Anthropic's computer use feature lets Claude control your macOS screen — no setup needed for Pro and Max subscribers. Here's what it does, what it can't, and who should try it now.
Anthropic has quietly rolled out one of its most ambitious features yet: computer use — the ability for Claude to take control of your macOS desktop, open apps, navigate your browser, click through interfaces, and fill in spreadsheets, all without you touching the keyboard. It's available right now as a research preview inside Claude Cowork and Claude Code for Pro and Max subscribers, with no extra setup required beyond granting permission.
This is not a polished, production-ready feature. Anthropic is calling it a research preview and has been explicit about its limitations. But for Mac-based agency owners, solo operators, and small dev teams, it's worth understanding what you're actually getting — and what you should avoid until it matures.
On March 23, 2026, Anthropic announced via the Claude official X account:
"You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets — anything you'd do sitting at your desk."
The feature shipped inside the Claude desktop app for macOS and is accessible immediately to anyone on the Pro plan (~$20/month) or Max plan ($100–200/month). There is no separate waitlist or API key required. You opt in, grant permission, and Claude can start acting on your screen.
Community discussion has been limited since launch — no significant Reddit threads, no Hacker News posts, and minimal reaction on X beyond the official announcement. That's likely a combination of the macOS-only limitation, the research preview label, and the Pro/Max paywall keeping it out of developer forums for now.
Research Preview — Not Production-Ready
Here's what the feature enables, according to Anthropic's support documentation and early coverage:
Crucially, Claude doesn't reach for screen control first. As The Tech Outlook notes, "Claude will reach for the most precise tool first... When there isn't a connector, Claude can directly control your browser, mouse, keyboard, and screen to complete tasks." In practice, this means Claude will prefer a native Slack or Google Calendar integration over clicking through the GUI — screen control is the fallback for everything else.
Dispatch is Anthropic's remote task handoff system. With computer use enabled, you can start a task on your phone, hand it off to your desktop via a QR code or mobile prompt, and Claude will execute it on your Mac while you're away from the machine. One YouTube demo showed this working for project file access and browser navigation — useful for solo operators who want to queue tasks remotely.
Your desktop app needs to be open and your Mac awake for this to work. No background-daemon magic here.
Screen control is a significant trust surface, and Anthropic has built in several guardrails:
The permission model is more conservative than, say, early Auto-GPT implementations from 2023, which would barrel through tasks without pausing. Whether the safety layer is sufficient for real-world use is an open question — no external privacy audits or third-party tests have been published yet.
Avoid Sensitive Data
For a 1-5 person team running on macOS, the practical near-term use cases are narrow but real:
What this doesn't replace yet: any workflow where speed and reliability are critical, any task touching sensitive data, or anything you need to run on Windows or Linux.
| Plan | Price | Computer Use Access | |---|---|---| | Free | $0 | ❌ Not available | | Pro | ~$20/mo | ✅ Research preview | | Max | $100–200/mo | ✅ Research preview + priority access + 5x–20x usage limits |
For most small teams, Pro is the right entry point to test this. Max makes sense if you're already hitting Pro's prompt limits on heavy Cowork or Code sessions — the computer use feature itself doesn't require Max, but sustained long sessions do.
This feature is part of a broader agentic push from Anthropic across its 2025–2026 product arc: Claude Code for terminal control, Projects for persistent memory, Opus 4.6 for complex reasoning, and now computer use for GUI-layer automation. The direction is clear — Anthropic wants Claude to be the operative layer for your entire workday, not just a chat window.
The competitive angle is real. No other major AI subscription at the Pro tier ($20/mo) currently offers GUI screen control. OpenAI has hinted at desktop agent capabilities, and tools like Cursor dominate IDE-level control, but a general-purpose screen agent at this price point is new territory. The catch: "new territory" also means unproven territory.
For context, Auto-GPT's early screen control pilots in 2023 were exciting and unreliable in roughly equal measure. They matured into more structured agent frameworks over 18+ months. Claude's computer use is starting from a stronger foundation — better models, tighter safety controls — but the reliability curve is still ahead of it.
Best First Test
Weekly AI tool reviews, news digests, and how-to guides.
Join 12,000+ builders