MCP Servers Guide: What They Are and How to Use Them
A plain-English MCP servers guide: what MCP is, the best servers to install, and how to set them up in Claude Code and Cursor in minutes.
TL;DR
MCP (Model Context Protocol) is a standard way to plug extra tools into an AI assistant so it can do things it normally can’t, like read your files, search the web, or talk to other services. Think of MCP servers as apps you install on a phone: each one gives your AI a new skill. This guide covers what MCP actually is, which servers are worth installing first, and how to get them running in Claude Code and Cursor.
What is MCP?
Your AI assistant is smart, but it lives in a box. By default it can only see what you paste into the chat. It cannot browse the web, read a file from your computer, check your calendar, or look up a row in a database (a table where apps store info). MCP is a simple rulebook that lets your AI step outside that box.
MCP stands for Model Context Protocol. A “protocol” here just means a shared language, a standard both your AI tool and an outside service agree to speak. Once they share that language, your AI can ask the outside service for information or tell it to do something.
An MCP server is the piece of software that speaks that language on behalf of some tool or service. You install it once, point your AI at it, and from that moment your AI can use that tool. No custom code. No manual copy-pasting.
The analogy that sticks: MCP servers are like browser extensions. Each one you install quietly adds a new power to the thing you are already using.
Why MCP servers matter for app builders
If you are building apps with AI tools like Claude Code or Cursor, you already know the rhythm: describe what you want, the AI writes code, you copy it in, something breaks, you explain the error, repeat. MCP servers can shorten that loop.
With the right MCP servers installed, your AI can read your project files directly instead of you pasting them in. It can search the web for documentation it doesn’t know yet. It can look up a value in your database without you having to fetch it yourself. The AI stops being a text-in, text-out tool and starts acting more like a teammate who can actually look things up.
If you are new to building with AI, the article on how to build an app without coding is a good place to start before going deeper here.
How MCP servers actually work
Here is the short version. When you send a message to your AI, the AI decides if it needs to call a tool to answer you. If an MCP server is available that covers what it needs, it sends a small request to that server. The server does the work (fetches a webpage, reads a file, runs a search) and sends the result back. The AI folds that result into its reply.
You do not see any of this. To you, it looks like the AI just knows things.
The MCP server runs locally on your computer or on a remote machine. Local servers are more common for developer tools because they can access your files and projects without sending everything to an outside service.
The best MCP servers to install first
There is no single right list here. The best servers depend on what you are building. But these are the ones that come up again and again as genuinely useful rather than just impressive in a demo.
Filesystem
The most immediately useful server for most builders. It gives your AI direct read access to files on your computer, so you can say “look at my project folder and tell me what’s broken” instead of pasting file after file into chat.
Many setups support write access too, though you should keep it scoped tightly to one project folder so the AI cannot accidentally touch files you care about.
Web search
AI models have a knowledge cutoff, meaning they stop knowing about new things at some point in the past. A web search server gives your AI access to current documentation, Stack Overflow threads, and recent news. If you are building with a framework that updates often, this one pays for itself fast.
Browser / web scraping
A step up from search. Instead of just looking something up, your AI can actually open a webpage and read what’s on it. Useful for pulling in competitor pricing, reading a terms-of-service page, or checking if a live site matches the code.
Database access
If your app stores data in a database (a structured table of info, like a spreadsheet), you can point an MCP server at it. Your AI can then run queries (questions you ask a database) against your actual data. This is especially useful for debugging why something looks wrong in your app.
GitHub
Lets your AI read issues, pull requests (proposed code changes), and code from your repositories (saved copies of your project’s history) on GitHub. Useful for context when you ask the AI to work on a feature that touches existing code.
Memory / knowledge base
Some MCP servers give your AI a persistent memory, meaning it remembers things across separate conversations. By default, AI assistants forget everything when you close the chat window. A memory server lets you store facts, preferences, or project context that carries forward.
Puppeteer / Playwright (browser automation)
These servers let your AI control a real browser, clicking buttons, filling forms, and navigating pages. This is less for everyday use and more for when you want to automate a repetitive web task or test an app by simulating a real user.
Setting up MCP servers in Claude Code
Claude Code is Anthropic’s AI tool for building software in your terminal (the text-based program launcher on your computer). Setting up MCP servers there is done through a configuration file (a settings document the tool reads on startup).
Step 1: Find or create your MCP config file
Claude Code looks for a file called claude_desktop_config.json in your app settings folder. On a Mac, that is inside ~/Library/Application Support/Claude/. On Windows, it is inside %APPDATA%\Claude\. If the file does not exist, create it.
Step 2: Add a server entry
The file holds a JSON object (a structured list of settings). Each MCP server gets its own named entry. Here is the pattern:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/your/project"]
}
}
}Replace /path/to/your/project with the actual folder you want to give the AI access to. The npx command is a way to run small packages (pre-packaged bits of code) without installing them permanently.
