Back to Blog

How to Promote Your MCP Server (So AI Agents Can Actually Find It)

You’ve built an MCP server. The endpoint responds. The handshake works. tools/list returns a beautiful collection of capabilities. Your terminal is full of reassuring green checkmarks.

There is just one tiny problem. Nobody knows your MCP server exists. Welcome to the next frontier of SEO 3.0 (ok, GEO if you like).

For years, we have helped search engines discover webpages. Then we started helping language models understand their content. Now, with AI agents learning to use tools, we face a new challenge. How do we make the capabilities behind a website discoverable? There are already plenty of surfaces where you can advertise your MCP (llms.txt, well-known, api directories and so on), however here we’re focusing on the so called on-page optimization. Can the humble webpage become the distribution channel for an MCP server?

At WordLift, we’ve been playing with this idea. Here’s what we’ve learned, along with a little experiment you can reproduce.

1. Your MCP server needs a front door

MCP makes APIs easier for AI assistants to use. But it doesn’t magically solve their distribution problem. Having a working endpoint is not the same as having a discoverable endpoint.

Take our free WordLift resolve() MCP end-point: https://wordlift.io/resolve/. It gives AI assistants access to a tool that identifies entities in text and links them to public Wikidata identifiers. Think Apple the company versus apple the fruit. Same word. Very different entity.

The MCP endpoint is public, read-only, and requires no authentication. That makes it a useful test subject. But a URL sitting in a README is not a particularly effective invitation. We need to make that URL visible, understandable, and easy to use.

2. Stop hiding your MCP behind a README

Imagine visiting a website and seeing a button: Add this capability to your AI assistant.

Not “Read our API documentation.” Not “Contact sales to discuss integration opportunities.” Just a way to connect a useful capability to your AI.

This is already possible in some environments. Claude supports a connector URL that can prefill its setup dialog. ChatGPT supports custom MCP connections in eligible accounts and workspaces, although its connection flow differs. Other applications, including Muse, have their own connector setup experiences.

There isn’t one universal deep link (yet) that guarantees installation everywhere. But here is what we can use.

How to create MCP connection links for Claude, Muse, and ChatGPT

Let’s take our Resolve MCP server as an example.

Claude: a prefilled connector link

Claude supports a URL that opens the custom connector dialog with the connector’s name and MCP address already filled in.

Here’s our actual example:

👉 Add WordLift Resolve to Claude

The pattern is straightforward:

const name = "WordLift Resolve";
const mcpUrl = "https://mcp.wordlift.io/resolve";

const claudeLink = new URL(
  "https://claude.ai/customize/connectors"
);

claudeLink.searchParams.set(
  "modal", "add-custom-connector"
);
claudeLink.searchParams.set(
  "connectorName", name
);
claudeLink.searchParams.set(
  "connectorUrl", mcpUrl
);

console.log(claudeLink.toString());

You can use the resulting URL in a standard HTML anchor. One click opens Claude’s setup flow. The user still reviews and approves the connection.

Not quite one-click installation, but considerably better than asking someone to copy configuration settings manually.

Muse: let the agent configure its own connector. Muse takes a different approach. Rather than relying on a documented URL template that prefills connector settings, we can give Muse the MCP address and ask it to create the connection and it will do it for us.

For example, a website could offer an Open in Muse link pointing to https://muse.ai/, alongside a copyable setup prompt:

Create a custom connector called WordLift Resolve using the remote MCP server at https://mcp.wordlift.io/resolve. It uses Streamable HTTP and requires no authentication. Discover its tools and confirm when the connection is ready.

The user opens Muse, pastes the prompt, and follows its connection flow. In other words, the link opens the assistant, while the prompt communicates how to connect. The important difference: Muse can help assemble the integration from instructions, rather than requiring a prefilled connector form.

ChatGPT: link to your approved plugin

For ChatGPT, assuming your plugin has been approved and published, the distribution mechanism is different again. You can link directly to its listing in the ChatGPT Plugin Directory.

The resulting URL typically looks like:

https://chatgpt.com/plugins/<approved-plugin-id>

Replace the placeholder with the actual URL of your published plugin.

Your website can then offer a button such as Install WordLift in ChatGPT.

The user opens the plugin listing, reviews its capabilities, and installs it, subject to their account and workspace permissions.

Unlike Claude, you don’t construct this public installation link by encoding an arbitrary MCP URL. You use the published plugin’s listing address.

If the plugin hasn’t been approved or published, use a link to the ChatGPT plugin directory or provide eligible users with custom MCP setup instructions instead.

