Claude Computer Use Setup: Pro + Max Guide (Mac & Windows)

Claude can now drive your Mac or PC. Not through an API, not through a dev preview you have to beg your way into — through the desktop app you already have, on the Pro or Max plan you already pay for. It’s one toggle. Flip it, and Claude starts clicking buttons and opening files on your behalf.

But it’s a research preview, which is Anthropic-speak for “this will sometimes break in embarrassing ways.” Here’s the full setup — every permission, every gotcha, and what actually works today versus what’s still rough around the edges.

What Claude Computer Use Actually Does

Claude watches your screen via screenshots, then clicks, types, and scrolls to finish a task. Three things make the April 2026 rollout different from the October 2024 API beta:

It’s inside the apps you already use. Claude Desktop, Cowork, and Claude Code — no custom code, no API setup.
It’s gated to Pro and Max subscriptions — not Team or Enterprise (yet). And yes, that’s the opposite of Claude for Word, which is Team/Enterprise only. Confusing. Welcome to 2026.
It pairs with Dispatch, Anthropic’s mobile app — you can message a task from your phone and Claude will do it on the Mac sitting on your desk.

Anthropic’s own framing: “When Claude doesn’t have access to the tools it needs, it will point, click, and navigate what’s on your screen to perform the task itself.”

And it won’t do all three at once. Claude tries the “most precise tool first” — connectors (Gmail, Slack, Drive) if they exist, browser control if not, screen clicking as the last resort.

Who Can Turn It On

Claude plan	Computer Use in Cowork/Code?
Free	No
Pro ($20/mo)	✅ Yes
Max ($100 or $200/mo)	✅ Yes
Team ($25/seat/mo)	Not yet
Enterprise	Not yet

Yes, you read that right. A $20/month Pro subscriber gets it. A $25/seat Team customer doesn’t. Anthropic is rolling consumer features to Pro/Max first, enterprise features to Team/Enterprise first. Claude for Word went the other way. Your wallet is the new plan chart.

Enable It on macOS (The 2-Minute Path)

Step 1: Update Claude Desktop.

Download or update from claude.com/download. If you already have Claude Desktop, launch it and let it auto-update. You need the April 2026 build or newer.

Step 2: Toggle Computer Use on.

In Claude Desktop, click the gear icon → Settings → General (under “Desktop app” section) → find “Computer use” → toggle on.

That’s it for the app side. Now the Mac needs to grant permissions.

Step 3: Grant macOS permissions when prompted.

The first time you ask Claude to do something on your computer, macOS shows two prompts:

Accessibility — lets Claude click, type, and scroll
Screen Recording — lets Claude see what’s on your screen

Click through each. The prompt has a “Open System Settings” shortcut that takes you straight to Privacy & Security → Accessibility. Add Claude Desktop and toggle it on. Same for Screen Recording.

Heads up: after granting Screen Recording, macOS often wants you to restart Claude Desktop before the permission takes effect. If Claude says “I can’t see your screen,” quit the app completely (⌘Q) and reopen.

Step 4: Open Cowork and start a session.

Cowork is the conversation surface where Computer Use shines. You can also use it inside Claude Code if you’re a developer.

Step 5: Try a first task.

Good first prompt: “Open my Downloads folder, find any screenshots from today, and move them into a new folder called Screenshots-today.”

Watch Claude request permission to use Finder. Click Allow. Then watch it click around. It’s eerie the first time.

Enable It on Windows

Step 1: Install Claude Desktop for Windows.

Same download page: claude.com/download. Windows x64 only — Windows on ARM (Surface Pro X, some Copilot+ PCs) is not supported at launch.

Step 2: Toggle Computer Use on.

Same path: Settings → General → Computer use. Toggle on.

Step 3: Allow the Linux VM to download.

This is where Windows setup diverges. Cowork runs its sandboxed environment inside a Linux virtual machine on your PC. The first run triggers a download of around 2 GB (per community reports — Anthropic hasn’t documented the exact size). It takes a few minutes. Don’t close Claude Desktop while it’s downloading.

Step 4: Grant app access as Claude requests it.

Windows doesn’t require the Accessibility/Screen Recording dance macOS does. Instead, Claude asks per-app inside the Cowork interface — “Can I use Excel?” — and you approve in the sidebar.

Step 5: First task.

Same idea. Ask it to organize files, draft an email in Outlook, or pull data from a PDF. Same permission prompts.

Grant Folder Access Inside Cowork

Separate from the OS permissions, Cowork itself needs explicit permission to touch your files.

