Claude can now operate your computer without you touching the keyboard. Anthropic shipped computer control to Code and Cowork as a research preview this week, letting Claude autonomously open files, navigate browsers, run dev tools, and execute tasks while you’re away.
This isn’t theoretical. It’s shipping to Claude Pro and Max subscribers on macOS right now.
What Computer Control Actually Does
Claude has had autonomous capabilities baked into the 3.5 Sonnet model since 2024. What’s new is the interface: you ask Claude to do something on your computer, grant permission once, and it executes. No setup required.
The flow looks like this:
- You describe a task: “Pull the Q1 financial data from our shared drive and create a summary spreadsheet.”
- Claude asks for permission to control your screen and interact with applications
- You approve
- Claude navigates your file system, opens apps, and completes the work while you step away
This is fundamentally different from traditional automation or scripting. Claude doesn’t need custom workflows, APIs, or configuration. It sees what you see and acts accordingly—the same way a person would if you handed them control of your mouse and keyboard.
The Constraints That Matter Right Now
Computer control is shipping in public preview, which means it has real limitations.
Platform: macOS only, for now. Windows and Linux support are presumably coming, but Anthropic hasn’t committed to a timeline. If you’re on Linux or Windows, this feature doesn’t exist yet.
Scope: The preview is limited to Code and Cowork—Claude’s code-focused and collaborative environments. Desktop Claude or the web interface don’t have this yet.
Audience: Claude Pro and Max subscribers only. If you’re on Claude Free, you’re waiting.
The permission model is per-session. Claude doesn’t get persistent access to your machine. You’re explicitly approving each autonomous action, which is the right security trade-off for a preview feature.
Why This Matters More Than Another AI Update
Most AI product announcements are incremental—faster inference, better reasoning, more context. Computer control is structural. It moves Claude from a tool that answers questions to a tool that executes tasks without human intervention in the loop.
The threshold for “good enough to use autonomously” is lower than for “good enough to give me an answer to verify.” Claude makes mistakes. It hallucinates. It misreads UI elements. You probably wouldn’t notice when it gets a research question 85% right. You absolutely will notice when it opens the wrong file or clicks the wrong button.
Anthropic is aware of this. The research preview framing—not a full release—suggests they’re testing failure modes before shipping at scale. That’s responsible. Watch for what breaks in the preview phase. That’s where the real product design decisions happen.
What to Test Today
If you have Claude Pro or Max and run macOS, the feature is available now in Code and Cowork. Start with tasks that are repetitive but low-stakes: downloading files, organizing folders, extracting data from web pages, renaming batches of documents.
Don’t start with “execute this critical business workflow unsupervised.” Start with “can Claude handle this task 9 times out of 10?” You’ll find the boundaries fast.
Document what fails. Anthropic is reading the preview feedback. What breaks here shapes whether computer control becomes reliable enough for production work.
