WebMCP is a proposal for a new web capability that lets AI agents execute actions on a website directly. It’s still in the early stages and currently only works in Chrome.
What does this mean?
AI chatbots like ChatGPT can already go through a website in a browser, click things and perform actions. But it isn’t smooth and it doesn’t always work. While an agent is clicking through a site, a pop-up might appear, or the page might shift because an element hasn’t finished loading. Pages can also be long and heavy, which makes it harder for an agent to find what it needs.
WebMCP aims to fix that by giving agents clear instructions for a set of pre-defined actions like filling a form, adding a product to a cart, filtering through products, booking a ticket and so on. Instead of working out how your site behaves, the agent is told. That makes the action smoother and quicker.
How would this work?
Think of a simple contact form. Right now an AI agent fills it the way a person does. It finds the page, works out what each field is for, types into each one, then submits.
With WebMCP you add a few extra tags to the form that tell the browser what the form is for and what each field expects. The agent asks the user for exactly the information the form needs, then the browser jumps to your form and fills it in correctly, on screen where the user can see it. The user usually still clicks Submit, or the control of that can be given to the agent as well .
Plain HTML form
<form action="/submit">
<label for="firstName">First Name</label>
<input type="text" name="firstName" id="firstName">
<label for="lastName">Last Name</label>
<input type="text" name="lastName" id="lastName">
<button type="submit">Submit</button>
</form>HTML form with WebMCP
<form toolname="createSupportRequest"
tooldescription="Submits a request for customer support."
action="/submit">
<label for="firstName">First Name</label>
<input type="text" name="firstName" id="firstName">
<label for="lastName">Last Name</label>
<input type="text" name="lastName" id="lastName">
<button type="submit">Submit</button>
</form>For things that aren’t forms, like filtering hundred products by different sizes, colors, etc., you can write a small JavaScript function instead and give it a name and a description. The agent calls that function directly and gets clean results back.
This doesn’t mean agents will act the way you want them to
You can pre-define how a form on your website should be filled, but there’s no guarantee an agent will use it. Just as a person can do whatever they like on your site, so can an agent. It might ignore WebMCP entirely and click through the form fields instead.
You can tighten control over the actions that matter most, but you can’t force agents to take that route. And right now, most don’t. Browsers are only just adding support, and the major AI agents aren’t calling these tools yet.
Where is this really useful?
Website with long forms with a lot of fields, multi-step forms, hundreds of products or services behind multiple filters, complex booking/reservations processes, etc. These are all actions you can describe for an agent using WebMCP, and they’re the ones where an agent clicking its way through is most likely to go wrong.
Does this matter to you?
For most websites, not yet. If your site is a handful of pages where people read something and maybe send an enquiry, an agent can already handle that well enough, and there’s nothing here worth acting on today. The sites that should be paying attention are the ones where something actually gets completed on the page such as booking systems, e-commerce checkouts, large catalogues with filters, multi-step applications, customer portals and internal tools. The more steps between arriving and finishing, the more there is for an agent to get wrong, and the more WebMCP has to offer.
Even then, this is worth watching rather than building right now. The agents that would use it haven’t shipped support yet fully, so implementing today buys you readiness, not traffic. The useful thing to do in the meantime is to notice which parts of your site would be hardest for someone to complete without seeing the screen. Those are the parts you’d describe first when this becomes real.