What is an MCP server? The Model Context Protocol explained

MCP lets AI tools like Claude and ChatGPT use outside tools and data. How MCP servers work, local vs remote, sign-in, security and a real example.

1. The short answer

An MCP server is a program that gives an AI tool new abilities, such as reading your files, querying a database or deploying a website, using the Model Context Protocol (MCP), an open standard that any compatible AI tool understands. Build the server once and it works in Claude, ChatGPT, Cursor, VS Code and many others.

The protocol's official site puts it neatly: “Think of MCP like a USB-C port for AI applications.” Before USB-C, every device needed its own cable. Before MCP, every AI tool needed its own integration for every service.

2. What MCP is, and why it exists

On their own, AI models only know what they were trained on and what you paste into the chat. To be useful for real work they need to look things up and take actions: read today's calendar, search a codebase, create a ticket, put a site online.

Anthropic introduced and open-sourced the Model Context Protocol in November 2024 to make that a standard. Instead of each AI app building a custom connection to each service, a service publishes one MCP server, and any AI app that speaks MCP can use it. The protocol's site lists support in AI assistants such as Claude and ChatGPT and in development tools such as VS Code and Cursor.

MCP doesn't decide how the model thinks. It only standardizes how context and actions pass between the AI app and the outside world.

3. How MCP works: hosts, clients and servers

MCP has three parts, in the protocol's own terms:

• Host: the AI application you use, such as Claude, Claude Code, Cursor or VS Code.
• Client: a connection the host makes for each server. If you've added three servers, your host keeps three clients.
• Server: the program that provides the abilities. It can run on your computer or on the internet.

Messages between them use JSON-RPC 2.0, a simple format for requests and replies. A server can offer three kinds of things, which MCP calls primitives:

• Tools: actions the AI can take, like create_site or run_query. This is what most people mean when they talk about MCP.
• Resources: information the AI can read, like a file or a database schema.
• Prompts: reusable templates that help the AI use the server well.

When you connect a server, your AI tool asks it which tools it has (a tools/list request). Each tool comes with a name, a description and the inputs it expects. When the model decides a tool would help, the host sends a tools/call request with the inputs, and the result goes back into the conversation. Good AI tools show you what's being called and ask before actions that change things.

4. Local and remote MCP servers

MCP defines two ways for clients and servers to talk, called transports:

• Local servers (stdio). Your AI tool starts the server as a program on your computer and talks to it through standard input and output. Good for things on your machine, like files or a local database. The server runs with your permissions, so only install ones you trust.
• Remote servers (Streamable HTTP). The server runs on the internet at an address like https://api.example.com/mcp, and messages travel over HTTPS. Good for online services, because there's nothing to install and the company running it keeps it up to date.

Remote servers need to know who you are. MCP recommends OAuth for this, the same “Sign in with…” flow you use elsewhere: your AI tool opens a browser, you sign in to the service and approve access, and the tool receives a token. You never copy an API key into a settings file.

5. Examples of MCP servers

• Files and code: read and search folders or a Git repository.
• Databases: run queries and describe tables.
• Work tools: calendars, documents, issue trackers and chat.
• Monitoring: look up errors and logs. The protocol's docs use Sentry's remote server as an example.
• Design to code: the protocol's docs mention Claude Code generating a web app from a Figma design.
• Hosting: create sites, deploy them and follow the build, like Cloudzyro's server below.

6. A worked example: Cloudzyro's remote MCP server

Cloudzyro hosts websites and apps, and it runs a remote MCP server at https://api.cloudzyro.com/mcp so your AI tool can put sites online for you. It's a good example of how a real server is designed.

What it offers. Ten tools, each doing one thing:

• check_name: checks whether an address like my-portfolio.cloudzyro.app is free, and suggests others.
• create_site: creates a site in your workspace and claims its address.
• get_upload_link and deploy_upload: send a project folder from your computer as a zip, then put it online.
• deploy_files: put a small site online from files sent directly.
• get_deployment: follow a deploy until it's live, or get the likely cause and a suggested fix if it fails.
• list_sites and get_site: see your sites, their addresses and their status.
• list_versions and restore_version: see earlier versions and put one back online.

What it deliberately can't do. It can never delete a site, see or set environment variables, or change a plan. Those stay in the console, where a person does them. That's least privilege in practice: the AI gets exactly the abilities the job needs, and the riskiest actions aren't available at all, however a conversation goes.

How you sign in. There are no keys to copy. You add the address to your AI tool, it opens your browser, you sign in to Cloudzyro, choose a workspace and select Allow. Each connection is tied to that one workspace. In Claude Code it looks like this:

claude mcp add --transport http --scope user cloudzyro https://api.cloudzyro.com/mcp
claude mcp login cloudzyro

Which tools work. It's proven with Claude, Claude Code, ChatGPT, Cursor and Codex, and documented for Gemini CLI, Windsurf and VS Code. Any tool that supports remote MCP servers with sign-in can use it. Setup for each: Claude, ChatGPT, Cursor, Codex, Gemini CLI and Windsurf.

Then you just ask: “Put this site online with Cloudzyro.” Your AI tool calls create_site, sends the files, checks get_deployment and tells you the address. The full walkthrough is in how to deploy a site built with Claude.

7. MCP security: what to watch for

An MCP server is a way for an AI to act on your behalf, so treat adding one like installing an app with access to your accounts. The protocol's security guide covers the details for builders. For users, the practical rules are:

• Trust the source. A local server runs code on your computer with your permissions. Install servers only from publishers you trust, and read the command before you approve it.
• Prefer narrow servers. A server that can only do what you need limits the damage if something goes wrong. Look for servers that leave destructive actions out.
• Check sign-in screens. With OAuth, check the address in your browser and the permissions being asked for before you select Allow.
• Watch for prompt injection. Text an AI reads, such as a web page, an email or a file, can contain instructions meant to trick it into calling tools. Keep approvals on for actions that change things.
• Keep secrets out of chats. Don't paste passwords or API keys into a conversation for a server to use. Good servers never need them.
• Clean up. Disconnect servers you no longer use. On Cloudzyro, Connect your AI tool in the console lists connected tools and lets you disconnect any of them.

For builders, the protocol's guidance includes accepting only tokens issued for your own server (never passing a client's token on to another API), asking for the smallest scopes you need, and verifying every request rather than trusting an ID the client sends.

8. How to start using MCP

1. Pick a server for something you already do by hand, such as deploying a site.
2. In your AI tool's settings, add the server's address, or find it in the tool's connector directory.
3. Sign in when asked and approve access.
4. Ask for the task in plain words. You don't need to name the tools; the AI picks them.

If you build things with AI, a hosting server is one of the most useful to add first, because it closes the gap between “it works on my machine” and a live link. Cloudzyro's Launch plan is free, allows business use and never puts sites to sleep.

9. Sources

• Model Context Protocol: What is MCP?
• Model Context Protocol: Architecture overview
• Model Context Protocol: Security best practices
• Anthropic: Introducing the Model Context Protocol
• Cloudzyro docs: What your AI tool can do