Anthropic Brings Computer Use to the Claude Code CLI, Letting AI Control Your Mac From the Terminal

Anthropic added computer use to the Claude Code CLI on March 23, 2026, enabling the coding agent to open apps, click buttons, type, and capture screenshots on macOS directly from a terminal session, no separate GUI required.

The feature, released alongside computer use support in Claude's Cowork product, makes Claude Code the first major CLI coding agent to integrate screen-level computer control natively. It is available to Pro and Max subscribers running Claude Code v2.1.85 or later on macOS.

How It Works

Computer use in Claude Code is implemented as a built-in MCP server called computer-use. It ships disabled by default and is activated per project through the /mcp command inside an interactive Claude session. Once enabled and macOS permissions are granted (Accessibility and Screen Recording), Claude can control any app you approve without workflow changes.

The tool follows a precision hierarchy: Claude tries a configured MCP server first, then Bash commands, then browser automation via Claude in Chrome, and only falls back to full computer control when none of those apply. Screen interaction is reserved for tasks that can't be reached any other way, such as native app UIs, hardware control panels, and tools with no CLI or API.

Access is per session and per app. Each time Claude needs to control a specific app, a terminal prompt shows which apps will be affected, what extra permissions are requested, and how many other apps will be hidden while Claude works. Approvals expire when the session ends.

What Claude Can Do on Your Screen

The practical use cases span anything that previously required leaving the terminal to handle manually. Anthropic's documentation highlights four primary workflows:

• Native app validation: Claude can compile a Swift or macOS app, launch it, click through every control, and screenshot the results, all in the same conversation where it wrote the code.
• End-to-end UI testing: Point Claude at a local Electron or iOS Simulator app with a plain-text instruction. Claude opens the app, runs through the specified flow, and reports what it found, no Playwright configuration or test harness required.
• Visual bug reproduction: Describe a layout issue (a modal clipping on narrow windows, for example) and Claude resizes the window, captures the broken state, patches the relevant CSS, and verifies the fix. Claude sees what you see.
• GUI-only tool automation: Interact with design tools, hardware panels, the iOS Simulator, or any proprietary application that has no CLI or API surface.

Built-In Safety Controls

Anthropic added several guardrails to the implementation. Apps that grant broad access (terminals, IDEs, Finder, System Settings) trigger a sentinel warning before approval, so developers know exactly what they're granting. The Esc key aborts computer use immediately from anywhere on the machine, and the key press is consumed so that on-screen prompt injection can't use it to dismiss dialogs.

A machine-wide lock prevents two Claude Code sessions from controlling the screen simultaneously. Claude's own terminal window is excluded from its screenshots, which blocks a class of prompt injection where malicious on-screen text could feed instructions back into the model's context.

Browser and trading platform access is view-only. Terminal and IDE access is click-only. All other apps get full control. These tiers apply regardless of which apps a user approves.

Reactions and Context

The release follows months of growing demand from Claude Code users running into GUI-only bottlenecks. Testing native app builds, interacting with iOS Simulators, and diagnosing visual regressions all previously required exiting the terminal workflow.

Computer use for Claude was first introduced by Anthropic in October 2024 as an API capability, aimed at developers building agents that could interact with graphical interfaces. The CLI integration brings that capability directly into the developer's own workflow rather than requiring it to be built into a product.

The feature is currently a research preview. It is not available on Team or Enterprise plans, on Linux or Windows, in non-interactive (headless) sessions with the -p flag, or when accessing Claude through third-party providers like Amazon Bedrock, Google Vertex AI, or Microsoft Foundry. Users on those platforms need a separate claude.ai account with a Pro or Max subscription to access it.

What's Next

Computer use in the Claude Code CLI currently lacks a few controls available in the Claude Desktop app: a configurable denied apps list, an optional auto-unhide toggle, and Dispatch integration. Anthropic's documentation notes these gaps explicitly, suggesting they are on the roadmap for parity.

The broader trajectory is toward agentic workflows where AI handles more of the build-test-debug loop without human hand-off points. Computer use in the CLI removes one of the last friction points in that loop for macOS developers.

Sources

• Claude Code computer use documentation, Anthropic official docs
• Anthropic announces Claude Code and Cowork computer use, TechTimes, March 23, 2026
• Claude release notes, March 23, 2026, ReleaseBot