Claude computer use works, with caveats — here’s what 24 hours of real testing found.

Claude computer use shipped March 23, 2026 — and I woke up that morning to find a spreadsheet populated, three browser tabs open with research, and a summary document waiting in my text editor. No macros. No Zapier chain. Just Claude, operating my Mac like a (slightly cautious) remote assistant.

This is Anthropic’s computer use feature, and it shipped as a research preview on March 23, 2026 for Claude Pro and Max subscribers. It’s the first time a leading AI assistant has offered mainstream users the ability to hand over actual desktop control — not through a plugin, not through an API integration, but by literally clicking around your screen.

I’ve been testing it for about 24 hours now. Here’s where it stands.

Quick Verdict

Aspect	Assessment
Overall Score	★★★★☆ (3.7/5) — genuinely useful, clearly unfinished
Best For	Repetitive desktop tasks you’d normally script or delegate
Availability	Claude Pro ($20/mo) and Max ($100/mo) subscribers only
Reliability	~70-75% task completion on structured workflows
Speed	Slower than you’d do it manually. Faster than teaching someone else to do it
Privacy Risk	Medium — Claude sees your screen. Anthropic says nothing is stored

Bottom line: This is the most practical computer-use implementation from any major AI lab. It works. It’s also slow, occasionally confused, and not something I’d run unsupervised on anything important. But for the right tasks, it’s already saving me real time.

How Claude Computer Use Actually Works

The setup is simpler than I expected. You open Claude on your phone (or the desktop app — but the phone workflow is the point), type a task description, and Claude takes control of your Mac or PC. It can open applications, navigate web pages, type text, click buttons, fill form fields, and move between windows.

No step-by-step scripting required. You describe the outcome — “find the Q1 revenue numbers from the finance dashboard and put them into the budget spreadsheet” — and Claude figures out the clicks.

Under the hood, it’s taking screenshots of your display at regular intervals, interpreting what it sees, deciding what action to take next, executing that action, then taking another screenshot to verify the result. It’s a perception-action loop running on your actual desktop, not a sandboxed simulation.

What Claude computer use can do right now:

1. Open and navigate apps — launch browsers, switch between windows, use menus
2. Browse the web — search, click links, fill forms, extract information from pages
3. Work in spreadsheets — enter data, apply formulas, create basic charts
4. Draft and edit documents — open text editors, write content, format text
5. Move data between apps — copy from a website, paste into a spreadsheet, repeat
6. Execute multi-step workflows — chain several of the above into a single task

What it can’t do (yet): anything requiring drag-and-drop precision, most keyboard shortcuts, complex image editing, or tasks that need sub-second timing.

What I Actually Tested

I ran Claude through a dozen real tasks over the past day. Not benchmarks — stuff I’d actually need done on a Monday morning.

Test 1: Research and Populate a Spreadsheet

I asked Claude to find pricing information for five project management tools from their websites and fill in a comparison spreadsheet I’d already set up in Google Sheets.

Result: It got four out of five correct. Missed one because the pricing page had a dynamic element that wasn’t visible in the screenshot capture. Took about 8 minutes for what would’ve taken me 15. Not faster for one-off tasks, but I was drinking coffee instead of clicking.

Test 2: Fill Out a Repetitive Web Form

I had a batch of 12 entries to submit through an internal tool with a web interface. Each entry had the same five fields, different values. I pasted the data into my message and told Claude to fill them in.

Result: 11 out of 12 correct. It stumbled on a dropdown menu that required scrolling, entered the wrong value, then corrected itself on the next entry. Total time: about 20 minutes. I would’ve hated every second of doing this manually.

Test 3: Open an App, Draft a Summary, Save It

I told Claude to open TextEdit, write a brief summary of a PDF I had open in Preview, and save the file to my Desktop.

Result: It worked. The summary was decent (not great — it was working from screenshots of the PDF, not parsing the actual text). Saved the file correctly. The whole thing felt like watching someone remote-desktop into my machine.

Test 4: Complex Multi-App Workflow

I asked Claude to check my email for a specific thread, extract a few data points from it, add them to a Numbers spreadsheet, then open Slack and send a message summarizing the update.

