What Is WebMCP? AI Agents Just Got a Proper Front Door to Your Website

When an AI assistant visits your website today, it does so without your permission and without a map.

It reads your HTML the way someone reads a wall of text — looking for patterns, inferring meaning, trying to guess what you do, where you are, and how to contact you from a page that was designed for human eyes. If there is a contact form, the AI must deduce that the “Name” field is for a name, the “Email” field is for an email, and the “Submit” button is how you send it. If your services are described in a paragraph rather than a structured list, the AI has to parse that paragraph and hope its interpretation is accurate.

This is how AI agents work today. It is effective enough for simple cases, but it is imprecise, error-prone, and completely uncontrolled from the website owner’s perspective. The AI decides what information it extracts. The website owner has no say.

WebMCP changes this. It gives the website owner a way to declare — in precise, machine-readable terms — exactly what information their site contains and exactly what actions an AI agent is allowed to take. Not by scraping. Not by guessing. Through a structured interface that the site itself provides.

The front door analogy

Think of the current state as a house with no front door. Visitors can get in — through windows, through gaps in the walls, by peering through the letterbox — but the owner has no control over how they enter or what they see. WebMCP is the front door: a controlled, deliberate entry point that the owner designed and that visitors are expected to use.

For a business website, WebMCP lets you say to an AI agent: “Here is what we do. Here is where we operate. Here is how you book an appointment. Here is how you check if we cover your area.” These are not page titles and paragraphs for the AI to interpret — they are named functions the AI can call directly.

The technical term for these functions is “tools.” A tool in the WebMCP sense is a piece of functionality your website exposes: get the business’s contact details, check availability for a date, search the services catalogue, submit a consultation request. When a browser-based AI agent asks your website what it can do, your website responds with its list of tools.

The agent can then use those tools to answer the user’s question accurately and completely — not from a guess at what your site’s text means, but from the structured data your site provides.

Who built this and how real is it

WebMCP is not a startup product or a vendor proposal. It is a draft standard being developed through the W3C Web Machine Learning Community Group — the same body that governs web standards for machine learning in browsers — and it was co-authored by engineers from Google and Microsoft.

Google announced it at their developer conference in May 2026 and opened it for testing in Chrome, the browser used by most web users worldwide. The standard is currently in what is called an “origin trial” — a structured testing period during which developers can implement it and provide feedback before it becomes a permanent part of the web.

This is not a feature that exists only in experimental builds. It is available in Chrome right now, for anyone who chooses to implement it.

The honest picture on who is using it

Here is where we will be direct with you, because the coverage of WebMCP is not always honest about this.

Very few websites are implementing WebMCP today — independent crawls of the top 200,000 websites found essentially none in production as of July 2026. Brands like Shopify and Expedia were cited at Google’s conference as early experimenters, but “experimenter” describes intent and demo participation, not live production systems.

More importantly: no mainstream AI assistant is currently calling these tools on live websites. Not ChatGPT, not Claude, not Gemini. They are all still operating through the old method — DOM parsing, screenshots, guessing.

The reason to care now is not that it is working today. It is that:

1. Google and Microsoft co-authored the standard, which is about as strong a guarantee of forward momentum as exists in web technology
2. The cost of implementing it is low — it is a relatively small amount of structured configuration
3. The first businesses to have their tools ready are the ones AI agents will be able to use properly when the consuming agent software ships
4. Google’s own Lighthouse testing tool now checks whether your site has WebMCP registered, which is a signal about the direction Google intends to take this

What this means for a KZN service business

Consider what a fully-capable AI agent would do when a user asks: “Can you find me a reliable plumber in Umhlanga who handles blocked drains and is available this week?”

Today, the agent searches the web, reads whatever pages it can find, and synthesises an answer from ambiguous text. It might get the right businesses. It might get their phone numbers wrong. It might describe a plumber who is out of business. It has no way to check availability.

With WebMCP tools registered on the right websites, the agent calls searchBusinesses({ category: 'plumber', area: 'Umhlanga', service: 'blocked drains' }), gets back a structured list of matching businesses with their accurate contact details, and calls checkAvailability({ week: 'current' }) on the top result to confirm they have capacity.

The plumbing company whose website has those tools registered answers the query with precision. The one that doesn’t is described from whatever the agent could scrape.

That is the competitive gap this technology creates — and it is the reason we implement it on every TVS site while it is still early, while the infrastructure cost is low, and while most competitors have not considered it yet.

---

TVS implements WebMCP tool registration, llms.txt, and structured schema data on every site we build. If you want to know whether your current website is AI-ready — or how far behind it is — run a visibility check or speak to us.