Claude Computer Use: 50% Success Rate and Still Worth It

Claude can control your Mac now. And it’s roughly 50% reliable.

That’s the honest number. MacStories tested 12 different operations — file management, data analysis, web research, email — and got about half of them done without issues. The simple stuff (open this app, find this file, summarize this document) works. The complex stuff (analyze this spreadsheet, build a presentation from the results, share it via email) breaks mid-task or produces incomplete results.

So why is this still worth paying attention to?

Because that 50% is doing things no other AI could do on your Mac three days ago.

What Claude Computer Use Actually Does

Since March 23, Claude Pro and Max subscribers can grant Claude permission to control their Mac. Not through an API. Not through a developer tool. Through the actual screen — cursor movements, clicks, typing, opening apps, browsing the web.

It works through Cowork, Anthropic’s desktop app. You give Claude a task, grant screen access, and watch it work. Or walk away and come back to the results.

The clever part is the connector-first design. When you ask Claude to send an email, it doesn’t click around in Gmail. It uses the Gmail connector — a direct API connection that’s faster and more reliable. Screen control only kicks in when no connector exists for the task. This reduces the error surface considerably.

What you need:

• Claude Pro ($20/month) or Max ($100-200/month) subscription
• macOS (Windows support coming, no date yet)
• Claude Desktop app (latest version)
• Willingness to let an AI move your cursor

What Works Reliably

After reading through early reviews from PCWorld, Tom’s Guide, CNBC, and a handful of Substack writers who’ve been running it for 72 hours, the patterns are clear.

Solid tasks (high success rate):

• Opening apps and navigating to specific features
• Finding and summarizing documents on your Mac
• Simple web searches and pulling data from pages
• File management — moving, renaming, organizing
• Reading spreadsheets and pulling specific data points
• Filling in forms with structured data
• Taking screenshots and annotating them

Flaky tasks (50/50 or worse):

• Multi-step workflows that span multiple apps
• Creating presentations from analyzed data
• Complex spreadsheet operations (formulas, charts)
• Sending results via email or Slack (often stalls)
• Any task that requires interpreting ambiguous UI elements
• Workflows that require scrolling through long pages

Broken tasks (don’t bother yet):

• Anything involving sensitive data (Anthropic explicitly warns against this)
• Tasks requiring precise pixel-level interactions
• Complex drag-and-drop operations
• Multi-monitor setups
• Anything on Windows or Linux (macOS only)

The 50% Success Rate in Context

Half the time sounds bad. But here’s the context that matters.

This is a research preview, not a finished product. Anthropic is calling it exactly that. They’re collecting data on what breaks and why. The 50% is a starting point, not a ceiling.

And the tasks it fails on are genuinely hard. Getting an AI to navigate a spreadsheet, create a chart, export it to a slide deck, and attach it to a calendar invite — that’s a multi-step workflow with visual interpretation, app switching, and context maintenance. Humans mess this up too, just differently.

The tasks it succeeds at — finding files, summarizing documents, simple data extraction — represent real time savings. If Claude can reliably handle the “find this, read this, tell me what’s in it” tasks, that’s meaningful even if it can’t yet do the end-to-end workflow.

How It Compares to Other Desktop Agents

Claude isn’t the only AI that can control your computer. Here’s how the options stack up:

The honest take:

• Claude has the best integration if you’re already in the Anthropic ecosystem. The connector-first approach is smarter than pure screen control, and running locally means your data stays on your machine. But macOS-only and 50% success limits the audience.

• Manus is the easiest to start with — no local setup, just a browser. But the unpredictable credit consumption means your $39/month can evaporate fast on complex tasks. And your data runs through their cloud.

• OpenClaw gives you the most control and it’s free — but you need to set it up, configure your own LLM, and maintain it. It’s a developer tool, not a consumer product.

• Perplexity Computer is the most expensive and least flexible, but it’s excellent for web-based research tasks specifically.

For most people reading this? Claude at $20/month is the best starting point. You get computer use plus everything else in Claude Pro (the chatbot, Cowork, Dispatch). The others are worth exploring only if Claude’s limitations block you.

What Early Users Are Actually Doing With It

The most interesting trend in the first 72 hours isn’t the enterprise use cases Anthropic demos in their marketing videos. It’s the scrappy, entrepreneurial use cases popping up on X.

People are using Computer Use for:

• Ghostwriting services — Claude reads a client’s notes, writes drafts, pastes them into Google Docs, formats them
• Competitor research — Open 10 competitor websites, screenshot pricing pages, compile into a comparison doc
• Data entry — Read a PDF invoice, enter line items into a spreadsheet, move to the next invoice
• Social media scheduling — Draft posts, open Buffer/Hootsuite, paste and schedule
• Lead research — Search LinkedIn profiles, summarize backgrounds, add notes to a CRM

None of these are flashy. All of them save real hours. And they’re the kinds of tasks that work within the 50% reliability range because each step is simple — the complexity comes from repetition, not depth.

The Security Question

Let’s address the elephant: you’re giving an AI full control of your screen.

Anthropic has built in a permission-first model. Claude asks before accessing each new app or system. It includes prompt injection detection — meaning if a webpage tries to trick Claude into doing something unexpected, the system catches it.

But Anthropic themselves say: don’t use this for sensitive information. No banking. No password managers. No confidential documents you wouldn’t want a third party to see.

That’s a reasonable precaution for a research preview. It should be a red line for you too, at least until the feature graduates from beta.

Should You Try It?

Yes, if:

• You have a Claude Pro subscription and a Mac
• You spend time on repetitive computer tasks (file organizing, data entry, simple research)
• You’re comfortable with a tool that fails sometimes and needs supervision
• You want to learn how AI agents work before they become mainstream

Not yet, if:

• You need 100% reliability for mission-critical tasks
• You’re on Windows or Linux
• You handle sensitive financial or health data on your computer
• You’re not comfortable granting screen access to an AI

The 50% success rate will improve. It improved from the 2024 API-only beta to this native Mac integration, and it’ll improve again as Anthropic collects more real-world usage data. But right now, treat it as a capable but unreliable assistant — check its work, start with simple tasks, and scale up as you learn what it handles well.

We built a full Claude Computer Use course that walks through setup, real tasks, security, and practical workflows if you want to go deeper.

Sources:

• Claude controlled my Mac for half an hour — PCWorld
• I tried Claude’s new Cowork feature — Tom’s Guide
• Claude Code and Cowork can now use your computer — Engadget
• Anthropic giving Claude the ability to use your Mac — 9to5Mac
• Anthropic Claude AI agent — CNBC
• Claude can now control your Mac — VentureBeat
• Let Claude use your computer in Cowork — Claude Help Center
• OpenClaw vs Perplexity vs Manus vs Claude — PC Build Advisor
• Is Claude Cowork an Agent Yet? Honest Review — Jock.pl