What Is MCP? A Beginner's Guide
This page is the short explanation, written for someone who has never hooked an assistant up to another app. The picture below follows the official introduction, What is the Model Context Protocol (MCP)? The names for the app, the connection, and the three things a connection can offer come from the architecture overview.
The missing calendar
You ask an assistant, "What is on my calendar tomorrow?" It can write a smooth answer from the words in the chat. Your calendar lives in another app. Until something joins the two, the answer is a guess.
The same gap shows up with notes, a company database, a design file, or a search box. Each one is its own system. For a while, every AI app invented its own plug for each system. A plug written for one assistant stayed inside that assistant.
MCP is the shared plug. Write the connection once, to the standard, and any AI app that speaks MCP can use it.
What the connection is for
The introduction names three things an AI app can reach through MCP.
- Data sources. Local files, a database, a page of notes.
- Tools. A search engine, a calculator, or a command that creates a page or sends a file onward.
- Workflows. A specialized prompt: a saved way to ask for one kind of result.
The same page gives a few pictures of what that looks like in practice:
- An assistant reads Google Calendar and Notion, so the reply can use your real day and your real notes.
- Claude Code reads a Figma design and builds a web app from it.
- A company chatbot queries several databases and answers in the chat.
- A model builds a 3D design in Blender and sends it to a 3D printer.
The model is still the part that reads and writes language. MCP is the part that lets it reach the other system and, when you allow it, take an action there.
Host, client, and server
Three names show up as soon as you open the architecture page. In daily use you choose two of them.
| Name | What it is | A familiar example |
|---|---|---|
| Host | The AI app you are talking to. It manages the connections. | Claude, ChatGPT, Visual Studio Code, Cursor |
| Server | A program that knows one outside system and shares that context. | A files server, a calendar server, a database server |
| Client | The connection the host opens to one server. One server, one client. | Created for you when you add a server |
Connect three servers, and the host holds three clients. You rarely meet the client as its own screen. The app creates it when you add the server.
"Server" here means the program that supplies context. The word does not say where that program runs. It can sit on your computer, or it can sit on the internet.
Tools, resources, and prompts
A server can hand the app three kinds of things. These are the core pieces the architecture overview describes.
| Piece | Plain meaning | Example |
|---|---|---|
| Tool | An action the app can call | Search, run a query, create a file |
| Resource | Something the app can read and use as context | The text of a file, a database row, the layout of a table |
| Prompt | A saved template for a kind of request | Wording for a task, plus a few examples of a good question |
A prompt is the "specialized prompt" the introduction lists under workflows. A database server might offer all three together: a tool that runs a query, a resource that shows the tables, and a prompt that shows how to ask.
The app asks the server what is available, then uses the pieces that fit the request. The list can change later. When it does, the server can tell connected apps.
On your computer, or on the internet
The messages mean the same thing either way. The path they travel on is different.
- On your computer. A local server talks through the program's standard input and output. The docs call this stdio. A files server is the usual example. Claude Desktop can start it on the same machine, and the messages stay on that computer.
- On the internet. A remote server talks over HTTP. The current form is called Streamable HTTP. Sentry's server works this way: it runs on Sentry's side, and many apps can connect to it. A remote server can ask you to sign in. The project recommends OAuth for that sign-in.
Why it is worth having
The introduction splits the benefit by where you sit.
- If you build the connection. You build it once, and you spend less time writing a separate plugin for every assistant.
- If you build the assistant. You can offer data, tools, and apps that other people already maintain. The assistant gets more capable from that catalog.
- If you use the assistant. It can see the information you connect, and it can take an action for you when the task needs one. The calendar answer can come from the calendar.
Where it already works
MCP is an open protocol. The introduction lists apps that support it, including Claude, ChatGPT, Visual Studio Code, Cursor, and MCPJam. A server written to the standard can show up in each app that speaks it.
That is the practical promise: build the connection once, and use it in more than one place. The same project also documents MCP Apps, interactive views that run inside the AI client, and guides for building a server or building a client.
MCP and agent skills
This site also has a guide to agent skills. The two ideas meet, and they answer different questions.
| MCP | Agent skill | |
|---|---|---|
| What it stores | A connection to an outside system | The steps for one kind of task |
| Example | Reach Notion, a database, or your files | Write the weekly status email in the team format |
| When it is used | When the assistant calls that system | When that kind of task comes up |
The connection is how the assistant reaches the system. The skill is how your team wants the work done once it is there. A calendar server can supply tomorrow's meetings. A skill can say how those meetings should be turned into the Monday email. Use both when the job needs a specific system and a specific way of working.
Questions people ask first
- Is MCP an app I have to install?
- MCP is a shared standard. The app you already use, such as Claude, ChatGPT, Cursor, or Visual Studio Code, is the host. You add a server for the system you want that app to reach.
- Do I have to write code to use MCP?
- Many apps let you turn on an existing server from settings or a directory. You write a server when you want to connect a system that does not have one yet.
- What is the difference between a host, a client, and a server?
- The host is the AI app. The server is the program that knows one outside system. The client is the connection the host opens to that server. One server gets one client.
- What can a server give the assistant?
- Three kinds of things. A tool is an action, such as searching or creating a page. A resource is something to read, such as a file or a record. A prompt is a saved template for a kind of request.
- Will one server work in more than one app?
- That is the point of the standard. Claude, ChatGPT, Visual Studio Code, Cursor, and others speak MCP, so a server written to the standard can be used in each app that supports it.
- How is MCP different from an agent skill?
- MCP connects the assistant to an outside system. An agent skill stores the steps for a kind of task. The connection is how the assistant reaches Notion. The skill is how your team wants the page written.
- Does a server have to run on the internet?
- A local server, such as one for your files, runs on your computer and talks through the program's input and output. A remote server runs on the internet, and the app talks to it over HTTP, usually after you sign in.
Related on this site
- What Are Agent Skills?, the folder of steps an assistant opens for one kind of task
- Agent Skills browser, a catalog of public skill files
- All guides