What Is MCP? The Model Context Protocol for Beginners
MCP is a standard way to plug tools and data into AI apps — a USB-C port for AI. What the Model Context Protocol is, what servers and clients are, what tools, resources, and prompts do, real examples, and the safety questions to ask before connecting one.
You'll see "MCP server" everywhere in AI tooling now: MCP servers for GitHub, databases, browsers, Figma, Notion, Sentry. It's become the standard way to give an AI assistant new abilities. Here's what it is, in plain terms.
The problem it solves
An AI model on its own can only read and write text. To be useful, AI apps need to do things: look up a GitHub issue, query a database, read a design file, check an error tracker.
Before MCP, every AI app built its own connection to every service. Ten AI apps and ten services meant a hundred custom integrations.
MCP — the Model Context Protocol — is an open standard, introduced by Anthropic in late 2024, that defines one common way for AI apps to connect to tools and data. Build an MCP server for your service once, and any AI app that supports MCP can use it. It's often described as a USB-C port for AI: one plug, many devices.
The pieces
- MCP host / client — the AI app you use: Claude Code, Claude Desktop, VS Code, Cursor, and many others.
- MCP server — a small program that exposes a service's abilities in the standard format. There are servers for GitHub, Postgres, Slack, browsers, file systems, and thousands more.
The client connects to one or more servers, learns what each offers, and lets the model use them.
What a server can offer
An MCP server can provide three kinds of thing:
- Tools — actions the model can take: "create an issue," "run this SQL query," "take a screenshot of this page." The most common kind.
- Resources — data the model can read: files, database schemas, documents.
- Prompts — ready-made instructions for common tasks, which you can pick from a menu.
When you ask "What are the open bugs labelled 'checkout'?", the model sees it has a GitHub tool for searching issues, calls it, reads the result, and answers.
How servers run
Servers connect in two main ways:
- Locally — the AI app starts the server as a program on your own computer and talks to it directly. Good for access to local files or tools.
- Remotely, over HTTP — the server runs on the internet, often hosted by the service's own company, and you sign in with OAuth. (Sign in with Google Explained covers how OAuth works.)
Using MCP with Claude Code
In Claude Code, you add a server with the claude mcp add command, and list what's connected with /mcp. Servers can be configured just for you, or shared with your team through a .mcp.json file in the project. Claude Code MCP Servers walks through setup and scopes.
Some genuinely useful examples:
- GitHub — read issues and pull requests, review comments.
- A database — let Claude inspect the schema and query a development database.
- A browser automation server — let Claude open your app, click through it, and see what breaks.
- Your error tracker — pull the real stack trace for a production bug.
- Documentation servers — up-to-date docs for libraries, reducing outdated suggestions.
The safety questions
An MCP server gives an AI model real abilities — sometimes including the ability to read private data or change things. Treat adding one like installing software with access to your accounts:
- Who made it? Prefer official servers from the service's own company, or well-known open-source projects. Anyone can publish an MCP server.
- What can it do? A read-only database server is very different from one that can run
DROP TABLE. Connect to development data, not production, where you can. (How to Stop an AI Agent From Deleting Your Production Database.) - What will it read? Content an MCP tool returns — an issue, a web page, an email — could contain text written to trick the model into doing something else. That's prompt injection, and it's the main new risk MCP brings. (Prompt Injection and Coding Agents.)
- What credentials does it hold? Use tokens with the narrowest permissions possible.
- Do you need it? Every connected server's tool descriptions take up context, even when unused. Connect what you'll actually use. (What Is a Context Window?.)
Claude Code asks your permission before using MCP tools unless you've approved them, which gives you a chance to see what's being called.
Building your own
If you have an internal tool or API you want AI assistants to use, you can build an MCP server for it using the official SDKs, available for TypeScript, Python, and other languages. A basic server exposing a couple of tools is a short program. Securing MCP Servers covers doing it safely.
EasySpawn runs Claude Code in a persistent, isolated workspace, so the MCP servers you connect — and the credentials they use — live in a sandbox for that project rather than on your personal machine. See how it works or join the waitlist.
Related: Claude Code MCP Servers · What Is an AI Coding Agent? · What Is an API?
Keep reading
What Is Claude Code? A Beginner's Guide
Claude Code is Anthropic's AI coding agent: instead of just suggesting code in a chat, it reads your project, edits files, runs commands, and checks its own work. What it is, where you can use it, what it costs, and how to get started without getting into trouble.
What Is an AI Coding Agent? (And How It's Different From a Chatbot)
AI coding tools come in three flavours: chatbots that answer, assistants that autocomplete, and agents that take actions — editing files, running commands, and checking their own work. What makes something an agent, why it matters, and what agents need to work well and safely.