Claude Code adds macOS UI “computer use” from the CLI

Anthropic has just rolled out “computer use” in Claude Code, letting it open apps, click through macOS UI, and test what it builds from a single prompt. The research preview is available on Mac for Pro and Max users, with early feedback surfacing rate-limit and app-support wrinkles.

mac-os cover

TL;DR

  • Claude Code “computer use” from CLI: Opens apps, clicks macOS UI, and tests builds without app-specific CLI hooks
  • End-to-end loop: Write code, compile, launch, navigate UI, find bugs, fix, and verify in one prompt
  • macOS UI coverage: Works with compiled SwiftUI apps, local Electron builds, and GUI tools lacking a CLI
  • Availability: Research preview on macOS, limited to Pro and Max plans
  • Enablement: Turn on via /mcp; docs: https://t.co/x9kSKedkjz
  • Early feedback: Interest tempered by usage/rate limits; reported app boundary with Brave prompting extension use

Anthropic is adding a new capability to Claude Code: “computer use” directly from the CLI. In practice, that means Claude can open apps, click through macOS UI, and test what it just built—without relying on the target app having its own CLI hooks.

What “computer use” looks like inside Claude Code

Claude frames this as an end-to-end loop that can happen from a single prompt: write code, compile, launch the app, navigate the UI, find a bug, fix it, and verify the fix. Rather than stopping at code generation or basic command execution, the agent is positioned to push further into the kinds of steps that usually require a human at the keyboard—especially when the only “API” available is the UI itself.

For AI-assisted coding workflows, the shift here is less about novelty and more about coverage: the last mile between “the code compiles” and “the feature actually works when clicked.”

macOS UI coverage: not just browser automation

A notable detail is the breadth of targets Claude claims to handle on Mac. The examples called out span:

  • a compiled SwiftUI app
  • a local Electron build
  • a GUI tool that doesn’t have a CLI

That “GUI tool that doesn’t have a CLI” line is doing a lot of work. In many dev environments, internal tools, legacy utilities, and one-off QA repro apps still live entirely behind windows, menus, and dialogs. UI-level interaction makes that surface area accessible to an agent in a way standard terminal tooling can’t.

Availability, plans, and how it’s enabled

Computer use in Claude Code is available in a research preview on macOS for Pro and Max plans. It can be enabled with /mcp, with documentation linked here: https://t.co/x9kSKedkjz.

Early reactions: excitement—plus friction around limits

The replies show a familiar split: interest in a CLI-driven agent that can actually operate UI, alongside frustration about usage limits and rate limiting (including requests for clearer communication on limits). There are also early notes about app support boundaries—for example, one user reports Claude Code refusing to interact with Brave and instead directing them toward a browser extension.

Original source: https://x.com/claudeai/status/2038663014098899416

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community