Claude in Chrome & Computer Use
Claude can do more than talk — it can operate software: read and act on a web page in the browser, or control a computer (move the cursor, click, type) to complete tasks across apps. Powerful, and worth understanding the guardrails.
Two related capabilities
- Claude in Chrome — a browser-side agent that can read the current page and take actions in it (navigate, click, fill forms) on your behalf.
- Computer use — more general control of a desktop (screenshots + mouse/keyboard) to operate native apps and multi-app workflows.
Use them for things that are otherwise tedious: pulling info across pages, filling repetitive forms, multi-step web tasks.
The safety model (read this first)
Letting an AI click and type on your behalf is exactly where prompt injection gets dangerous: a malicious web page can contain instructions that try to hijack the agent.
:::warning Treat links and pages as untrusted
- Be present and supervise for anything consequential — don't walk away mid-task.
- Don't point it at sensitive accounts (banking, admin) casually.
- Verify destinations — an agent shouldn't follow a suspicious link or act on instructions hidden in a page.
- Least privilege — restricted/permissioned modes exist precisely so the agent can look but not act where it shouldn't. :::
Good vs risky tasks
| Good fit | Be careful |
|---|---|
| Research across pages, read-only | Anything that spends money |
| Filling a known, benign form | Logging into sensitive accounts |
| Extracting info you supervise | Acting on emails/pages from strangers |