10 MCP Servers Every Developer Should Connect to Their AI

November 6, 202610 min read
10 MCP Servers Every Developer Should Connect to Their AI

10 MCP Servers Every Developer Should Connect to Their AI

Your AI coding agent is smart. But it is also blind.

It cannot see your database. It cannot open your app in a browser. It cannot read your GitHub issues or your error logs. So it guesses.

It guesses your table names. It guesses how a library works today. It guesses why the page is broken. Sometimes the guess is right. Often it is not, and you lose an hour.

MCP servers fix this. They give your AI a direct line to the tools you already use. In this guide, I list 10 real MCP servers that developers use. I checked each one on its official page in November 2026.

Liked our post on 5 AI agent skills for mobile app developers? Then this is part two. Skills teach your AI how to work. MCP servers let it see and touch your real tools.

The 10 MCP servers at a glance

Server

What it connects

Best for

GitHub

Repos, issues, pull requests, Actions

Daily code and PR work

Playwright

A real browser it can click and type in

Testing web pages and forms

Chrome DevTools

Chrome's console, network and performance tools

Debugging slow or broken pages

Filesystem

Folders you choose on your computer

Chat apps without file access

Git

A local Git repository

Reading history and diffs

Context7

Up-to-date library documentation

Stopping old, wrong API code

Supabase

Your Postgres database and Supabase project

Schema and data questions

Sentry

Live errors, traces and logs

Fixing real production bugs

Figma

Design files and their details

Turning designs into UI code

Linear

Issues, projects and comments

Planning and tracking tasks

You do not need all ten. Start with two or three that match your daily work. More servers also means more tools for the AI to choose from.

What is MCP? (In simple words)

MCP stands for Model Context Protocol. A protocol is just a shared set of rules for how two programs talk.

The official MCP site calls it an open-source standard for connecting AI apps to outside systems. It also gives a nice picture: MCP is like a USB-C port for AI apps.

Think about your phone charger. One USB-C cable works with many devices. In the same way, one MCP server works with many AI tools. Claude, ChatGPT, VS Code and Cursor all support MCP, as the official site lists.

Here are the key words to know:

  • Host: the AI app you use, like Claude Code or VS Code.

  • Client: the small connector the host creates for each server. You rarely see it.

  • Server: a small program that gives the AI access to one thing, like GitHub or a browser.

  • Tools: the actions a server offers, like "create an issue" or "run a SQL query".

A server can run on your own computer. That is a local server. It can also run on the company's cloud. That is a remote server, and you usually just log in to it. The MCP architecture page explains both types.

1. GitHub MCP Server

This is GitHub's official MCP server. It lets your AI read code across repos. It can also manage issues and pull requests. It can check GitHub Actions runs and security alerts.

You can use the remote version that GitHub hosts. Then there is nothing to install. You can log in with OAuth. That is a "Log in with GitHub" style sign-in. Or you can use a personal access token, a password-like key made for apps.

Two settings are worth knowing. Toolsets let you turn on only the groups you need, like issues or pull requests. If you run the server yourself, the --read-only flag blocks every write action.

Why I would add it: ask "summarise the open bugs in this repo" and get a real answer.

Best for: anyone who works on GitHub every day, from students to open-source contributors.

2. Playwright MCP

Playwright MCP comes from Microsoft. It gives your AI a real browser to control. It can open pages, click buttons, fill forms and move between pages.

It does not look at screenshots to decide where to click. It reads the page's accessibility tree. That is a structured list of every button, link and field on the page. So it does not need an AI that reads images.

The README is honest about limits. It says Playwright MCP is not a security boundary. So do not point it at sites or accounts you would not trust it with.

Why I would add it: your AI can test the login page it just built. No more "it should work".

Best for: web developers and anyone writing tests that click through your app like a real user.

3. Chrome DevTools MCP

Chrome DevTools MCP is from the Chrome DevTools team. It lets your AI control and inspect a live Chrome browser.

It goes deeper than clicking. It can record a performance trace and pull out speed insights. It can check network requests. It can also read console messages. It shows which line of your own code caused an error.

Note one privacy point from its README. Some speed tools may send page URLs to Google's CrUX service. CrUX holds real-user speed data. The tool also collects usage statistics by default. You can turn these off with the --no-performance-crux and --no-usage-statistics flags.

Why I would add it: the AI can measure why a page is slow.

Best for: frontend developers who fight slow pages and strange console errors.

4. Filesystem MCP Server

The Filesystem server is one of the official reference servers. It lets the AI read, write, move and search files.

The key safety feature is simple. You list the folders it may use. It cannot touch anything outside them.

Coding agents like Claude Code can already read your project files. So this server matters most for chat apps like Claude Desktop. They cannot see your disk on their own. The official quickstart uses it as its first example.

Why I would add it: it is the easiest way to learn how MCP works, step by step.

Best for: beginners and anyone using a chat app as their main AI helper.

5. Git MCP Server

The Git server is another official reference server. It lets the AI check status, read diffs, view the log, create branches and make commits.

Two honest warnings come from its own page. First, it is still in early development, so tools may change. Second, always set the --repository option. Without it, the AI can reach any Git repo on your computer.

Why I would add it: ask "what changed since yesterday?" and get an answer from real history.

Best for: students learning Git, and chat apps that cannot run Git commands themselves.

6. Context7 (up-to-date docs)

AI models learn from old data. So they often write code for an old version of a library. Context7, from Upstash, tackles this problem.

It fetches current, version-specific docs and code examples for a library. Then it hands them to your AI.

You can use it without a key. The README says a free API key gives you higher rate limits. A rate limit is how many requests you can make in a set time.

Why I would add it: fewer "this function does not exist" errors with new library versions.

Best for: anyone using fast-moving libraries like Next.js, React or new Python packages.

