Automating Desktop Tasks with Computer Use

Last verified: 21 April 2026 | Applies to: Pro, Max (research preview, macOS Apple Silicon and Windows x64)

In 30 seconds

Computer Use lets Claude see your screen and control your mouse and keyboard, clicking buttons, filling forms, navigating between apps, and completing tasks that have no API or connector. Available on macOS (Apple Silicon) and Windows (x64) as of 3 April 2026. It is the fallback for when a connector does not exist and the Chrome extension cannot reach it (because it is a native desktop app, not a website). Think: vendor portals with no export button, desktop accounting software, or multi-app workflows where you are copying data between windows. It works, but it is slow and imperfect, with roughly 50% success rate on complex multi-step tasks. Start with low-stakes, repetitive work.

When to use what

graph TD
    A{Does a connector exist?} -->|Yes| B[Use the connector]
    A -->|No| C{Is it web-based?}
    C -->|Yes| D[Use Chrome extension]
    C -->|No| E{Is it a desktop app?}
    E -->|Yes| F[Use Computer Use]
    E -->|No| G[Manual or build a connector]

This decision tree is the key to choosing the right automation approach:

Connector exists (Gmail, Slack, HubSpot, Google Sheets): always use the connector. It is faster, more reliable, and does not require screen access.
Web-based with no connector (vendor portals, niche SaaS tools, internal web apps): use the Chrome extension for browser automation. It reads the page structure directly and is more reliable than screen-based control.
Native desktop app with no connector (Xero desktop, legacy ERP systems, local databases, PDF tools): this is where Computer Use fills the gap. Claude sees your screen and operates the app as a human would.

Computer Use is the option of last resort, not the first choice. It is slower and less reliable than connectors or the Chrome extension. Use it when nothing else can reach the tool you need.

How to set it up

Step 1: Enable Computer Use

In Claude Desktop, go to Settings → Cowork → Computer Use and toggle it on. macOS will prompt you to grant screen recording permission — approve it.

Step 2: Open the app you want to automate

Computer Use works with whatever is visible on your screen. Open the application, log in, and navigate to the starting point of the task you want to automate.

Step 3: Describe the task

In Cowork, describe what you want Claude to do:

I have the Xero desktop app open on the Invoices screen. For each of the 15 unpaid invoices listed, I need you to: click into the invoice, copy the invoice number, amount, and due date, then go back to the list. Compile all 15 into a CSV file on my desktop.

Claude takes a screenshot of your screen, identifies the UI elements, and begins executing the task step by step. You will see your cursor move and windows interact — Claude is controlling your Mac.

Step 4: Monitor the first run

Watch the first execution closely. Computer Use is not hands-off automation. It is closer to a capable but fallible assistant operating your machine. If Claude gets stuck (wrong button, unexpected dialog box, app layout change), it usually pauses and asks for help.

How operators actually use it

Vendor portal data extraction

Many vendors provide data only through a web portal with no API and no export button:

I have the BuildSupply vendor portal open in Safari. Navigate to Reports → Monthly Statement → March 2026. Screenshot each page of the statement (there are usually 4-5 pages). Save the screenshots to my Desktop/vendor-statements/ folder, then extract the line items into a CSV.

Computer Use navigates the portal, captures the data, and structures it — a task that would take you 15 minutes of clicking and copy-pasting.

Multi-app reporting

When your report pulls data from multiple desktop applications:

I need to create a weekly summary. Steps: (1) Open the Numbers spreadsheet at ~/Documents/weekly-metrics.numbers and read the revenue row for this week. (2) Switch to the Slack desktop app and check #sales for the top deal update. (3) Open Mail and find the latest email from our logistics partner with the shipping update. Compile all three into a summary and save it to my Reports folder.

Claude switches between apps, reads data from each, and produces a compiled output. This is the kind of task where you are the integration layer. Computer Use replaces that manual toggling.

CRM data entry from a spreadsheet

When your CRM does not have a bulk import and you are entering records one at a time:

I have a spreadsheet open in Excel with 20 new leads. For each row, switch to the Pipedrive desktop app, click New Contact, fill in: name, email, company, phone number, and source. Save each contact before moving to the next. The spreadsheet columns are: A=Name, B=Email, C=Company, D=Phone, E=Source.

This is tedious, error-prone manual work that Computer Use handles well, provided the form layout is consistent.

Expense receipt processing

I have 12 receipt PDFs open in Preview. For each receipt, read the vendor name, date, amount, and category. Then switch to our expense tracker spreadsheet and enter a new row for each receipt. Save after every entry.