Step 3: Restart Claude Code
Save the file and restart the application. Claude Code will read the config on startup. If the server is configured correctly, you will see it listed under available tools when you start a new conversation.
Step 4: Test it
Ask Claude something that would require the new tool. If you added the filesystem server, try: “List all the files in my project folder.” If the server is working, Claude will use it rather than saying it doesn’t have access.
For more on working with Claude Code day-to-day, the Claude Code slash commands reference covers the commands that make it faster to use.
Setting up MCP servers in Cursor
Cursor is an AI-assisted code editor. It added MCP support in late 2024. The setup is similar to Claude Code but lives inside Cursor’s own settings panel.
Step 1: Open Cursor settings
Go to Settings (the gear icon or press Cmd+Shift+J on Mac, Ctrl+Shift+J on Windows). Look for the “Features” section, then find “MCP Servers.”
Step 2: Add a new server
Cursor has a UI (a visual panel you click through) for adding servers, which makes this a bit friendlier than editing a raw config file. Click “Add new MCP server,” give it a name, and paste in the command.
For example, to add the filesystem server, the command is:
npx -y @modelcontextprotocol/server-filesystem /path/to/your/project
Step 3: Enable it and reload
Toggle the server on. Cursor will prompt you to reload the window. After reload, the server is active.
Step 4: Check the tool shows up
Open a chat in Cursor (the Agent or Composer panel). You should see a tools icon. Click it to confirm your new server appears. If it does, the AI can now call it mid-conversation.
To understand how Cursor compares to other options for AI-assisted coding, the Cursor vs Claude Code comparison is a good read.
Common things that go wrong
The server does not show up. Usually a JSON formatting error in the config file. Paste your config into a JSON validator (search for one online) to check for missing commas or brackets.
Permission errors. The server is running but cannot read your files. Double-check that the path you gave it actually exists and that you have read permissions on that folder.
The AI does not use the server. MCP servers are tools the AI can call, but it decides when to use them. If the AI does not seem to be using a server, try being explicit: “Use the filesystem tool to read my package.json file.”
The server crashes silently. Check your terminal or Cursor’s output panel for error messages. Missing packages or an old version of Node (the environment that runs JavaScript-based servers) are the most common culprits. Running node --version in your terminal tells you what version you have; most MCP servers need version 18 or higher.
MCP servers and no-code AI builders
MCP is mainly aimed at tools where you write or review code. If you are using a fully no-code AI builder, you probably will not need to set up MCP servers manually because those tools handle their own connections behind the scenes.
For example, omg.dev builds all the connections in from the start: data storage, real-time updates, sign-in, and AI features all work without you wiring anything up. That is a different philosophy from the MCP approach, where you pick and assemble the tools yourself. Both are valid. The question is whether you want control over the plumbing or whether you want someone else to handle it so you can focus on the app itself.
If you are weighing those approaches, the best AI app builder guide for 2026 has a clear breakdown.
Picking the right MCP servers for your project
Start with one server, not five. The filesystem server is the best first install for almost everyone because the benefit is immediate and obvious: the AI can see your project without you describing it.
Add a web search server once you notice the AI is working from outdated information. Add the database server only when you are actively debugging data issues and the copy-paste approach is slowing you down.
Resist the urge to install every server you find. Each one adds a small overhead to every conversation and gives the AI more tools to choose from, which can make responses slower. Install what you use.
FAQ
What is an MCP server in simple terms?
An MCP server is a small program that gives your AI assistant a new skill, like reading files, searching the web, or talking to a database. You install it once and your AI can use it without you having to copy-paste information manually.
Do I need to know how to code to set up MCP servers?
You need to be comfortable editing a text file and running a command in your terminal. It is closer to following a recipe than writing code. If you have never opened a terminal before, it will take 20-30 minutes to get comfortable with the process.
Are MCP servers safe?
It depends on what permissions you give them. A filesystem server that can only read one project folder is low risk. A server with write access to your entire computer is higher risk. Start with the minimum permissions you need and expand only if necessary.
Which AI tools support MCP?
Claude Code and Cursor are the two most popular tools with solid MCP support as of mid-2026. Other editors and AI tools are adding support quickly as MCP becomes a wider standard.
Can I write my own MCP server?
Yes. The MCP specification is public and there are libraries in several programming languages to help. If you have a specific tool or internal service you want your AI to access, building a custom server is an option. It is a beginner coding project, not an advanced one.
How is MCP different from just giving the AI a long prompt with all my context?
A long prompt only works if you know exactly what context the AI needs upfront. MCP lets the AI fetch the context it needs during the conversation, on demand. It also scales better: you cannot paste a 10,000-file codebase into a prompt.
Will MCP slow down my AI responses?
Each tool call adds a small delay, usually a second or two. If the AI calls three tools in a row, you might wait an extra five seconds for a response. For most tasks that is a reasonable trade-off for the extra capability.