7. Supabase MCP Server

Many students build projects on Supabase, which gives you a hosted Postgres database. The Supabase MCP server connects your AI to your Supabase projects. It can list tables, run queries and fetch project settings.

It has two safety options I strongly suggest:

  • Read-only mode: the AI can look, but it cannot change data.

  • Project scoping: the AI can see one project only, not your whole account.

The Supabase MCP guide also warns about prompt injection. That is when bad instructions hide inside your data, like a support ticket. They can trick the AI into a harmful query.

Why I would add it: the AI writes queries for your real tables, not imaginary ones.

Best for: full-stack students and anyone building on Supabase.

A small note: the old reference Postgres server is now archived, according to the official servers repo. So prefer a server that its vendor maintains.

8. Sentry MCP Server

Sentry tracks errors in live apps. The Sentry MCP server is a remote server that connects your coding agent to that data.

Your AI can search issues, errors, traces and logs. A trace is a step-by-step record of a slow request. The project says it is built for human-in-the-loop coding agents. That means you stay in charge and approve the steps.

The Sentry MCP page explains setup. You log in with OAuth, so you do not paste a token anywhere. You can limit it to one project, which Sentry recommends.

Why I would add it: "fix last night's top crash" becomes a real task with real data.

Best for: developers with an app in production (the live app your users see) that already uses Sentry.

9. Figma MCP Server

The Figma MCP server lets your AI read the details of your Figma designs. So the UI code it writes can match the design.

Figma recommends the remote server. As of November 2026, it works on many Figma plans. This includes the free Starter plan. Usage limits depend on your plan and seat.

It can also write designs back to the canvas. Figma says this is free during the beta. Later, you will pay based on how much you use it.

Why I would add it: your AI copies the real design, not a rough idea of it.

Best for: frontend and mobile developers who work with designers.

10. Linear MCP Server

Linear is a popular issue tracker for software teams. The Linear MCP server lets your AI find, create and update issues, projects and comments.

It is hosted by Linear, and you log in with OAuth. There is also a read-only address. Use it if you only want the AI to look, not edit.

Why I would add it: say "pick up issue ENG-42" and the AI reads the full ticket first.

Best for: teams and startups that plan work in Linear.

How to add an MCP server

Most AI apps use a small JSON config file. The exact file name depends on your app. Here is the format the official MCP quickstart uses for Claude Desktop:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "C:\\Users\\yourname\\Projects"
      ]
    },
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

Each entry has a name and a command to start the server. The folder path at the end is the only folder the Filesystem server may use. Most local servers like these need Node.js installed.

In Claude Code, one command adds a remote server:

claude mcp add --transport http linear-server https://mcp.linear.app/mcp

Then run /mcp inside Claude Code to log in.

Two small gotchas, as of November 2026:

  • VS Code supports two formats. One uses mcpServers and the other uses servers. Check the VS Code MCP docs before you copy a config.

  • Remote servers usually ask you to log in the first time you use them.

Always copy the setup from each server's own README. Formats change, and the README is the source of truth.

Stay safe: MCP security basics

An MCP server is not just a plugin. It is code that acts with your permissions. The official quickstart warns that a local server runs with your user account's rights. So it can do any file action you can do.

Follow these simple rules:

  1. Install only trusted servers. Prefer servers from the company that makes the tool, like GitHub, Sentry or Figma. The official MCP Registry (still in preview) checks who published a server. But it says it does not scan the code itself.

  2. Give the least access possible. Use read-only modes, toolsets and project scoping. Only list the folders the AI truly needs.

  3. Never give write access to production blindly. Test on a copy or a development branch first. Supabase's guide suggests exactly this.

  4. Protect your secrets. Keep tokens in environment variables (settings stored outside your code). Never paste them into chat or commit them to Git. A shared config like .mcp.json can read ${VAR} values from your environment instead.

  5. Read before you approve. Many apps ask before a tool runs. Do not click "allow" without reading what it will do.

  6. Watch for prompt injection. Web pages, tickets and comments can hide instructions. Claude Code's docs warn that servers that fetch outside content carry this risk.

Even the official reference servers come with a note. The servers repo calls them educational examples, not production-ready tools. So add your own safeguards.

Final thoughts

MCP servers do not make your AI smarter. They make it better informed. And a well-informed AI guesses far less.

Here is a simple plan for this week:

  • Add one server you would use daily, like GitHub or Context7.

  • Use it on a small personal project first.

  • Turn on read-only mode wherever you can.

  • Add a second server only when you feel a clear need.

AI tools still need a developer who understands the code they write. Keep your basics strong with our free tutorials and coding problems. Those skills are what let you check the AI's work.

FAQ

What are MCP servers in simple words?

MCP servers are small programs that connect an AI app to one outside tool. Examples are GitHub, a browser or a database. They let the AI read real data and take actions, with your approval.

Are MCP servers free?

Many are free and open source, like Playwright MCP and the Filesystem server. But the service behind a server may have its own plans. Sentry, Figma and Supabase each have their own pricing and limits.

Which AI tools support MCP?

The official MCP site lists Claude, ChatGPT, Visual Studio Code and Cursor, among many others. Claude Code, Codex, Windsurf and Zed also support it, as Linear's docs show.

Are MCP servers safe to use?

They are as safe as the server and the access you give it. Install servers from trusted publishers. Use read-only modes. Keep production databases and secrets out of reach.

What is the difference between MCP servers and AI agent skills?

Skills are instructions that teach the AI how to do a task well. MCP servers give the AI access to real tools and data. Many developers use both together.

What is the difference between a local and a remote MCP server?

A local server runs on your own computer. Your AI app starts it for you. A remote server runs on the company's cloud. You connect over the internet and usually log in first.