Claude Code now controls your Mac desktop

Claude Code just got a new trick: it can take direct control of your computer when it needs to. Anthropic says the tool can now “point, click, and navigate what’s on your screen” to open files, use the browser, and run dev tools automatically, but only as a research preview for now.

That matters because this is a shift from an assistant that suggests actions to one that can execute them on your desktop. Anthropic is rolling the feature out first to Claude Pro and Max subscribers on macOS, and the company is already warning that the feature is slower and more error-prone than direct app integrations.

What Claude Code can actually do on your screen

Anthropic is splitting its automation strategy into two paths. When possible, Claude will still use Connectors to talk to apps and data sources directly. If that route is unavailable, Claude can ask permission to interact with the machine itself, including scrolling, clicking, and exploring the interface.

That sounds simple, but the practical difference is huge. A connector can query a file system or a service through an API. Desktop control means the model has to interpret pixels, identify buttons, and avoid making mistakes in a live interface. Anthropic’s own wording is pretty blunt: the company says “computer use” takes longer and is more error-prone than using connectors.

• Available first on macOS
• Limited to Claude Pro and Max users
• Released as a research preview, not a full launch
• Can open files, use the browser, and run dev tools
• Can be managed remotely through Dispatch if the target Mac stays powered on

There is also a more casual companion product in the mix: Claude Cowork. Anthropic says both tools can use the desktop when needed, which makes this feel less like a one-off feature and more like a broader push toward computer-operating agents.

Anthropic says the guardrails are real, and limited

Security is the part that will make or break this feature. Anthropic says it has safeguards against prompt injection and blocks some apps by default, including investment and trading platforms and cryptocurrency tools. It also says the model is trained to avoid risky operations like moving money, modifying files, scraping facial images, or entering sensitive data.

But the company is not pretending those controls are airtight. Anthropic warns that the safeguards “aren’t perfect” and “aren’t absolute,” which is the kind of sentence you only write when you know the model can still go off script. The company also says Claude can see anything visible on-screen during computer use, including personal data, sensitive documents, and private information.

“The safeguards aren’t perfect and aren’t absolute,” Anthropic warns in its support material for computer use.

That warning should be read literally. If you hand an agent desktop access, you are also handing it whatever appears in that workspace. For developers, that means the sensible starting point is a clean test machine, dummy credentials, and low-stakes workflows. For everyone else, it means the first demo should probably be a boring task, not your bank account or tax folder.

How Claude Code compares with other AI desktop agents

Anthropic is entering a crowded race. Perplexity recently rolled out Personal Computer, Manus launched My Computer, and