What Is an MCP Server? MCP in AI Explained, With a Real Example
An MCP server gives AI apps like Claude and ChatGPT tools and data through the Model Context Protocol. How it works, MCP vs API, and real server examples.
An MCP server is a program that gives an AI app, such as Claude, ChatGPT or Cursor, tools and data it can use for you: your files, a database, your Search Console numbers. MCP in AI stands for Model Context Protocol, an open standard that apps and servers both follow, so you connect a server by pasting a URL or a command into the app. Nobody writes integration code. When your request needs one of the server's tools, the model calls it and the result comes back into the chat. To set one up in Claude, follow how to add MCP servers to Claude.
TL;DR
- MCP (Model Context Protocol) is an open standard for connecting AI apps to outside tools and data. Anthropic open-sourced it in November 2024.
- An MCP server offers tools (actions), resources (read-only data) and prompts (templates). The AI app is the host, and it runs one MCP client per server.
- Local servers run on your computer over stdio. Remote servers run on the web over Streamable HTTP and usually sign you in with OAuth.
- A developer writes code against an API. An MCP server describes its tools so a model can find and call them.
- A connected server's tools can read, change and delete data in the service behind it. Connect only servers you trust.
What is MCP in AI?
Anthropic open-sourced the Model Context Protocol on November 25, 2024, with a specification, SDKs, local server support in the Claude Desktop apps and a repository of example servers. The protocol's documentation is at modelcontextprotocol.io.
Before MCP, each AI app needed its own connector for each data source, and each service had to build one per app. With MCP, a service builds one server, and every app that supports MCP can use it. The official docs compare it to a common plug: "Think of MCP like a USB-C port for AI applications."
What is an MCP server?
The server is the program on the service's side of that plug. It runs on your computer or on the provider's servers, and it can offer three kinds of building blocks, which the MCP docs call primitives:
| Building block | What it is | Who decides when it's used | Example |
|---|---|---|---|
| Tools | Functions the model calls to fetch data or take an action | The model | start_seo_audit starts a site crawl in Encited |
| Resources | Read-only data the app adds to the conversation | The app | A file's contents or a database schema |
| Prompts | Saved instruction templates | You, from a menu | "Summarize my meetings" |
Encited's MCP server offers tools only, plus usage guidance that the AI app reads when it connects, so the model knows which tool fits which question.
To see those tools work on your own data, connect the Encited MCP to Claude or ChatGPT and ask where your site ranks for your main search.
How does an MCP server work?
Every connection has three parts:
- Host: the AI app you use, such as Claude Desktop, Claude Code or Visual Studio Code.
- Client: a connection the host opens for one server. Connect three servers and the host runs three clients.
- Server: the program that provides the tools.
They exchange JSON-RPC 2.0 messages, a plain JSON format for asking another program to run a function. A tool call goes through four steps:
- Discovery. The client asks the server for its tools with
tools/list. Each tool comes back with a name, a description written for the model, and an input schema (JSON Schema) that lists the arguments it takes. - Choice. You ask something. The host gives the model the tool list, and the model decides whether a tool would help and which one.
- Call. The client sends
tools/callwith the tool name and its arguments. - Result. The server does the work and returns content, usually text. The model reads it and writes the answer.
Local and remote MCP servers
MCP messages travel over one of two transports, the channel between client and server:
| Local server | Remote server | |
|---|---|---|
| Where it runs | On your computer, started by the AI app | On the provider's infrastructure |
| Transport | stdio: the app and the server pass messages through standard input and output | Streamable HTTP: HTTP POST requests, with optional streaming |
| How you add it | A command, such as npx -y @playwright/mcp@latest | A URL, such as https://encited.com/api/mcp |
| Sign-in | Runs with your user account's permissions; API keys go in environment variables | Usually OAuth in your browser, which the MCP docs recommend |
| Serves | One client, as a rule | Many clients |
| Good for | Files, local databases, a browser on your machine | Hosted services you use from more than one device |
An MCP server example, step by step
On November 9, 2026, we asked Claude, with the Encited MCP server connected, where encited.com ranks on Google for "what is mcp".
- Claude picked
check_serp_positionfrom the tools Encited's server lists. Its description says it fetches the first twenty Google results for one query, finds the site's positions, adds matching Search Console pages, and costs 3 workflow credits. - Claude's MCP client sent the call. Trimmed to the parts that matter:
{"jsonrpc": "2.0","id": 1,"method": "tools/call","params": {"name": "check_serp_position","arguments": { "domain": "encited.com", "query": "what is mcp" }}}
- The server saved a run and answered at once:
{ "runId": "…", "workflowSlug": "serp-position", "status": "queued", "estimatedCredits": 3 }
- Claude called a second tool,
get_workflow_result, with that run ID. The result held 16 results from Google's top 20 positions, all about the Model Context Protocol, and encited.com was not among them. modelcontextprotocol.io's intro page ranked first, a YouTube explainer second and Google Cloud's MCP guide third. People Also Ask included "How is MCP different from API?" and "Does ChatGPT use MCP?", and Search Console matched one encited.com page to the query,/seo-mcp.
