MCP server (Model Context Protocol)
An MCP server lets AI assistants query your data directly through an open standard, the Model Context Protocol, instead of guessing from your web pages.
An MCP server is a small piece of software that sits in front of your data and answers questions from AI assistants. MCP stands for Model Context Protocol. Anthropic released it as an open standard in late 2024, and other major AI companies, OpenAI among them, have since added support. Because it is a shared standard, one server works with any assistant that speaks it.
Without one, an assistant that wants to know about your business reads your web pages the way a person would, then guesses. It might miss a product, mix up two details, or quote a page that is years old. With one, it can ask a direct question, such as which products fit a need, and get a structured answer back with a working link.
Why it matters to you
More buyers now start with an AI assistant instead of a search box. They ask for a recommendation, and the assistant goes looking. If the only way it can learn about you is by reading pages built for people, you are relying on it to read carefully. An MCP server removes that guesswork for the information you choose to expose.
It sits next to the other AI visibility work, not instead of it. llms.txt gives an assistant a summary. Structured data labels the facts on each page. An MCP server lets an assistant query the source directly, which matters most when you have a lot of structured information, like a product catalogue.
How I set it up
I start read only, over data that is already public on the site, and I build it from the same system of record as the pages, so the server and the site do not drift apart. I keep the set of tools small, such as searching the catalogue and fetching one product, and every answer carries a real link back to the page.
Then I make it findable. The server I built for virginiamitchell.ca publishes a signed discovery record, so an agent can locate it without being told where to look. Anything that would take an action or touch private data is a separate build, with authentication designed in from the start.
What it looks like in practice
virginiamitchell.ca has a 128 product catalogue with a live MCP server beside it. When I called that server during an audit, it returned 128 products and 186 pages.
What I cannot tell you yet is how much it changes how often assistants recommend that catalogue. I have not measured a before and after for any client, and I would rather say so than sell you a number I do not have. What it does guarantee is narrower and still useful: when an assistant does ask, it gets the right answer.
Questions I get about this
- Is an MCP server the same thing as an API?
- It is built on the same idea. An API is a way for software to ask your system for data. MCP is a shared standard for describing that access, so any AI assistant that speaks MCP can discover what your server offers and use it without custom code written for each one.
- Does my business need one?
- Most do not, yet. It earns its keep when you have structured information people ask about, such as a product catalogue, a service list, inventory or locations, and your buyers are starting to ask AI assistants before they search. A simple brochure site gets more from getting its facts right first.
- Is it safe to connect my data to AI assistants?
- Start read only, and expose only what is already public. A catalogue server that answers questions about products you already publish carries little risk. Anything that takes an action, touches customer data or needs a login is a different build, with authentication and limits designed in from the start.
Where this fits in our work
Want this set up properly for your business?
This is the kind of thing I build every week. Grab a time and we will talk through what fits.