We’re excited to announce the Computer Controls API, a new way to give AI agents OS-level control of Kernel browsers.
Today, most AI agents rely on CDP to control browsers. CDP is great for many tasks, with a few caveats: it can’t reliably interact with some websites, and its interaction patterns are more likely to trigger bot detection.
That's why we built the Computer Controls API. Agents can now behave like real users, performing the same mouse and keyboard interactions that humans use:
- Move the mouse to on-screen coordinates
- Click at precise screen positions
- Type text as real keyboard input
- Scroll through pages
- Press keyboard shortcuts
- Capture screenshots of the UI
This gives vision-based agents a Kernel browser that works the way they do. Claude Computer Use was trained to navigate screens like a person would, and now it has a browser built for that. Anything that appears on the screen is interactive, no matter what’s happening under the hood.
Scrapybara fans, rejoice
Scrapybara was the first project to offer virtual machines for Claude Computer Use, but it was sunset on October 15. If you're looking for a replacement, we’re glad you found us!
Kernel offers very similar functionality to Scrapybara, with the addition of stealth mode, custom proxies, and session replays. Check out our migration guide for all the details.
With Computer Controls, your agent can extract data from visual charts, work with Figma canvases, or interact directly with the underlying VM environment—something CDP can’t do.
Get started with our docs, we can’t wait to see what you build!