No part of Claude was written for Encited. It read the tool descriptions, chose two tools and turned the results into an answer.
MCP vs API
An MCP server often sits on top of an existing API: Encited's MCP tools call the same services as the Encited dashboard. What changes is the caller.
| API | MCP server | |
|---|---|---|
| Who calls it | Code a developer writes for that API | An AI app, when the model picks a tool |
| How the caller learns what it does | Documentation a developer reads | tools/list returns names, descriptions and input schemas |
| Message format | Each API picks its own (REST, GraphQL and others) | JSON-RPC 2.0 for every server |
| Sign-in | Each API picks its own | OAuth recommended for remote servers |
| Adding a new one | New integration code in every app | A URL or a command in the AI app's settings |
A web server answers HTTP requests from browsers and other programs. An MCP server answers MCP messages from AI apps, and a local one is often a small program that the AI app starts on your computer and talks to through stdio, with no network involved.
Use an API when you write software that calls a service directly. Use an MCP server when you want an AI app to work with the service while you talk to it.
MCP server examples
The MCP project keeps a short list of reference servers, maintained by its steering group. They are teaching examples for developers, and the project says they aren't production-ready. Published servers are listed in the MCP Registry, and Claude has its own connector directory. The servers Anthropic shipped at launch for Google Drive, Slack, GitHub and Postgres are now archived; GitHub runs its own remote server, listed below.
| Server | Local or remote | What it gives the AI app |
|---|---|---|
| Filesystem (reference) | Local | Read, create, move and search files in folders you allow |
| Git (reference) | Local | Read, search and change Git repositories |
| Fetch (reference) | Local | Fetch web pages and convert them for the model |
| Memory (reference) | Local | A knowledge graph that persists between chats |
| Playwright | Local | A browser the model can open, click through and read |
| Notion | Remote, https://mcp.notion.com/mcp | Your Notion workspace |
| Sentry | Remote, https://mcp.sentry.dev/mcp | Your Sentry projects and issues |
| GitHub | Remote, https://api.githubcopilot.com/mcp/ | Repositories, issues and pull requests |
| Encited | Remote, https://encited.com/api/mcp | Search Console data, technical SEO audits, crawl logs, AI visibility tracking, keyword and SERP research |
Encited's page checks also run in a browser with no AI app. Run one page through the free SEO audit to see its noindex and canonical tags and whether it has enough text without JavaScript.
Which AI apps support MCP?
- Claude: claude.ai, the Claude desktop and mobile apps through connectors, and Claude Code. The setup for each is in how to add MCP servers to Claude.
- ChatGPT: connects to remote MCP servers. Encited's ChatGPT steps are on the Encited MCP page.
- Visual Studio Code: GitHub Copilot's agent can use MCP servers, so Copilot in VS Code is an MCP client.
- Cursor: reads servers from
~/.cursor/mcp.json, or.cursor/mcp.jsoninside a project.
A remote server such as Encited's uses the same URL in each of them.
Are MCP servers safe?
Once connected, a server's tools can read data in the service behind it, create, change or delete that data, and take actions for you. Anthropic's connector docs and the MCP docs give these rules:
- Connect only servers from developers and organizations you trust.
- During sign-in, read the permissions the server asks for and refuse what it doesn't need.
- Watch for prompt injection: a malicious server, or a page a server fetches, can carry hidden instructions for the model.
- Read each tool approval request. Use Always allow only for servers you trust, and block tools you don't need.
- A local server runs with your user account's permissions, so it can do anything you can do in the folders and systems you give it.
On a remote server, access control is the provider's job. Encited's sign-in asks which account the AI app acts on and whether it gets read or write access, every request is checked again against your current role and plan, and Settings > Connections in Encited revokes a connection.
FAQ
Do I need to code to use an MCP server?
No. A remote server needs its URL and a sign-in. A local server needs a command, and many run through npx, which requires Node.js on your computer.
Are MCP servers free?
The protocol is open and free to use. Each server is run by its provider, and the service behind it may need its own account or a paid plan.
Can an AI app use more than one MCP server?
Yes. The host opens one client per server, so Claude can use a file server, a GitHub server and the Encited server in the same conversation.
