"MCP server" shows up in almost every AI agent tutorial now, usually without an explanation. This guide covers what MCP is, how the pieces fit together, how to connect a server to the AI tools you already use, and what to check before you trust one. At the end you'll connect a real, free MCP server in under a minute. ## What MCP is, in one paragraph The **Model Context Protocol (MCP)** is an open standard, introduced by Anthropic in late 2024, for connecting AI applications to outside tools and data. Before MCP, every AI app needed its own custom integration for every service. With MCP, a service builds one **MCP server**, and any app that speaks MCP (Claude, Claude Code, Cursor, VS Code and a growing list of others) can use it. Think of it as a standard plug: build the socket once, and every compatible device fits. ## The three pieces - **Host:** the AI application you use, such as Claude Desktop, Claude Code or Cursor. - **Client:** the part of the host that holds a connection to one server. - **Server:** a program that offers capabilities to the AI. It can run on your computer or on the internet. A server can offer three kinds of things: | Capability | What it is | Example | |---|---|---| | **Tools** | Actions the AI can call, with typed inputs and outputs | "check this URL", "create a calendar event", "search these docs" | | **Resources** | Data the AI can read | files, database rows, documentation pages | | **Prompts** | Reusable prompt templates the user can pick | "review this pull request" | Tools are by far the most common. When you ask the AI something, it sees the list of available tools with their descriptions, decides whether one helps, and calls it. The server runs the action and returns the result, and the AI uses that result in its answer. ## Local vs remote servers MCP servers connect in one of two ways: - **Local (stdio):** the host starts the server as a program on your machine and talks to it directly. Good for things that need local access, like your files or a local database. You install and update it yourself. - **Remote (Streamable HTTP):** the server lives at a URL, like a web API. Nothing to install, it's always up to date, and it works from web-based AI apps too. This is how most hosted services offer MCP. ## Connect a remote MCP server (example) Stormap runs a free remote MCP server at `https://stormap.ai/mcp`. It has tools for checking how AI assistants see a website (an AI visibility score, AI crawler access, llms.txt and robots.txt generators) and for searching Stormap's AI agent guides. It works without signing up, so it's a handy one to practice with. **Claude Code:** ```bash claude mcp add --transport http stormap https://stormap.ai/mcp ``` Then start Claude Code and ask something like: *"Use stormap to score how visible example.com is to AI assistants."* **Claude (web or desktop):** open Settings → Connectors → Add custom connector, and enter `https://stormap.ai/mcp`. **Cursor, VS Code and other clients** usually take a JSON config: ```json { "mcpServers": { "stormap": { "url": "https://stormap.ai/mcp" } } } ``` The exact file location differs by app; check your app's MCP settings page. After connecting, most apps show the server's tools in a list, so you can see exactly what the AI is allowed to call. ## What to check before you trust an MCP server An MCP server can do whatever its tools do, so treat adding one like installing an app: 1. **Know who runs it.** Prefer servers from the service's own company or well-known open-source projects. The official MCP Registry (registry.modelcontextprotocol.io) is a good place to look them up. 2. **Read the tool list.** A "weather" server that also has a "send email" tool is a red flag. 3. **Give the minimum access.** Use read-only API keys or tokens with narrow scopes. Don't hand a server your admin credentials. 4. **Keep confirmations on.** Most hosts ask before running a tool. Leave that on for anything that writes, sends, deletes or spends money. 5. **Watch for prompt injection.** Text that comes back from a tool (a web page, an email, a document) is data, not instructions. A well-built agent won't follow orders hidden in tool results, but it's worth knowing the risk exists, especially when combining a server that reads untrusted content with one that can take actions. ## Building your own If you have an API or internal tool your team keeps copying into chats, it's a good MCP candidate. Official SDKs exist for TypeScript, Python and other languages, and a minimal server is a few dozen lines: describe each tool's name, purpose and input schema, then write the function that runs it. Good tool descriptions matter more than you'd expect, because the AI decides when to use a tool based on that text alone. ## Quick recap - MCP is a standard plug between AI apps and outside tools and data. - Servers offer tools (actions), resources (data) and prompts (templates). - Remote servers are a URL you add; local servers are programs you run. - Check who runs a server, what its tools can do, and keep permissions tight. Want to try it now? Add `https://stormap.ai/mcp` to your AI app and ask it to check your own website. The full tool list and setup instructions are on the [Stormap developers page](/developers).