What is an MCP server? It is a small program that lets an AI agent use a tool or a service. It tells the agent what it can do, such as “search my email” or “run an analytics report”, and does the work when the agent asks. MCP stands for Model Context Protocol, the open standard that both sides speak.
I built 20 MCP servers this month, one for each Google service I use, from Gmail to BigQuery. So this is the answer to “what is an MCP server” that I wish I had read first, with how it works and what to watch out for.
What an MCP server does, in one example
I type “how many people visited my site yesterday?” into Claude Code. Claude cannot see my Google Analytics account on its own. But my GA4 MCP server is installed, so Claude asks it what it can do and sees a tool called run_report. It calls that tool with yesterday’s date. The server signs in to Google with my account, gets the number and hands it back. Claude answers in a sentence.
That is the whole job. The AI model does the thinking. The MCP server gives it hands.
What MCP stands for and where it came from
MCP is the Model Context Protocol. Anthropic introduced it as an open standard on November 25, 2024, to fix a simple problem. Every AI app needed its own custom connection to every data source, so nothing scaled. With one shared protocol, a server written once works in every app that supports it. The specification is public, and its current version is dated July 28, 2026.
How an MCP server works
Three pieces take part in every conversation:
- The host is the AI app you use, such as Claude Code, Cursor or VS Code with Copilot.
- The client is a connector inside that app. It keeps one connection to one server.
- The server is the program that offers the capabilities.
They talk in JSON-RPC 2.0, a simple format for requests and answers. The app first asks the server for its list of tools. Each tool comes with a plain-language description and a schema of the inputs it needs. When the model decides a tool will help, the app sends a call with those inputs and gets the result back.
A server can run in two places. A local server runs on your own computer, and the app starts it and talks to it directly. A remote server runs on the web and is reached over HTTP. All 20 of mine are local, so your data goes from your machine to Google and nowhere else.
What an MCP server can offer
- Tools: actions the model can take, such as sending an email, adding a row to a sheet or pausing an ad campaign.
- Resources: data the model or you can read, such as a document or a file.
- Prompts: ready-made instructions you can pick from a menu.
Most servers you will meet are mostly tools. My Gmail server, for example, has tools to search, read, draft and send mail, plus labels and filters.
MCP server examples
To make it concrete, here is what a few of my servers let an agent do:
| MCP server | What you can ask |
|---|---|
| Gmail MCP | “Find the invoice from last week and draft a reply” |
| Google Calendar MCP | “When am I free on Thursday afternoon?” |
| Google Sheets MCP | “Add these three leads to the sales sheet” |
| GA4 MCP | “Which pages lost traffic this month?” |
| Search Console MCP | “Is this page indexed by Google?” |
Outside Google, there are MCP servers for databases, code hosting, browsers, design tools and much more. The idea is always the same. One server per service, written once, usable from any MCP app.
What is an MCP server compared with an API or a regular server?
An API is the door a service offers to programs. An MCP server usually stands in front of that door and translates for the AI model. My GA4 server, for example, calls Google’s Analytics Data API underneath. I wrote a full comparison in MCP vs API.
A regular web server waits for browsers or apps to ask for pages and data. An MCP server waits for an AI app, and the first thing it does is describe itself so the model knows how to use it. Many MCP servers are not “servers” in the usual sense at all. They are small programs that start when your AI app opens and stop when it closes.
Are MCP servers safe?
An MCP server can read your data and take actions, so treat it like any software you give access to your accounts. The specification is clear on this. It treats tools as code that runs on your behalf, and says the app must get your consent before calling one. In practice, your AI app shows the call and asks before it runs.
Three habits keep you safe:
- Install servers from sources you trust, the way you would choose a browser extension.
- Check which accounts and permissions a server asks for during setup.
- Prefer servers that ask before destructive actions. In mine, deleting or sending anything also needs an explicit confirmation.
How to start using one
Pick a service you already use every day and install its MCP server in the AI app you already have. Most of mine install from the VS Code Marketplace or Open VSX in one click and walk you through signing in to Google. For Claude Code or another app outside VS Code, each one also runs from npm with a single command. The full list is on my Google MCP servers page.
Common questions
What does MCP stand for?
Model Context Protocol. It is the open standard that AI apps and MCP servers use to talk to each other.
Why is an MCP server needed?
An AI model only knows what it was trained on and what you paste into the chat. An MCP server gives it live access to your tools and data, such as your calendar or your analytics, and lets it act on them with your approval.
What is the difference between an API and an MCP server?
An API is built for programs that a developer writes. An MCP server is built for AI models: it describes its own tools so the model can find and call them without custom code. Most MCP servers use an API underneath.
Does ChatGPT use MCP?
Yes. OpenAI’s help center says ChatGPT can connect to remote MCP servers (on Business, Enterprise and Edu plans as of November 2026). Claude, Cursor, VS Code Copilot and many other apps support MCP too.
Do I need to know how to code to use an MCP server?
No. Installing one is like installing an app, and you use it by asking questions in plain language. Building one takes programming.
Are MCP servers free?
Many are, including all 20 of mine. The service behind a server may still have its own costs or limits, such as API quotas.