#038

Claude Code Now Controls Your Entire Computer

Claude Code can open apps, click, and take screenshots from your terminal. Computer use is the last link in the chain toward the autonomous agent.

From zero to professional: Claude Code tips, tutorials and deep dive guides.

Explore the hub

TL;DR Claude Code can open apps, click, type, and take screenshots — directly from your terminal. It compiles a Swift app, launches it, verifies buttons work, and goes back to the code to fix what fails. All in the same conversation.

Claude Code tries the precise tool first

Some tasks a terminal just can't solve: verifying a native app works visually, debugging a layout that only breaks at a certain window size, or interacting with an iOS simulator. Until now, that meant leaving Claude Code and doing it by hand.

Computer use changes that. It's a built-in MCP server called computer-use that gives Claude visual access to your screen. But it's not the first thing Claude tries — it's the last.

Claude only touches the screen when nothing else works

Claude Code has a clear tool hierarchy. It always uses the most precise tool available:

  1. MCP server — if there's an MCP server for the service, it uses that
  2. Bash — if the task is a terminal command, it executes it
  3. Chrome — if it's browser work and you have Chrome set up
  4. Computer use — only when none of the above work

Screen control is the nuclear option: powerful but slow. Claude reserves it for what nothing else can reach: native apps, simulators, tools without an API.

The missing piece in the autonomous agent

Computer use isn't an isolated feature. It's the last link in a chain Anthropic has been building for months — each piece reducing the distance between "I do" and "Claude does for me":

Capability What it solves
Headless mode Claude works without you present
Notifications Alerts you when it's done
Remote Control Control it from your phone
Channels Talk to it from Telegram/Discord
/loop Watch on a recurring basis
Computer use Controls your screen when the terminal isn't enough

The combination of all of them: an autonomous agent that works on your machine, contacts you through your preferred channel, and can now also see and touch your screen.

How to enable it

1. Enable the MCP server

/mcp

Find computer-use in the list. Select it and choose Enable. Setting persists per project.

2. Grant macOS permissions

The first time, macOS will ask for two permissions:

  • Accessibility: so Claude can click, type, and scroll
  • Screen Recording: so Claude can see what's on your screen

3. Ask for something that needs a GUI

Build the MenuBarStats target, launch it, open preferences,
and verify the interval slider updates the label.
Screenshot when you're done.

Claude compiles, launches, interacts with the UI, and reports back.

Reference

Aspect Detail
MCP server computer-use (built-in, disabled by default)
Enable /mcp → select → Enable
Platform macOS (CLI). Windows only in Desktop app
Plans Pro and Max (not Team or Enterprise)
Minimum version v2.1.85+
macOS permissions Accessibility + Screen Recording
App approval Per session — Claude asks permission for each app
Stop Esc anywhere or Ctrl+C in terminal
Hierarchy MCP → Bash → Chrome → Computer use (last resort)

Official docs: Computer use — Let Claude use your computer from the CLI

Requirements

  • macOS (not available on Linux or Windows via CLI)
  • Claude Code v2.1.85+
  • Pro or Max plan (doesn't work with API key, Bedrock, Vertex, or Foundry)
  • Interactive session (doesn't work with -p)

Get only what matters

If I have nothing worth saying, you won't hear from me. When I do, you'll be the first to know. 7,000+ professionals already trust this.

Are you a professional Web developer?
No

Unsubscribe at any time.