For twenty years a website had one kind of visitor: a person with a browser. It now has a second kind. An AI agent, acting for a person, wants to search your site, ask it questions and, increasingly, take an action on it: book, order, check a policy. The Model Context Protocol (MCP) is how most of those agents will do it, and in the first week of October 2026 our founder went looking for every service, platform and open-source project that puts an MCP server in front of a website. He found 44 tools and two standards.
The full comparison, with every tool checked against its own documentation on 7 October 2026, is on his site. This note is the business reading of it.
The question that sorts everything: who runs the server?
An MCP server for a website can be run by four different parties, and that choice decides what an agent can see and who controls it.
| Who runs it | What it means for you |
|---|---|
| The platform | Wix, Shopify, GitBook and Mintlify switch one on for every site they host. No setup, but only for sites hosted there. |
| A hosted service | You point a service such as Cloudflare AI Search, CustomGPT.ai, SiteSpeakAI, Kapa.ai or Azure AI Search at your site and it hosts the server. Each ties you to its own stack. |
| You, the site owner | You run open-source software, a CMS plugin or a docs framework, and you decide what it exposes. |
| The agent’s user | Someone crawls your site into their own index and serves it to their own agent. You are not involved, and you do not see what is asked. |
Only the last one leaves the owner out entirely. It is also the one that happens by default when you have done nothing.
You may already have one
Every Wix site has an MCP endpoint at a fixed address that any visitor’s agent can use without authentication: search, business details, and actions such as booking an appointment. Every Shopify store answers policy questions on one endpoint and moved product search and cart actions to a newer one in 2026. Every published GitBook and Mintlify site gets a read-only docs server.
If your site is on one of these platforms, the first question is not whether to build an MCP server. It is what the one you already have is saying about you, and whether that is what you want.
Most of the owner-run options are not for your visitors
Of the 44 tools, 24 are in the three groups the site owner runs. Fourteen of those serve documentation sites or a CMS’s own editorial team: they are editing tools behind a login, built so the site’s team can use an agent, not so a visitor’s agent can use the site. The project built as a public “ask this site” endpoint that the owner publishes is NLWeb, an MIT-licensed protocol led at Microsoft that builds on the Schema.org data sites already publish. Retrieval platforms and search engines can also serve a site over MCP once you have crawled it into them.
Discovery has only just started
For an agent to use your server, it has to find it. MCP Server Cards became a final proposal on 6 October 2026, and WebMCP is in a Chrome trial. For now, someone usually still has to give the agent the address. That will change, and when it does, sites with a well-designed server will be found and the rest will be crawled.
The layer that is still missing
The survey ends with a gap. Nothing established today works on any platform, is run by the site owner, keeps content in sync, lets the owner choose what is exposed, shows what agents ask, and publishes the discovery files. One closed beta pitches it. For most businesses that means the honest options in October 2026 are a platform’s built-in server, a hosted service with its trade-offs, or a server built for the purpose.
What we recommend
- Find out what you already expose. If you are on Wix, Shopify, GitBook or Mintlify, read what the built-in server returns before anyone else does.
- Decide who should run it. If the answers an agent gives about your business matter, the owner-run or purpose-built options are the ones where you hold the controls: what is exposed, who is authenticated, what is logged.
- Start from the questions, not the API. A good server has a few tools that answer what people actually ask an agent, with compact results. That is the design work in our MCP server service.
- Watch discovery. Server Cards and WebMCP are the signals that agents will start finding sites on their own.
The full research, with all 44 tools and the two standards, is at the link above. Our view of whether a product should have an MCP server at all is in a separate note.