12 MCP Servers Worth Installing, and the Rules First

7 min read · Last updated: October 24, 2026

MCP servers are worth installing, but only a few, and only with read-only access at first. The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems. Twelve servers cover most developer work: code, browser, docs, database, payments, deploys, design, notes, errors, files, web and memory.

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The protocol

• What is the Model Context Protocol? (modelcontextprotocol.io)
• Reference servers repository (GitHub, checked Oct 24, 2026)

The official server docs

• GitHub MCP Server
• Playwright MCP (Microsoft)
• Context7 (Upstash)
• Supabase MCP guide
• Stripe MCP docs
• Vercel MCP docs, last updated Sept 15, 2026
• Figma MCP server guide
• Notion MCP docs
• Sentry MCP

Key takeaways

• MCP is described by its own docs as "a USB-C port for AI applications": one standard way to plug an AI into external systems.
• Of the twelve servers below, nine are hosted remote servers you connect to with OAuth. Three are local reference servers.
• The reference servers are "reference implementations," not production-ready, per the official repository.
• Stripe stops accepting full-access and untagged restricted keys on November 30, 2026.
• Connecting Vercel MCP gives the AI the same access as your Vercel user account.
• Start read-only, scope to one project, and keep human approval on.

Why do I think you should install MCP servers at all?

Because an AI that can only see the code you paste is guessing. An MCP server lets it read the real thing: your repo, your database, your error log. MCP is an open standard that lets AI apps like Claude or ChatGPT connect to data sources, tools and workflows, according to the official MCP documentation.

That is the whole take. The model is not the bottleneck on most tasks anymore. Context is.

The official intro lists what this looks like in practice: agents reading your Google Calendar and Notion, and Claude Code generating a web app from a Figma design. It also says the protocol is supported by Claude, ChatGPT, Visual Studio Code and Cursor, so you build the setup once and it follows you between tools.

If you are still choosing the AI tool itself, start with our ranking of AI coding tools in 2026. This post is about what you plug into it.

Which 12 MCP servers should every developer know?

The twelve below cover the daily loop of a web developer: write code, test it, read docs, store data, take payments, deploy, match the design, track tasks, and catch errors. Each one is documented by its maker. The table gives the quick view, and the sections after it give an example prompt for each.

#	Server	What it connects the AI to	Where it runs	Auth
1	GitHub	Repos, issues, pull requests	Remote or local Docker	OAuth or personal access token
2	Playwright	A real browser	Local, via npx	None
3	Context7	Current library docs	Remote	API key as Bearer token
4	Supabase	Your Supabase projects	Remote or local CLI	Per Supabase docs
5	Stripe	The Stripe API and docs	Remote	OAuth or agent API key
6	Vercel	Projects, deployments, logs	Remote	OAuth
7	Figma	Design files	Remote or desktop	Per Figma plan
8	Notion	Pages and databases	Remote	OAuth
9	Sentry	Errors and performance data	Remote	OAuth
10	Filesystem	Local files	Local	Configurable access controls
11	Fetch	Web pages	Local	None
12	Memory	A persistent knowledge graph	Local	None

Auth cells say "per" where the page I read did not spell it out for that server.

1. GitHub

The official GitHub MCP Server "connects AI tools directly to GitHub's platform," letting them read repositories and code, manage issues and pull requests, and automate workflows. GitHub hosts it at https://api.githubcopilot.com/mcp/. Tools are grouped into toolsets, so you can switch on only issues or pull_requests.

Example prompt: "List open issues labelled bug, group them by file, and draft a fix plan for the oldest one."

2. Playwright

Microsoft's Playwright MCP lets an AI control a browser. It works through "structured accessibility snapshots" of the page, not pixels, so it does not need screenshots or a vision model. It needs Node.js 18 or newer and launches with npx @playwright/mcp@latest.

Example prompt: "Open localhost:3000, sign up with a test account, and tell me which step breaks."

3. Context7

Context7 "pulls up-to-date, version-specific documentation and code examples straight from the source," per its README. It fixes the classic problem of an AI writing code for an old version of a library. Its own example prompt is: "How do I set up a Next.js 16 proxy? use context7".

4. Supabase

Supabase's hosted server at https://mcp.supabase.com/mcp lets an AI "interact with and query your Supabase projects on your behalf." Supabase warns that prompt injection is "the primary attack vector unique to LLMs." Use read_only=true and scope the server to one project with project_ref.

Example prompt: "Show me the schema for the orders table and explain why this query is slow."

5. Stripe

Stripe's server at https://mcp.stripe.com lets agents call the Stripe API and search its documentation. For risky writes such as refunds and outbound payments, Stripe requires a human to click an approval URL. The approval expires after 24 hours if you ignore it.

