window of the agent's own.
- Mac apps: create events in Calendar, write in Notes, work through menus and file dialogs.
- Documents: sign a contract in Preview with a real pen stroke.
- Several projects at once: Claude on your invoices and Codex testing your web shop, side by side, each with its own cursor.
Limits to know
Background computer use is not magic, and some things depend on how an app is built:
- Minimized and hidden windows cannot be seen or operated. Covered windows can.
- Dragging files, tabs or text between windows is a system drag that follows the real mouse, so ShadowCursor cancels it instead of dropping. Sliders, canvases and signatures work.
- Hover without a click reaches pages in Safari but not in Chrome or Electron apps.
- An agent cannot use a background tab in your browser window, because browsers only draw and accept input for the selected tab. It opens its own window instead.
- If an agent works in the same app you are typing in, it waits for your pauses.
FAQ
Is background computer use the same as browser automation?
No. Browser automation tools such as Playwright drive a browser through its developer interface and only reach web pages. Computer use operates the apps themselves, through the same buttons, fields and menus you use, so it also works in native Mac apps, file dialogs and desktop software.
Which AI agents can use ShadowCursor?
Claude Code, Codex, Cursor, VS Code and any other client that speaks MCP (Model Context Protocol). Several agents can work at the same time, each with its own cursor.
Does my screen leave my Mac?
ShadowCursor runs locally. An agent only receives screenshots of the windows it works in, and only the agent you connected sees them.
What does it cost?
ShadowCursor starts with a 7-day free trial without a card. After that it costs $12.99 per month or $69.99 per year, with unlimited agents on up to three devices.
Which Macs are supported?
Macs with macOS 15 Sequoia or later. A version for Windows 10 and 11 is in beta.