Computer Use beta

Eyes and hands.

A capability where Claude looks at screenshots and returns mouse + keyboard actions. You run a loop: capture the screen, send to Claude, execute the actions Claude returns, repeat until done. Useful for automating apps with no API.

Where it earns its keep.

• Filling browser forms across sites that block scraping.
• Driving legacy desktop software during a migration.
• QA — visiting a list of pages and reporting what's broken.
• Bulk-processing tasks in apps with no batch UI.

Sandbox first.

1. Run it inside a VM or container. Anthropic publishes a Docker reference image. Don't point it at your daily-driver desktop.
2. Pull the reference container from ghcr.io/anthropics/anthropic-quickstarts.
3. Set ANTHROPIC_API_KEY and start the container; it serves a web UI you chat with.
4. Give a small task first — "open Firefox, search for cat pictures" — to feel it out.

What works.

• Discrete tasks with clear endings. "Open this PDF, copy the table to a CSV, save and quit."
• Verify checkpoints. After each subtask, ask Claude to confirm what it sees on the screen before continuing.
• Snapshot on failure. When the loop bails, save the last screenshot. Most "why did it stop" questions answer themselves with the picture.
• Cap step count. A runaway computer-use loop can rack up tokens fast.

Don't reach for it first.

If the target system has an API, an MCP server, or a Playwright/Puppeteer wrapper, those are cheaper, faster, and more reliable than Computer Use. Reach for Computer Use when the only available interface is a GUI.

Related products.

Every Anthropic surface has its own page on this guide.