Anthropic adds computer use to Claude Code for GUI testing and app control on macOS

What Anthropic posted on X

On March 30, 2026, Anthropic said computer use is now available in Claude Code. The X post frames the feature as a way for Claude to open applications, click through interfaces, and test what it built without leaving the terminal. In practical terms, Anthropic is extending Claude Code beyond file edits and shell commands into direct interaction with graphical interfaces.

That matters because many real development tasks still break out of the terminal. Native macOS apps, Electron builds, simulators, internal tools, and proprietary desktop software often have no clean CLI or API. Until now, a coding agent could write code and run commands, but the last mile of validating what appears on screen still required a human operator or a dedicated test harness.

What the Claude Code docs add

The official docs say computer use runs as a built-in MCP server called computer-use. It is a research preview on macOS, requires a Pro or Max plan, and needs Claude Code v2.1.85 or later in an interactive session. Anthropic also notes that the feature is not available in non-interactive mode with the -p flag.

The documentation positions computer use as the broadest but slowest tool in Claude Code’s toolbox. Claude is supposed to prefer a more precise interface first, such as Bash, an MCP integration, or the Chrome extension. When those options do not apply, computer use can take over the screen to handle tasks like validating a native build, reproducing a layout bug, driving the iOS Simulator, or operating GUI-only software.

Anthropic’s examples show the intended workflow clearly. Claude can compile a Swift app, launch it, click through controls, capture screenshots, and verify the result in the same conversation where it wrote the code. That makes the feature more than remote desktop automation; it is a bridge between code generation, UI inspection, and bug fixing inside one agent loop.

Why this matters

The bigger signal is that coding agents are moving from code assistants toward end-to-end software operators. If an agent can both edit source files and validate the rendered interface, it can close more of the feedback loop on its own. That is especially relevant for UI regressions, native apps, and integration flows that are difficult to cover with static tests alone.

There are still clear limits. Anthropic labels the feature as a research preview, restricts it to macOS and interactive sessions, and reserves it for cases where more precise tools are unavailable. Even so, bringing computer use directly into Claude Code changes the practical scope of what a terminal-based coding agent can do.

Sources: Claude X post · Claude Code docs