Glossary
What is Model Context Protocol (MCP)?
Definition
Model Context Protocol (MCP) is an open standard that lets an AI assistant call external tools and read external data through a defined interface. Unlike a website, which an assistant reads by crawling, an MCP server is connected on purpose by a user or a developer. It is a way to be used, not a way to be found.
At a glance
- CategoriesTechnical
- Related fieldsTechnical
- DifficultyAdvanced
At a glance
- MCP is an open standard with a published specification and reference implementations.
- An MCP server exposes tools the model can call, resources it can read and prompt templates it can use.
- Nothing discovers a server by crawling. Somebody connects it deliberately.
- MCP changes what an assistant can do with your product, not whether it mentions your brand.
Server, client, and what gets exposed
Three parts. A host application such as Claude or a code editor, a client inside it managing the connection, and a server that you run.
Your server declares what it offers. The specification sets out the categories: tools the model can call, resources it can read, and prompts it can reuse. A retailer’s server might expose a tool that checks stock at a named store and a resource holding the returns policy.
The assistant guesses at none of this. It asks the server what exists, reads the descriptions, and calls whatever fits the request in front of it. The descriptions you write are the interface.
Transport is plumbing, a local process or an HTTP endpoint. What matters is the contract you publish and the permissions the user granted when they connected.
Being connectable is not the same as being discoverable
MCP has started turning up in AI visibility pitches. It answers a different question.
Crawling is how a brand gets found. Somebody asks an open question, the system retrieves pages, and your content is in the pool or it is not. Crawlability and extractability govern that end to end.
MCP begins after the user has already chosen you. They connect your server on purpose, and from that point the assistant can act against your systems: look up an order, check availability, make a booking.
Worth building if you have a product people use repeatedly and a real API behind it. Worth nothing if the goal is being named when a stranger asks which brands make a thing, because nobody discovers a server by asking a question.
Decide whether you need a server at all
Work through these in order. Most teams stop at step two.
- Name the task a user would want done, not read about. Check an order, reschedule a delivery, pull a report. If every answer is informational, the work belongs on your website.
- Confirm you have an API that already does it. An MCP server is a wrapper with good descriptions. With no API underneath, this is an API project wearing a hat.
- Write the tool descriptions before the code. A model picks a tool by reading its description, which makes that text a product surface rather than documentation.
- Scope permissions to the task. People connect servers into assistants with broad session access, so least privilege matters more here than usual.
- Ship it to your own team first. Internal use exposes vague tool descriptions faster than any spec review.
Treat the result as one piece of agent experience rather than a marketing channel.
A website vs an MCP server
| Your website | An MCP server | |
|---|---|---|
| How it is reached | Crawled and retrieved | Connected on purpose by a user |
| What it provides | Text an assistant can quote | Actions an assistant can take |
| Who it serves | Anyone asking about your category | Existing customers doing a task |
| Who owns it | Marketing and content | Engineering and product |
| Helps you get cited | Yes | No |
Frequently asked questions
-
Is MCP an AI visibility tool?
No. MCP defines how an assistant connects to tools and data a user has deliberately given it. Visibility is about retrieval: whether your pages are found and quoted when somebody asks a question. Building a server does not put your brand into a single additional answer.
-
Do I need an MCP server to appear in ChatGPT?
No. Appearing in an assistant’s answers depends on your pages being crawlable, readable and worth quoting. An MCP server is a separate capability for users who connect it, and most brands will never need one.
-
Is MCP a standard or a proposal?
A published open standard with a versioned specification, reference implementations and adoption across several assistant products. That puts it well ahead of conventions like llms.txt. It also means the spec moves, so build against a version and expect to keep up.
-
Who should actually build one?
Companies whose product is a system people operate: logistics, banking, ticketing, project tools, anything with an account and repeat tasks. If your relationship with a customer ends at the purchase, a server gives them nothing to call.