In brief, an agent involves your web site, discovers what WebMCP instruments you provide, reads what every one does, and calls the one that matches the duty.
There are two methods to set this up: HTML for types or JavaScript for extra advanced makes use of.
Keep in mind, these instruments solely work whereas the web page is open; there’s no headless or background mode, and the browser has to maintain working.
Declarative API (HTML)
The declarative API is the better implementation, meant for actions that already stay in a type. For instance, a reservation, signup, or contact request.
To implement this, you merely want so as to add a few attributes to the shape you have already got:
- toolname provides the instrument a reputation
- tooldescription explains what it does
- toolparamdescription, elective, and tells the agent what that discipline expects
From there, the browser turns the shape right into a instrument. By default, solely your individual website can use it.
To see the declarative API in motion, I loaded Google’s Le Petit Bistro demo with the Mannequin Context Device Inspector open. It’s a Chrome extension that lists the instruments a web page has registered and allows you to enter a immediate and watch its Gemini-powered check agent name them.
Only a heads up, you’ll want a Gemini API key for this (you need to use the free tier).
The demo is a restaurant reservation web page with one reserving type. The fields embody identify, cellphone quantity, date and time, and so forth.
Certain sufficient, the shape confirmed up within the Inspector as a registered instrument, full with its schema. The panel reveals precisely which WebMCP instruments an agent can see on a website.
