You do not need 50 MCP servers. You need the few that match your work, set up properly, and nothing extra eating your AI's memory.
This guide gives you 9 that are worth it in October 2026, with the exact command to add each one to Claude Code, when to use it, and what it needs. The same servers work in Cursor and Codex; each official README has those configs.
Install what you use. Turn off the rest.
// Build this with your AI agent
Build this with your AI agent
Want to try this? Copy the setup prompt into your coding agent with this guide's link.
Read this guide: https://codewithnishant.dev/free/mcp-servers-setup-guide/ and the official READMEs it links to. Help me set up the MCP servers from the guide that fit my work. First ask me which of the 9 I need (or suggest 2 or 3 from what you see in my project), and check what is already installed with claude mcp list so you do not add duplicates. Install each one with the command from its official README, let me type any API key or token myself in my own terminal, and never print or store secrets. Start database servers read-only. After each install, run a small test that proves it works (for example, ask Context7 for a library's current docs, or open a page with Playwright). Then show me the final list, how to use each one, and anything still incomplete. Ask before paid services or publishing.
01. What an MCP server is
One line
An MCP server is a plugin for your AI. It lets Claude Code, Cursor or Codex use a real app: your browser, GitHub, Figma, your database. MCP (Model Context Protocol) is the open standard they all speak, so one server works in many tools.
The catch: every server you add puts its tool descriptions into the AI's context before you type anything. Too many servers means less room for your actual work. That is why section 04 matters.
02. Which one for which job
Pick by the problem
| Your problem | Install |
|---|---|
| Your AI writes code for an old library version | Context7 |
| A page is broken in the browser | Chrome DevTools MCP |
| Test a flow like sign-up or checkout | Playwright MCP |
| Read issues, open PRs, check CI | GitHub MCP |
| Big codebase, AI wastes context | Serena |
| Turn a Figma design into code | Figma Context MCP |
| Research or pull content from websites | Firecrawl MCP |
| Check data or schema in your app | Supabase MCP |
| Manage DNS, Workers or R2 on Cloudflare | Cloudflare MCP |
03. Setup for all 9
Copy, paste, done
Run these in your terminal, not in a chat. Replace anything in CAPITALS with your own value. After adding, run claude mcp list to check it is connected, or /mcp inside Claude Code.
1. Cloudflare MCP
What it does: the whole Cloudflare API (DNS, Workers, R2, Zero Trust and more) through one small server. Official, by Cloudflare, checked 23 October 2026.
Use it when: You run anything on Cloudflare and want your AI to read or change it.
Claude Code
claude mcp add --transport http cloudflare https://mcp.cloudflare.com/mcpLog in with Cloudflare when your client asks (OAuth). For scripts or CI, use a Cloudflare API token as a Bearer header instead. Official repo: github.com/cloudflare/mcp
2. Context7
What it does: fresh, version-correct docs for the library you are using, so your AI stops guessing old APIs. 63.5K GitHub stars, checked 23 October 2026.
Use it when: Your AI keeps writing code for an old version of a library.
Claude Code
npx ctx7 setup --claudeThe setup logs you in (OAuth), makes an API key and installs it for you. Use --cursor or --opencode for those tools. Manual route: the server URL is https://mcp.context7.com/mcp with your key as a Bearer header. Official repo: github.com/upstash/context7
3. Chrome DevTools MCP
What it does: your AI opens Chrome and reads the console, network calls and performance of a live page. 54K GitHub stars, checked 23 October 2026.
Use it when: A page is broken in the browser and you want the AI to see the real error.
Claude Code
claude mcp add chrome-devtools -- npx -y chrome-devtools-mcp@latestNeeds Node.js and Chrome. No key. Official repo: github.com/ChromeDevTools/chrome-devtools-mcp
4. Playwright MCP
What it does: your AI clicks, types and moves through your app in a real browser, like a tester. 38.5K GitHub stars, checked 23 October 2026.
Use it when: You want the AI to check a sign-up flow or a form end to end.
Claude Code
claude mcp add playwright npx @playwright/mcp@latestNeeds Node.js. No key. Official repo: github.com/microsoft/playwright-mcp
5. GitHub MCP
What it does: issues, pull requests, code search and CI runs, straight from chat. 33.9K GitHub stars, checked 23 October 2026.
Use it when: You want the AI to read an issue, open a PR or check why CI failed.
Claude Code
claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'Make a GitHub personal access token with only the access you need and paste it in place of YOUR_GITHUB_PAT in your own terminal. GitHub's guide notes the older claude mcp add --transport http form if add-json fails on Windows. Official repo: github.com/github/github-mcp-server
6. Serena
What it does: edits code by symbol (functions, classes) instead of text search, so big codebases cost fewer tokens. 30.6K GitHub stars, checked 23 October 2026.
Use it when: Your project is large and the AI wastes context reading whole files.
Claude Code
uv tool install -p 3.13 serena-agent
serena init
claude mcp add --scope user serena -- serena start-mcp-server --context claude-code --project-from-cwdNeeds uv (Astral). Serena's README says not to install it from an MCP marketplace, because those listings use outdated commands. Official repo: github.com/oraios/serena
7. Figma Context MCP (Framelink)
What it does: gives your AI the real layout data of a Figma frame, so it builds the design instead of guessing from a screenshot. 16K GitHub stars, checked 23 October 2026.
Use it when: You have a Figma design and want it turned into code.