Result: Partial success. It found the email, got the data, updated the spreadsheet correctly. Then it opened Slack but typed the message into the wrong channel. I caught it before it sent. This is the kind of near-miss that makes supervision non-optional right now.

Where Claude Computer Use Struggles

I need to be direct about the limitations, because the demo videos make this look smoother than it is.

It’s slow. Each screenshot-analyze-act cycle takes a few seconds. A task that involves 30 clicks might take 3-4 minutes. You’re not going to watch Claude do something faster than you could. The value is doing it while you’re not at your desk, or while you’re focused on something else.

Visual interpretation fails happen. Claude reads your screen through screenshots. If a UI element is ambiguous, overlapping, or requires scrolling to reveal, it can misclick. Dark mode and non-standard UI layouts trip it up more than standard interfaces.

No undo awareness. If Claude makes a mistake (wrong cell in a spreadsheet, wrong field in a form), it doesn’t always recognize the error. It just keeps going. This is the single biggest reason you can’t walk away and trust it completely.

Browser-heavy tasks are inconsistent. Modern web apps with dynamic content, modals, cookie banners, and authentication flows create edge cases that Claude handles unpredictably. Simple static websites? Fine. A SaaS dashboard with a dozen floating elements? Coin flip.

Privacy is a real consideration. Claude is seeing your screen. All of it. Anthropic states that screenshots aren’t stored or used for training, but you should think carefully about running this with sensitive data visible. I closed my password manager and banking tabs before every test. You should too.

Claude Computer Use vs. the Competition

This isn’t the first computer-use agent. But it’s the first from a tier-one AI lab that’s available to regular subscribers rather than developers or enterprise buyers.

Feature	Claude Computer Use	Manus My Computer	GPT-5.4 Computer Use
Availability	Pro/Max subscribers ($20-100/mo)	Separate subscription	Not yet launched for consumers
Setup	Built into Claude app	Separate agent install	N/A
Reliability	~70-75% on structured tasks	~65-70%	N/A
Mobile trigger	Yes — send from phone	No	N/A
Speed	Slow but steady	Similar	N/A
Platform	Mac and Windows	Mac, Windows, Linux	Expected Mac/Windows

The mobile trigger angle is what separates Claude’s implementation. Manus and OpenClaw both require you to be at your computer to set up the task. Claude lets you message from your phone while you’re on the train and come back to a completed task. That’s a different product category.

Claude Code Channels: The Developer Side

Anthropic shipped a second feature alongside computer use: Claude Code Channels. This connects Claude Code (Anthropic’s developer tool for writing and editing code) to messaging platforms like Telegram and Discord.

The use case: you’re away from your dev machine, you realize you need to fix something or run a script, and you message Claude Code through Telegram. It executes on your machine. No VPN, no SSH, no remote desktop.

I’m a Claude Code user, and this immediately made sense to me. The number of times I’ve wanted to trigger a build or run a test from my phone is embarrassing. This solves that specific itch.

It’s a developer-only feature and requires Claude Code to be running on your machine, so it’s not for everyone. But if you’re already in that ecosystem, it’s a natural extension of the agentic AI shift that dominated GTC last week.

Who Should Try Claude Computer Use Right Now

Yes, try it if:

• You already pay for Claude Pro or Max — it’s included, no extra cost
• You have repetitive desktop tasks you dread (form filling, data entry, research compilation)
• You want to delegate tasks while away from your desk
• You’re comfortable supervising AI output before it matters (checking the spreadsheet before you send it)

Not yet if:

• You need precision on high-stakes tasks — the error rate is too high for unsupervised financial data, client communications, or anything legal
• You work primarily on mobile or tablet — computer use requires a Mac or PC running the Claude desktop app
• Your workflows involve complex authentication flows or enterprise SSO — Claude trips on multi-step logins
• Privacy is paramount — until there’s more transparency on what happens with screen captures, cautious users should wait

The Bigger Picture: AI Moves From Conversation to Action

Computer use is the logical next step in a trend that’s been building all month. Agentic AI stopped being a concept and became the default at GTC 2026. Google shipped Gemini agents that act on your behalf across Gmail and Drive. OpenAI has been building toward its own computer-use offering.