We’ve created a test page that brings these connection paths together, using the same WordLift Resolve MCP server.

The interesting question isn’t whether the button gets clicked. It’s whether we can reduce the distance between discovering a capability and actually using it.

3. Give your MCP server a semantic identity

An MCP server has a URL. It may have a name, a description, and tools. How do we connect those capabilities to the website publishing them?

One approach worth experimenting with is Schema.org JSON-LD.

Schema.org already includes WebAPI, EntryPoint, and UseAction. Together, they can describe a service, what it does, and where its API can be invoked. We can use those concepts to make an MCP endpoint part of the website’s structured description.

Does adding JSON-LD guarantee that ChatGPT, Claude, or Google will discover your MCP server? Absolutely not. Does it constitute a new ranking factor? We have no evidence of that either. But it gives the capability an explicit, machine-readable description. And that is an interesting starting point.

Try it: generate Schema.org markup for your MCP server

Rather than asking you to write JSON-LD by hand, we built a small generator. We analyzed a proposal written on the official MCP repository that promotes the idea of leveraging JSON-LD to help AI crawlers find MCP application from web pages.

Paste the public HTTPS address of your Streamable HTTP MCP server, optionally add a name, provider, and description, and copy the generated markup into your website.

Free tool · No signup

Give your MCP server a name tag.

Paste a public Streamable HTTP MCP URL. We’ll write the Schema.org JSON-LD for your website. You can copy it, validate it, and publish it in your page’s HTML.

A small but important distinctionThis creates a semantic description of an MCP capability, not an MCP Server Card or an automatic installation link. Publishing the snippet in your site’s HTML does not force AI assistants to discover or connect to the server. Never include private URLs, credentials, or API keys.

You can validate the result with the Schema.org Validator.

A small but important distinction: this markup describes the API. It is not an MCP Server Card and does not automatically register the server with an AI assistant. Those are separate pieces of the puzzle.

4. Leave some breadcrumbs for the agents

Humans appreciate buttons. Machines prefer predictable discovery mechanisms.

This is where things are getting interesting in the MCP ecosystem. The emerging MCP Server Card work describes how a server can publish its identity and connection details. An AI Catalog at /.well-known/ai-catalog.json can advertise those cards at the domain level. And the official MCP Registry provides another route for publishing public servers to the ecosystem.

These mechanisms serve different purposes. A webpage explains why someone might want the capability. Structured data describes it semantically. A discovery document tells compatible clients where to find connection metadata. MCP handles the actual interaction.

We don’t yet have universal automatic discovery across AI assistants. But this is starting to look like a familiar pattern. The Semantic Web gave us identifiers and descriptions. The agentic web is adding executable capabilities. The challenge is to connect the two without forcing every agent to guess what a service does.

5. Don’t just advertise your MCP. Test whether it works.

For the WordLift experiment, we’re making the process observable.

You can inspect the MCP connection, discover the exposed tools, and send a short passage to resolve_demo.

Under the hood, we’re testing three important operations:

  • initialize: Can the client establish an MCP protocol session?
  • tools/list: Can it discover the available capabilities?
  • tools/call: Can it invoke a tool and interpret the result?

Try a passage mentioning Apple, Cambridge, or another ambiguous name. Resolve() should help ground the mentions to specific identities, or report that a mention could not be resolved.

The point isn’t to ask an AI to sound confident. It’s to give it an opportunity to use evidence.

And the same principle applies well beyond entity resolution. Imagine a merchant making its product catalogue searchable through MCP. Or a publisher exposing licensed content. Or a company allowing agents to ask questions against its corporate knowledge graph.

Those capabilities deserve to be discoverable, understandable, and usable.

From ranking webpages to discovering capabilities

I suspect the SEO 3.0 & GEO conversation is about to get more interesting. Today we ask whether AI systems mention our brands, cite our content, or recommend our products. Tomorrow we may also ask whether agents can discover and use the capabilities we expose.

Not every interaction needs to end with a webpage visit. Sometimes the most valuable outcome is a grounded answer, a reliable API call, or a transaction completed with the user’s permission. That doesn’t make traditional SEO irrelevant. Quite the opposite. It expands the meaning of discoverability.

And for those of us who have spent years working on linked data, knowledge graphs, and ontologies, there’s something beautifully familiar about this transition. We wanted machines to understand what things mean. Now we want them to act on that meaning.

An MCP server without discoverability is just another well-kept secret on the internet. Let’s see if we can change that.

Try the experiment, publish the markup for your own MCP server, and tell us what happens.

After all, the web was always meant to be connected. Maybe it’s time we started connecting its capabilities, too.