In Cowork: Settings → Add Folder (or Grant Access on some builds) → pick the folder you want Claude to work with. Your Documents folder, a specific project folder, or your Desktop are reasonable starts.

Cowork runs in a Linux VM, so file access is mediated. Claude can’t wander outside the folders you approved. Even if it tried, the VM sandbox blocks it.

Anthropic’s official guidance: “While the VM provides isolation, Claude does have access to local files you grant it permission to access.” Translation: the VM is the prison wall, the folder grants are the keys, you’re the warden.

The Safety Model (What Claude Asks Permission For)

Every action outside the VM goes through a permission prompt. Here’s what’s automatic, what’s per-app, and what’s blocked entirely.

Category	Behavior
VM-internal tasks	No prompt — Claude works freely inside the sandbox
File read in granted folder	No prompt after initial folder grant
File write in granted folder	No prompt after initial folder grant (review planned actions for sensitive files)
File delete anywhere	Per-action prompt — must click Allow
External app access (Safari, Chrome, Excel, etc.)	Per-app prompt the first time
Banking, investment, crypto apps	Blocked by default
Screen view	Automatic when Computer Use is active (Claude sees anything on screen)

That last one is the real catch. Anthropic’s own warning: “Claude will be able to see anything visible on-screen, including personal data.” Screenshots are how Claude understands what to click. If your bank’s website is open in a background tab, Claude will see the balance. That’s not a bug — it’s how the feature works.

Pro vs Max Plan Limits

The short version: usage is shared across Claude, Claude Code, and Cowork. Everything counts against your plan’s budget.

Plan	Price	Computer Use sessions (reported)
Pro	$20/mo	Shared usage pool with Claude. No specific session count documented.
Max 5x	$100/mo	Around 50 sessions/month (per independent community testing)
Max 20x	$200/mo	Around 50 sessions/month — higher token budget per session

Community reports on Reddit suggest Pro users hit limits faster when running long Computer Use tasks. Max 5x is the sweet spot for frequent use. Max 20x gives bigger token budgets per task — useful if you run Claude across 50-page documents, long browser research sessions, or multi-step automations.

If you blow through your limit, Anthropic offers “enable extra usage” (pay-per-token via Console) or an upgrade prompt.

What Actually Works Today (and What Doesn’t)

Being honest about the research preview — this isn’t production-grade yet. Here’s what works and what’s still rough.

Works well:

Opening and organizing files in the Finder/Explorer
Simple browser research on well-structured sites (news, docs, basic search)
Drafting emails in Gmail web, Outlook web
Data entry into forms where each field is clearly labeled
PDF reading and summarization
Working inside spreadsheets when layouts are standard

Still rough:

Heavy-JavaScript single-page apps (often misclicks or loses state)
Long browser sessions with many tabs (Claude loses context)
Anything requiring drag-and-drop (limited support)
Apps with non-standard UI (custom windowing, dense tables)
Scrolling through long documents — Claude sometimes misreads position
Google Drive in streaming mode — Cowork sees Drive files as empty stubs. You must set files to “Available offline” first, then Cowork can actually read them. This single limitation has caused some of the most viral “Cowork doesn’t work” posts of the launch week.

Reality-check numbers from the community:

About 50% of Cowork desktop sessions break irreversibly per one developer’s week of logs. Anecdotal, but a specific number that matches multiple other posts describing “had to restart Cowork mid-task” frustration.
Dispatch fails around 70% of the time per another developer’s week — usually on scheduled tasks hitting permission prompts while the user is away from the desktop.
macOS VM disk image fills silently. The VM’s sessiondata.img or rootfs.img can grow until all new sessions fail with RPC error: process with name already running. No warning. Fix: delete the image from ~/Library/Application Support/Claude/ and restart.
API 401 auth split between Chat and Cowork. Users report the Claude chat works fine but Cowork returns Failed to authenticate. API Error: 401. Usually fixed by signing out of the desktop app entirely and back in.
Windows gotchas: Cowork writing to a different Windows user context makes folders read-only. Cowork can save to the OneDrive-mirrored Desktop instead of your real Desktop (files “disappear”). Residual corporate SharePoint policy entries in the Windows registry can block writes entirely — you may need registry edits to clear them.

Crashes and outages:

Desktop app crashes after 30+ minute sessions
Stalled research jobs needing manual abort
Straight API Error: 500 outages during the rollout window

Anthropic frames this honestly: “Computer use is still in its early stages compared to Claude’s proficiency in coding or text interaction.” They’re right. Use it, but don’t trust it blindly.

When to Use Computer Use vs MCP

This confuses a lot of developers. Here’s the simple version.

