Omniscio

OmniscioGuidesClaude Code computer use

Claude Code computer use: Claude drives your Mac

Computer use is the broadest way Claude can act on a machine: not a shell command and not a browser tab, but the actual apps on your desktop. It is how Claude compiles a macOS menu bar app, launches it, and clicks every control to check that it works.

This page covers what it can do, the three steps that turn it on, the per-app approval prompt, what happens on screen while it works, and the parts of your machine it deliberately cannot reach.

Last checked against Anthropic's Claude Code documentation on November 18, 2026. By the Omniscio team. Updated .

What it can do

Computer use is for the tasks that need a GUI: anything you would normally leave the terminal to do by hand.

  • Build and validate a native app. Ask for a macOS menu bar app and Claude writes the Swift, compiles it, launches it, and clicks through every control before you open it.
  • Run end-to-end UI tests. Point it at a local Electron app and ask it to test the onboarding flow; Claude clicks through signup and screenshots each step, with no Playwright config.
  • Debug visual and layout bugs. "The modal clips on small windows" is a task it can resize a window to reproduce, screenshot, patch and re-check.
  • Drive tools with no API. Design tools, hardware control panels, the iOS Simulator and proprietary apps all become reachable.

Source: "What you can do with computer use" in Anthropic's computer use page, read November 18, 2026.

When Claude reaches for it

Claude has several ways to touch an app or a service, and computer use is the broadest and the slowest, so it tries the most precise tool first:

  1. An MCP server for that service, if one is connected.
  2. Bash, for anything a shell command can do.
  3. Claude in Chrome, for browser work when it is set up.
  4. Computer use, when none of those reaches the thing you asked about.

That order is why a request can finish without a single click: screen control is reserved for native apps, simulators, and tools that have no API.

Source: "When computer use applies" in Anthropic's computer use page, read November 18, 2026.

Turn it on in three steps

Computer use is available as a built-in MCP server — one of the tool connections Claude Code can load — called computer-use. It is off until you enable it.

  1. Open the MCP menu. In an interactive session, run /mcp and find computer-use in the server list. It shows as disabled.
  2. Enable it. Select computer-use and choose Enable. The setting persists per project, so you do it once for each project that needs it.
  3. Grant two macOS permissions. The first time Claude tries to use your computer you get a prompt asking for Accessibility (click, type and scroll) and Screen Recording (see the screen). The prompt links to the right System Settings pane; grant both and choose Try again. macOS may require a restart of Claude Code after Screen Recording.

Then ask for something that needs the screen: build a target, launch it, click through each tab, and screenshot any error state.

Source: "Enable computer use" in Anthropic's computer use page, read November 18, 2026.

Approving apps, one session at a time

Enabling the server does not hand Claude every app on the machine. The first time it needs a specific app in a session, a prompt appears in your terminal listing the apps it wants to control, any extra permission such as clipboard access, and how many other apps will be hidden while it works. You choose Allow for this session or Deny, and the approval lasts until the session ends.

Three app categories carry an extra warning, because approving them grants more than it looks like:

  • Equivalent to shell access: Terminal, iTerm, VS Code, Warp and other terminals and IDEs.
  • Can read or write any file: Finder.
  • Can change system settings: System Settings.

None of them are blocked — the warning exists so you can judge whether the task is worth that access. Control also varies by category: browsers and trading platforms are view-only, terminals and IDEs are click-only, and everything else gets full control.

Source: "Approve apps per session" in Anthropic's computer use page, read November 18, 2026.

What happens on your screen

Four behaviours explain most of what you will see while it runs.

  • One session at a time. A session takes the lock at its first computer use action and releases it when the session exits, not when the task finishes. A second session fails with an error naming the holder; exit that session first.
  • Other apps are hidden. Visible apps are hidden while Claude works, so it interacts only with the approved ones. Your terminal stays visible but is excluded from screenshots. Hidden apps are restored when the turn ends.
  • Screenshots are downscaled. Every capture is scaled down before it reaches the model — 3456x2234 becomes roughly 1372x887 on a 16-inch Retina MacBook Pro, keeping the aspect ratio. There is no setting for the target size; if text is too small for Claude to read, enlarge it in the app.
  • Two notifications, and one key. A macOS notification says Claude is using your computer the first time it acts each turn, and another says it is done. Press Esc anywhere to abort the current action, or Ctrl+C in the terminal; either way Claude stops, unhides your apps and gives control back.

Source: "How Claude works on your screen" in Anthropic's computer use page, read November 18, 2026.

The trust boundary

This is the part that makes computer use different from the sandboxed Bash tool. Computer use runs on your actual desktop, with access to the apps you approve, rather than inside an isolated filesystem and network. Anthropic's documentation says so plainly, and points at a separate safety guide for best practice.

The guardrails that ship without configuration:

  • Per-app approval: Claude controls only apps you approved in the current session.
  • Sentinel warnings: apps that grant shell, filesystem or system-settings access are flagged before you approve.
  • Terminal excluded from screenshots: on-screen prompts in your session cannot feed back to the model.
  • A global escape: Esc aborts from anywhere, and the key press is consumed so injected content cannot use it to dismiss a dialog.
  • A lock file: only one session controls the machine at a time.

Claude also checks each action and flags potential prompt injection from what is on screen. That is a mitigation, not a guarantee: approve the app you actually need, and stop the session when it is doing something you did not ask for.

Source: "Safety and the trust boundary" in Anthropic's computer use page, read November 18, 2026.

When it will not turn on

Three failures cover almost every report.

  • "Computer use is in use by another Claude session." Another session holds the lock, and it keeps it until it exits. Exit that session; if it crashed, the lock is released automatically once Claude sees the process is gone.
  • The macOS permission prompt keeps reappearing. Quit Claude Code completely and start a new session, because macOS sometimes needs the requesting process restarted after Screen Recording is granted. If it still persists, confirm your terminal app is listed and enabled under Privacy & Security, Screen Recording.
  • computer-use is not in /mcp at all. The server only appears on an eligible setup: macOS rather than Linux or Windows, a Pro, Max or Ultra plan, sign-in through claude.ai rather than Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry, and an interactive session rather than -p.

If you are on Windows, the CLI is not the road: the same engine runs in the Claude Desktop app, which supports macOS and Windows, with a settings toggle instead of a server list.

Sources: "Troubleshooting" and "Differences from the Desktop app" in Anthropic's computer use page, read November 18, 2026.

Where these answers come from

Common questions

Does computer use work on Windows or Linux?

Not from the CLI, which is macOS only. On Windows the same engine is available in the Claude Desktop app, which runs on macOS and Windows.

Why is computer-use missing from the /mcp list?

Eligibility. It needs macOS, a Pro, Max or Ultra plan, a claude.ai sign-in rather than a third-party provider, and an interactive session instead of -p mode.

Can two sessions use the computer at once?

No. The lock is taken at the first computer use action and held until that session exits, so exit the other session first.

Does Claude see my terminal?

No. The terminal stays visible to you but is excluded from screenshots, so Claude never reads its own output or your on-screen prompts.