Example prompt: "Find failed payments from the last 7 days and summarise the top three reasons."

6. Vercel

Vercel MCP, at https://mcp.vercel.com, can search Vercel documentation, manage projects and deployments, analyse deployment logs and query web analytics. Setup is one command: npx -y add-mcp https://mcp.vercel.com -g.

Example prompt: "Why did my last deployment fail? Read the build logs and tell me the first error."

7. Figma

Figma's server "gives AI agents the design context they need to generate code from your Figma files." It can pull variables, components and layout data into your IDE. Copy a link to a frame, paste it into your client and ask it to implement the design.

Example prompt: "Implement this frame as a React component using our existing button styles."

8. Notion

Notion MCP is a remote server hosted by Notion. Connected clients can search, read, create and update pages and databases, after you approve access with OAuth. The client sees content you have permission to view.

Example prompt: "Find the spec for the billing page and turn the acceptance criteria into a checklist."

9. Sentry

Sentry's server connects assistants to Sentry for "searching errors, analyzing performance, triaging issues." All connections use OAuth. You can scope the URL to one project with /mcp/{organizationSlug}/{projectSlug}, which Sentry recommends.

Example prompt: "What is the most frequent unresolved error this week, and which file is it in?"

10. Filesystem

The reference Filesystem server offers "secure file operations with configurable access controls." You choose which folders the AI may touch. Give it your project folder, not your home directory.

Example prompt: "Read every .env.example in this folder and list the variables my README does not mention."

11. Fetch

The reference Fetch server does "web content fetching and conversion for efficient LLM usage." It turns a web page into text the model can read.

Example prompt: "Fetch this changelog page and tell me what breaks in my project."

12. Memory

The reference Memory server is a "knowledge graph-based persistent memory system." It lets the AI carry facts between sessions instead of starting cold every time.

Example prompt: "Remember that this project uses Convex and that we never run builds while the dev server is up."

Are MCP servers a security risk?

Yes, and that is the strongest argument against my take. An MCP server gives an AI real access, and untrusted text can try to steer it. The risk is real, but every vendor above documents a control for it, so the answer is to narrow access, not to skip the tools.

The evidence comes straight from the docs:

• Supabase says prompt injection "might trick an LLM into following untrusted commands," and that its protective wrapping "is not foolproof."
• Vercel says that connecting its MCP "grants the AI system you're using the same access as your Vercel user account."
• Stripe tells you to turn on your client's tool-approval setting and to be careful combining its server with other servers.
• The reference servers repository says they are "not as production-ready solutions" and that developers "should evaluate their own security requirements."

The counter-argument says: do not give an AI that much power. I agree with the worry and disagree with the conclusion. Reading is cheap and safe to start with. Writing is where the damage happens.

What should you actually do this week?

Install three servers, not twelve, and set each to the narrowest access it offers. Pick one for code (GitHub), one for what is running (Sentry or Vercel), and one for docs (Context7). Add the rest only when a real task needs them.

1. Add one server over OAuth where possible, since Stripe notes OAuth lets you revoke a client without sharing an API key.
2. Scope it: one project for Supabase and Sentry, one folder for Filesystem, specific toolsets for GitHub.
3. Turn on read-only mode where it exists, as Supabase offers.
4. Keep human approval on for every write.
5. Review what the AI did before you trust it. Prompts that state a clear goal and constraint help here, as in our magic prompt formula.
6. If you use Stripe keys, migrate to agent API keys or OAuth before November 30, 2026.

If your editor setup is still bare, our list of VS Code extensions every developer needs in 2026 is a good companion.

FAQ

What is an MCP server?

An MCP server is a program that exposes data, tools or workflows to an AI application using the Model Context Protocol, an open-source standard. Your AI client connects to it and can then search, read or act in that system, within the permissions you grant.

Do MCP servers work with ChatGPT and Claude?

Yes. The official MCP documentation names Claude and ChatGPT among supported assistants, along with Visual Studio Code and Cursor. Vercel's docs list setup steps for Claude Code, ChatGPT, Codex CLI, Cursor and VS Code with Copilot.

Are the official reference MCP servers safe for production?

Not by default. The reference servers repository calls them reference implementations that demonstrate MCP features and the official SDKs. It says they are not production-ready solutions, so you must add your own safeguards.

Which MCP server should I install first?

Start with the one that matches where you lose the most time. For most developers that is GitHub for code or Context7 for current docs. Both are low risk when set to read-only or limited toolsets.

Pick one server today, scope it tightly, and run one real task through it. Then browse the free tools on Art of Code, and see how AI is changing the work in AI coding tools by category in 2026.

Written by Adithya Guttha, Founder of Art of Code.