Three ways for an agent to use a website

Read all the HTML

<div class="hdr-wrap">
<button class="btn-primary">?
<button class="btn-ghost">?
<div class="col-8 grid">
<div class="nl-form">…

Guesses

Which button adds to the cart and which one sends the newsletter? The class names say nothing.

Take screenshots

click(412, 288)?

Slow and fragile

Clicks on coordinates like a human – only worse, and wrong when the interface changes.
Call tools
- WebMCP · tools
- search_products
- compare_products
- add_to_cart
The page tells it

The tools have a name, a description and a schema. The agent calls the function instead of searching.

Three ways for an agent to use a website. Reading the HTML and guessing which button does what. Taking screenshots and clicking on coordinates. Or using tools that the page itself describes with WebMCP, such as search_products, compare_products and add_to_cart.