David manages finance for a 20-person construction company that uses three desktop applications with no APIs — a legacy estimating tool, a project management app from the early 2010s, and a local database for supplier records. Each month-end, he spent two to three hours manually copying data between these apps to compile his close package. He started using Computer Use for the most repetitive part: extracting project cost data from the estimating tool into a spreadsheet. The first attempt failed halfway through. Claude clicked the wrong tab and got lost. The second attempt, with more specific instructions about the UI layout, completed successfully. David now uses it for three monthly tasks and estimates it saves about 90 minutes per month. He still watches the first few steps of each run before walking away.

Safety and risk management

Computer Use gives Claude control of your screen, which is powerful but demands care. This section covers how to assess risk, test safely, and avoid the most common problems.

Risk tiers

Risk level	Description	Examples	Recommended approach
Low	Data export from a single, isolated app. No writes, no external systems.	Extracting invoice data from a desktop app into a CSV. Reading values from a spreadsheet. Screenshotting a report.	Good starting point. Watch the first run, then let it work.
Medium	Multi-app workflows or tasks that create or modify data within internal systems.	Copying data between two apps. Filling forms in a CRM. Organising files across folders.	Watch the entire first run. Verify outputs before acting on them. Keep a backup of any files being modified.
High	Anything touching payments, credentials, external-facing systems, or irreversible actions.	Submitting invoices. Sending emails. Interacting with banking portals. Posting to social media. Modifying shared databases.	Not recommended during the research preview. If you must, watch every step in real time and be ready to intervene.

Safe testing checklist

Before running your first Computer Use task, work through this list:

Close sensitive windows. Close any apps showing passwords, banking credentials, personal messages, or confidential information. Claude sees everything on your screen while screen control is active.
Set your display to never sleep. Go to System Settings, Energy, and set display sleep to “Never” (or at least a long interval). A locked screen kills the task.
Start with a non-critical task. Your first Computer Use task should be low-risk: exporting data from one app into a file. Do not start with anything that sends, publishes, or pays.
Watch the first run from start to finish. Sit at your desk and observe. Note where Claude hesitates, misclicks, or gets confused. These patterns help you write better instructions next time.
Maximise the target app window. Claude identifies UI elements visually. A maximised, unobstructed window with standard controls gives the best results.
Keep your desktop clean. Close unnecessary windows and remove desktop clutter. Fewer visual distractions mean fewer misclicks.
Have a way to interrupt. Move your mouse or press a key to interrupt Claude if it starts doing something unexpected. You always retain control.

What to watch out for

~50% success rate on complex tasks. Anthropic is transparent about this. Multi-step tasks with many UI interactions will fail roughly half the time. Start simple and build complexity gradually.
It is slow. Computer Use takes screenshots, analyses them, decides on an action, executes it, then takes another screenshot. A task you could do manually in 5 minutes might take Claude 10-15 minutes. The value is in freeing your attention, not raw speed.
macOS Apple Silicon and Windows x64. Intel Macs and Windows ARM64 are not supported. Windows support was added on 3 April 2026.
Your computer must stay awake. If your screen locks or your machine sleeps during a Computer Use task, Claude loses access and the task fails. Adjust your energy or power settings for longer tasks.
Screen layout matters. Claude identifies UI elements by their visual appearance. If an app rearranges its layout, uses non-standard controls, or has tiny clickable elements, accuracy drops. Maximise the app window and use a clean desktop.
Do not leave sensitive information visible. Computer Use sees everything on your screen. Close any windows containing passwords, personal messages, or confidential data you do not want Claude to process.
Research preview. This feature launched on 23 March 2026 (Mac) and expanded to Windows on 3 April 2026, both as a research preview. Expect improvements but also expect bugs. Do not rely on it for time-critical or high-stakes tasks yet.
Falls back to screen control when no connector exists. This is by design. Computer Use is meant to be the catch-all. But if a connector becomes available for a tool you are automating via Computer Use, switch to the connector immediately. It will be faster and more reliable.

For a deeper guide on staying safe with screen automation, including enterprise considerations and incident response, see Computer Use Safety.

Browser Automation: for web-based tasks, the Chrome extension is faster and more reliable
Connecting Your Tools: connectors are always preferable to screen automation
Cowork: the environment where Computer Use runs
Batch Processing: for repetitive file-based tasks that do not require screen control
Scheduled Tasks: combine with Computer Use for recurring desktop automation (with caveats about reliability)

Something wrong or outdated? Let us know →

Get weekly workflows. Subscribe to the newsletter.