What makes Anthropic’s move interesting is the packaging. This isn’t a research paper. It isn’t a limited beta for enterprise partners. It’s a feature that any Claude Pro subscriber can turn on today. That’s a deliberate choice to put autonomous computer control in the hands of millions of paying users, not just developers and researchers.

Whether that’s bold or premature depends on how the next few weeks go. If the error rate drops and the speed improves, this becomes the feature that justifies the subscription. If people lose work because Claude deleted the wrong file or sent a message to the wrong person, Anthropic will hear about it loudly.

My bet: it’ll land somewhere in between. Good enough to be useful for specific tasks, not reliable enough to trust blindly. That’s exactly where most AI agent tools sit right now — and exactly the trajectory that leads to something much more reliable in 6-12 months.

How to Get Started With Claude Computer Use

1. Confirm your subscription. You need Claude Pro ($20/month) or Claude Max ($100/month). Free-tier users don’t have access
2. Install the Claude desktop app on your Mac or PC if you haven’t already
3. Enable computer use in Settings — it’s under the new “Computer Use” section (labeled as research preview)
4. Start small. Give Claude a simple, low-stakes task: “open Chrome and search for the weather in Denver.” Watch it work. Get a feel for the speed and accuracy
5. Scale up gradually. Try a spreadsheet task, then a multi-app workflow. Always review the output before acting on it
6. Keep sensitive information off-screen during computer use sessions. Close password managers, banking apps, and anything you wouldn’t show a new coworker

The Bottom Line

Claude computer use works, and it’s the first mainstream implementation of something the AI industry has been promising for years. It’s also slow, imperfect, and not ready for unsupervised operation on anything that matters.

That sounds like a lukewarm endorsement. It isn’t. The fact that I can message Claude from my phone and come back to a populated spreadsheet — even one I need to double-check — represents a genuine shift in what an AI assistant can do. Not a theoretical shift. A practical one, available today, included in a subscription I already pay for.

Give it a test this week. Start with something you’d normally put off because it’s tedious. If Claude handles it, you’ll immediately see the value. If it doesn’t, you’ve lost five minutes and learned where the ceiling currently sits.

Either way, computer use is where AI assistants are headed. Anthropic just got there first with something you can actually use.

Frequently Asked Questions

What is Claude computer use?

Claude computer use is a feature that lets Claude autonomously control your Mac or PC — opening apps, browsing the web, typing text, clicking buttons, and moving data between applications. You describe a task in natural language, and Claude executes it by visually interpreting your screen and taking actions, without requiring step-by-step instructions or pre-built integrations.

How much does Claude computer use cost?

It’s included with Claude Pro ($20/month) and Claude Max ($100/month) subscriptions at no additional charge. It launched March 23, 2026 as a research preview. Free-tier Claude users do not have access.

Is Claude computer use safe? Can it see my passwords?

Claude takes screenshots of your display to interpret what’s on screen. Anthropic states these screenshots are not stored or used for training. However, Claude can see everything visible on your screen during a session, including sensitive information. Close password managers, banking tabs, and private documents before starting a computer use session.

How does Claude computer use compare to Manus or OpenClaw?

Claude’s implementation is the first from a major AI lab available to regular consumers (not just developers). Its key differentiator is the mobile trigger — you can send tasks from your phone. Manus My Computer and OpenClaw both require you to be at your computer to initiate tasks. Reliability is comparable across all three, in the 65-75% range for structured workflows.

What are Claude Code Channels?

Claude Code Channels, launched alongside computer use on March 23, 2026, connects Claude Code (Anthropic’s developer tool) to messaging platforms like Telegram and Discord. Developers can trigger code execution, builds, and scripts from their phone without needing SSH or remote desktop access to their development machine.

Can Claude computer use replace a virtual assistant?

For specific, repetitive desktop tasks — data entry, web research, form filling — it can handle the work at a basic level. For anything requiring judgment, context about your preferences, or real-time communication with other people, a human assistant is still necessary. Think of it as a complement to your workflow, not a replacement for a person.

---

Last updated: March 24, 2026. Based on hands-on testing during the first 24 hours of the research preview. Features and reliability may change as Anthropic updates the preview.

Related reading: Claude Opus 4.6 Review | Computer Use Agents Compared: Manus vs OpenClaw vs Claude | Agentic AI Is the New Default