What Is an MCP Server? A Plain-English Guide with Examples

An MCP server is a small program that gives an AI app a new ability, such as reading your database or checking Google Trends. The AI app connects to it using the Model Context Protocol (MCP), asks what it can do, and calls those actions when your request needs one. You do not need to code to use one.

What is an MCP server, in plain terms?

Claude and ChatGPT know a lot, but they cannot see your data. They cannot open your database, read your Search Console, or check your App Store reviews. An MCP server fixes that. It sits between the AI and one tool or data source, and it hands the AI a short menu of things it is allowed to do.

Think of a waiter. You (through the AI) say what you want. The waiter (the server) knows what the kitchen (your database, your app, your inbox) can make, and brings it back. The AI never walks into the kitchen itself. Each server is one connector, and you can use several at once.

In AI products, an MCP server is the piece that turns a chat assistant into something that can act on your own systems. Without servers, the AI only answers from what it already knows. With them, it can fetch live numbers, check a page, or send a message. You stay in the conversation the whole time.

The protocol behind this is open, so any AI app that speaks MCP can use any MCP server. For the bigger picture, read What is MCP?. This page stays on the server side.

What does an MCP server do? A concrete example

For example, say you run a small app and keep your users in a PostgreSQL database. You connect a Postgres MCP server to Claude and then type:

How many people signed up last week, and how does that compare with the week before?

Here is what happens behind the scenes:

1. Claude asks the server which tools it offers. The server lists them, for example one that describes tables and one that runs a read-only query.
2. Claude looks at your table structure through the describe tool.
3. Claude writes a query that counts rows by signup date and sends it through the query tool.
4. The server runs it against your database and returns the rows.
5. Claude turns the rows into a plain answer: "412 signups last week, up 9% from 378 the week before."

You typed one sentence. The server did the database work, and Claude did the thinking and the explaining. The same pattern applies to every other server: ask in plain English, the AI picks the right action, the server carries it out.

What does an MCP server expose: tools, resources and prompts?

The MCP documentation describes three things a server can offer:

• Tools are actions the AI can run, like "run a query" or "inspect this URL". They are the part most servers are built around. See MCP tools for how they are described and called.
• Resources are data the AI can read, such as a file, a table schema or a document.
• Prompts are ready-made instructions the server provides, which you can pick to start a task a certain way.

Most servers you meet in practice are mostly tools. Resources and prompts are less common, and many AI apps use tools first. When someone says "this server has 19 tools", they mean 19 separate actions on that menu.

Server, client and host: who is who?

Three roles show up in every setup, and they are easy to confuse:

• The host is the app you talk to, such as Claude, ChatGPT, Claude Code or Cursor.
• The client is the connector inside that app. It holds one connection to one server. See MCP client.
• The server is the program on the other end that does the work.

The architecture overview lays out the same split. It helps when something breaks: a failed tool call usually points at one server, not at the whole app.

Are MCP servers local or remote?

A server can run in two places. A local server runs on your own computer, started by your AI app. A remote server runs somewhere else and your app connects to it by URL. The remote server guide covers the connection side, and remote MCP server goes deeper.

A third option sits on top of remote: a hosted server, where a company runs the real server for you and you only enter your credentials. Here is how the three compare:

TypeWhere it runsSetupWho maintains itWorks from web apps

Local

Your computer

Install Node, Python or Docker, edit a config file

You

No, desktop and terminal apps only

Remote (self-run)

A server you rent

Deploy it, secure it, keep it online

You

Yes

Hosted

Run for you

Pick a server, save credentials, paste a URL

The host

Yes

If a local setup has ever sent you into a terminal, that is the reason hosted servers exist. Local servers stop working when your laptop sleeps. They cannot be reached from a browser-based AI app at all.

What are examples of MCP servers, and how do I find them?

Thousands of servers exist, most of them open source. The official examples page and the servers repository list reference implementations. A few well-known kinds:

• PostgreSQL lists tables, describes columns and runs queries, so you can ask questions about your data in plain English.
• GitHub reads and manages repositories, issues and pull requests.
• Playwright drives a real browser so an AI can click through a site and test it.
• Filesystem reads and writes files in folders you allow.
• Google Trends shows interest over time, compares search terms and lists trending topics by region.
• Google Search Console reads performance, inspects URLs and checks indexing and sitemaps.
• App Store Connect reaches apps, versions, TestFlight, reviews and sales reports.
• WhatsApp reads chats and contacts and can send messages.

For a ranked and explained list, see best MCP servers.

To find more, start with the official servers repository. Vendors such as GitHub publish their own servers, and marketplaces gather servers in one place so you do not hunt through repositories. The MCPifex marketplace is one example. Before you add any server, check what it can reach, which tools it exposes and who maintains it.

How do I use an MCP server without installing anything?

Use a hosted one. With MCPifex the steps are:

1. Create a free account and pick a server from the marketplace, for example PostgreSQL or Google Trends.
2. Save your own credentials for that server, if it needs any.
3. Choose which tools the AI may use.
4. Copy the endpoint URL and API key into Claude, ChatGPT, Claude Code or Cursor.

MCPifex runs the real server package for you, so nothing is installed on your machine. One endpoint covers every server you add. The Free plan includes 3 instances with no card, and gateway calls are unlimited on every plan.

Are MCP servers safe to use?

An MCP server can only do what its tools allow, so safety comes down to which tools you turn on. Read tools, like "run a SELECT query", are low risk. Write tools, like "execute any SQL" or "delete a WhatsApp message", can change or delete real things.

Good habits:

• Start with read-only tools and add write tools only when you need them.
• Give the server the narrowest credentials that work, such as a read-only database user.
• Check the call log to see what the AI actually did.
• Trust servers from sources you can verify, and remember that text inside your data can try to steer the AI.

With a hosted gateway such as MCPifex, write and destructive tools start off (WhatsApp's send tools are an exception: they start on). For PostgreSQL the write tool (execute) is off by default, and you turn on only what you want.

← Back to glossary