Your Website as an MCP Service

Putting your website behind MCP means a model can search your content and read it back to someone in chat, instead of handing them a link and hoping they click it.

This site runs one. Everything below is what it exposes and why, so you can copy the shape rather than the theory. The server side of it covers the protocol itself.

What To Expose First

Start with whatever your support inbox answers most: pricing, what the product does, and how to begin. Those are the questions already being asked, so coverage there pays back immediately.

Leave out anything that moves faster than you can sync it, and anything you would not publish on a page. An MCP server is a second front door to the same content, not a private one.

Build one area completely before adding the next. A half-covered topic is worse than an absent one, because the model answers confidently from the half it has.

Search Then Fetch

The pattern that works is two tools, not a list of pages. One searches and returns short previews; the other takes an id and returns the whole item.

This site's server does exactly that with content_search and content_get_full. Search returns ranked previews, and content_get_full takes a page_id and a type of either article or faq. Requiring both means the model cannot guess its way into a fetch.

Page the search results. Ours defaults to five per call and allows up to a hundred, because a model with a small context window drowns in more and a large one wants them in a single pass.

Write Descriptions The Model Reads

A tool description is not documentation for a developer. It is the only thing the model sees when choosing what to call, so it has to say what the tool returns and what to do next.

Here is the live description from the search tool on this site:

Search articles and FAQ content for relevant SEO and documentation
information. Returns lightweight previews ranked by relevance; use
content_get_full to read a result in full. Results are paginated
(default 5 per call) so low-context clients stay lean - raise "limit"
to pull more in one pass, or page with "offset".

It names the next tool by name. That one sentence is the difference between a model that searches and stops, and one that searches and then reads the thing it found.

Put the same care into each parameter. Our limit field carries its own note about small and large context models, because the model picking a value is the one that has to live with it.

Keep Responses Whole

Someone in a chat window wants the answer, not a redirect. If they ask what a plan includes, the tool should return the features and the price, not a sentence pointing at the pricing page.

Save URLs for the things a person has to leave the conversation to do: signing up, buying, downloading. A link is an action, not an answer.

Draw The Line At Actions

Reading is safe to automate. Doing something is not, and the boundary belongs in the tool list itself rather than in a policy document nobody reads.

Of the seven tools on this site's public server, six only read. The seventh creates an account from an email address the user supplies, and a confirmation email to that address is what makes the account real.

That is the whole rule. Let the model answer anything, let it do one clearly labelled thing, and route the rest to a human through a general channel rather than somebody's direct line.

Keep It In Sync

MCP content goes stale the same way a page does, and a wrong price quoted confidently in chat costs more trust than a wrong price on a page nobody read.

Update both at once or drive both from the same source. Retire discontinued products, expired offers and dead processes in the same pass.

Try It On This Site

The public endpoint is https://seolinkmap.com/mcp and it needs no account. Connect it to your LLM and its tools appear in your client.

Ask it about pricing, or about this article, and watch which tool it reaches for. The search-then-fetch chain is visible from the outside, which is the fastest way to judge whether your own descriptions are doing their job.

Every SEOLinkMap account also gets a private server for conversational access to its own SEO data, behind OAuth, built the same way: the same search-and-fetch pair for the clients that require it, plus one tool per dataset, with the line between reading and doing drawn in the tool list.