Giving AI Agents a Proper Interface to the Web

In WebMCP, Alexandra Klepper introduces a proposed web standard that lets websites expose structured tools directly to AI agents.

“Instead of an agent reviewing the element, the website declares the element’s purpose.”

Today, browser agents often rely on screenshots, the DOM, accessibility trees, and simulated clicks. That is slow and brittle. A button moves, a label changes, or the agent misunderstands a field, and the entire workflow can fail.

WebMCP would let a site explicitly provide tools such as checkout, filter_results, submit_application, or run_diagnostics. Developers can register them through JavaScript or annotate ordinary HTML forms, giving agents structured parameters and clear descriptions instead of making them guess how the interface works.

The especially interesting part is that these tools operate inside the open browser tab. They can use the application’s existing state and authentication while keeping the interface visible to the user. That is much more cooperative than an agent bypassing the website through a separate backend integration.

It also creates a powerful new security boundary. An agent may be acting inside an authenticated session, and malicious tool descriptions or contaminated responses could attempt to manipulate it. Chrome’s own guidance acknowledges that prompt-injection safety cannot be guaranteed and recommends origin isolation, strict permissions, limited tool exposure, confirmation for sensitive actions, and explicit labels for untrusted content.

WebMCP is still experimental and is currently available through a Chrome 149 origin trial. It is not ready to depend on across browsers yet.

Still, this feels like an important direction. If agents are going to use the web, I would rather websites provide explicit, inspectable actions than have models blindly click around interfaces built only for humans.

The same standard could make browser agents dramatically more useful. It could also give a compromised agent more reliable access to everything we can do while signed in.

Read Chrome’s introduction to WebMCP and the evolving WebMCP specification.