Claude Code
claude mcp add framelink -- npx -y figma-developer-mcp --figma-api-key=YOUR_FIGMA_KEY --stdioNeeds a Figma personal access token. Then paste a Figma frame link into your chat and ask the AI to build it. Official repo: github.com/GLips/Figma-Context-MCP
8. Firecrawl MCP
What it does: turns any website into clean text your AI can read, plus search and crawl. 7.7K GitHub stars, checked 23 October 2026.
Use it when: You want the AI to research a site or pull content from pages.
Claude Code
claude mcp add firecrawl -e FIRECRAWL_API_KEY=fc-YOUR_KEY -- npx -y firecrawl-mcpNeeds a Firecrawl API key (it has a free tier; check its pricing page). Firecrawl also runs a hosted server at https://mcp.firecrawl.dev/v2/mcp. Never put the key in the URL or in a chat. Official repo: github.com/firecrawl/firecrawl-mcp-server
9. Supabase MCP
What it does: query your database, manage tables and read logs by chat. Official, by Supabase, checked 23 October 2026.
Use it when: Your app runs on Supabase and you want the AI to check data or schema.
Claude Code
claude mcp add --transport http supabase "https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_REF&read_only=true"Log in with Supabase when asked. Start with read_only=true and one project_ref, as Supabase's security notes recommend, before you let the AI change anything. Official repo: github.com/supabase-community/supabase-mcp
04. The token-saving setup
Why MCP got way better this year
The old way loads every tool into the AI's context. Cloudflare measured what that costs for its own API, which has more than 2,500 endpoints:
| Setup | Tools | Tokens |
|---|---|---|
| Every endpoint as its own tool (full schemas) | 2,594 | about 1,170,523 |
| Every endpoint, required fields only | 2,594 | about 244,047 |
| Code Mode | 3 | about 1,100 |
These numbers are Cloudflare's own, from the cloudflare/mcp README. Code Mode works like this:
- your AI searches the API for the one endpoint it needs,
- writes a tiny program to call it,
- the program runs in a sandbox and only the answer comes back.
Cloudflare's blog post, Code Mode: give agents an entire API in 1,000 tokens, explains the idea. For your own setup, three habits keep things lean:
- Only add what you use. Remove a server with
claude mcp remove NAME. - Prefer servers that load few tools, like Cloudflare's Code Mode server or Serena's symbol tools.
- Turn on feature groups where a server offers them, like Supabase's
featuresoption.
05. Common problems
Quick fixes
| What happens | Try this |
|---|---|
npx not found | Install Node.js, then open a new terminal |
| Server added but the AI does not use it | Run claude mcp list to check it connected, then ask for it by name in your prompt |
| 401 or permission errors | Your key or token is missing access. Make a new one with the scopes the README lists |
| Serena install fails | Install uv first, and use the command from its README, not a marketplace listing |
| The AI feels slow or forgets things | Too many servers. Remove the ones you are not using this week |
06. FAQ
What is an MCP server?
An MCP server is a plugin that lets your AI assistant use a real app or tool, like your browser, GitHub, Figma or your database. MCP (Model Context Protocol) is the open standard Claude Code, Cursor, Codex and other tools use to connect to them.
Which MCP server should I install first?
Start with the one that matches your daily work. For most builders that is Context7 (fresh docs) and Playwright or Chrome DevTools (see the browser). Add GitHub when you work with pull requests, and Supabase or Figma only if you use them.
Do too many MCP servers slow my AI down?
Yes. Every server adds its tool descriptions to the AI's context, which uses tokens before you ask anything. Cloudflare measured its full API as about 1.17 million tokens when every endpoint is a tool. Keep only the servers you use, and turn the rest off.
What is Code Mode in MCP?
Code Mode is Cloudflare's way of giving an AI a huge API cheaply. Instead of loading 2,500+ endpoints as separate tools, the server gives the AI a few tools to search the API and run a small program in a sandbox, and only the result comes back. Cloudflare puts that at about 1,100 tokens.
Are MCP servers safe?
They can read and change real things, so treat them like any app you give access to. Use tokens with the least access you need, start database servers in read-only mode, install from the official repo, and never paste keys into a chat.
Do these work in Cursor and Codex too?
Yes, all nine are standard MCP servers. The commands here are for Claude Code. Each official README, linked below, has the config for Cursor, Codex and other tools.
07. Sources and what I checked
Checked 23 October 2026
Commands and star counts change. Recheck the README before you rely on a command. These are the primary sources for this guide.
- cloudflare/mcp and Cloudflare's Code Mode post: the server URL, OAuth and token setup, and the token table.
- upstash/context7:
npx ctx7 setupand the manual server URL. - ChromeDevTools/chrome-devtools-mcp: the npx config.
- microsoft/playwright-mcp: the Claude Code command.
- github/github-mcp-server and its Claude install guide: the remote server and
add-jsoncommand. - oraios/serena and its client docs: the uv install and Claude Code command.
- GLips/Figma-Context-MCP: the npx config and Figma token.
- firecrawl/firecrawl-mcp-server: the npx command and hosted server.
- supabase-community/supabase-mcp and Supabase's MCP docs: the server URL, read-only mode and security notes.
What I checked, and what I did not
I read each official README on 23 October 2026 and took every command from it. Star counts are from GitHub on the same day. I did not run all nine servers end to end for this guide, and I give no speed or quality numbers of my own. The token figures are Cloudflare's. Want my older picks too? See Best MCP servers for vibe coders and 5 Claude MCPs to set up.
Install what you use. Turn off the rest. Your AI gets smarter when it has room to think.
// Free newsletter
I send out guides like this every week
Real setups, real sources, no hype. Drop your email and I'll send you the next one.