Use MCP (Model Context Protocol) connectors when:

The tool has a connector or API
You need reliability and auditability
You’re automating repeated tasks
The workflow involves sensitive data

Use Computer Use when:

The tool is GUI-only with no API (legacy desktop software)
You’re clicking through internal dashboards with no integration
The task is one-off exploration
You need to handle browser-only workflows that don’t expose APIs

Anthropic’s internal priority confirms this: Claude reaches for the most precise tool first — “connectors to services like Slack or Google Calendar” — and falls back to screen clicking only when nothing else works. If you catch yourself saying “there should be a connector for this,” there probably is. Check the Integrations panel first.

Dispatch Integration (Phone → Desktop)

Dispatch is the mobile companion. You pair your iPhone or Android once, then message Claude from your phone while your laptop is back at your desk.

Pairing:

Open Claude Desktop → Cowork → find the Pair with your desktop option
On your phone, open the Claude/Dispatch app → tap Pair with desktop
Scan the QR code shown in Claude Desktop
Done

What you can do:

Send a task from your phone: “Convert the pitch deck on my desktop to PDF and attach it to my 3pm meeting invite.”
Claude executes it on your desktop using Computer Use
Same permission prompts apply — your desktop buzzes with approval requests
When the task finishes, Claude replies in the mobile thread

Anthropic’s engineer Felix Rieseberg described Dispatch as: “One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work.”

Yes, it works. Yes, it’s weird watching your laptop move on its own. No, you probably shouldn’t run Dispatch tasks on your work machine without clearing it with IT first.

What This Means for You

If you’re a knowledge worker who spends time in repetitive desktop tasks: Pro at $20/month is the cheapest way to try this. Point it at file organization, inbox triage, or browser research. Budget 5-10 hours to learn its limits. You’ll save real time on simple, repeated workflows.

If you’re already on Max for the context window: Computer Use is a bonus, not a headline feature yet. Don’t upgrade for it. But do enable it and experiment — the Max 20x token budget makes long Computer Use tasks actually feasible.

If you’re a developer weighing MCP vs Computer Use for a new workflow: Build MCP first, always. Computer Use is the fallback for legacy GUIs. A 2-hour MCP build today will save you 200 hours of flaky screen automation over 12 months.

If you’re in an organization that hasn’t approved it: Don’t install on work machines without IT sign-off. Claude will see anything on your screen — including things your security team probably wants kept private. The safety model is good, but corporate policies come first.

If you’ve never used AI to do real work before: Start with simpler tools (ChatGPT, Claude chat). Computer Use isn’t a beginner’s AI. The permission prompts, VM setup, and Dispatch pairing are more infrastructure than most first-time users want.

Two quieter things worth knowing:

Anthropic penalizes telemetry opt-out. A community post with 4,000+ likes this week surfaced that disabling telemetry in Claude Desktop drops your cache from 1 hour to 5 minutes — a 12x performance hit. Opt out if you want, but know the tradeoff.
Shared payment cards trigger instant bans. Multiple research labs this week got all their Max subscriptions suspended for paying with the same institutional card. If you’re setting up Claude for a team, use distinct payment methods per user.

What It Can’t Do Yet

A few honest limits worth stating:

No ARM Windows support. Surface Pro X, Snapdragon-based Copilot+ PCs are out.
No Linux desktop support. Server Linux can run the API beta, but Cowork desktop is Mac/Windows only.
No iOS or Android direct execution. Dispatch bridges mobile to desktop, but Claude can’t drive a phone directly.
No sensitive-vertical apps. Banking, investment, crypto are blocked. Anthropic explicitly advises against using it for financial, legal, or medical workflows.
No reliable long-session execution. 30+ minute sessions sometimes crash. Break large tasks into smaller prompts.

The Bottom Line

Claude Computer Use in Cowork is the first consumer-grade “AI controls my computer” feature that doesn’t require custom engineering. The setup is 5 minutes on Mac, slightly longer on Windows (thanks to the VM download). The safety model is solid — VM sandbox, per-app permissions, no access to sensitive categories by default.

But it breaks. A lot. Research preview is the honest label. Treat it as a power-user experiment, not a production tool. If you’ve been waiting for the “just tell AI what to do on my computer” moment, this is the closest we’ve gotten — and the next six months will determine whether it crosses into reliable territory or stays as a cool demo.

For $20/month on Pro, it’s worth the flip of the toggle.

Want to go deeper? Our Claude Cowork Essentials course walks through the full Cowork workflow. For developers building real automations, Claude Code Mastery covers the Computer Use + MCP stack together.

Sources: