MCP is your agent's hands.

Top MCP servers

MCP (Model Context Protocol) is the standard for connecting external tools to Claude. Without MCP, it works blind: no files, no repos, no database. With the right servers, Claude becomes part of your stack instead of an isolated chat.

Who this guide is for

MCP servers work in Claude Code (the terminal app for developers) and the Claude Desktop App (claude.ai/download). They don't work in the browser version of claude.ai. If you're just getting started, begin with sign-up and quick start or setting up Claude in 10 steps.

How to connect it in 5 minutes

One file, and Claude sees your whole context.

Create .mcp.json in your project root. Claude Code picks up the servers automatically on the next launch; you'll see a Connected list at the start of the session.

``json { "mcpServers": { "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/project"] }, "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." } }, "memory": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] } } } ``

npx downloads the servers automatically, so there's nothing extra to install.

Five servers we use every day

Not all MCPs are equal. These are must-haves.

Click a card and we'll show you step-by-step install instructions.

Filesystem

Access to local files and folders

INSTALLED BY 95% OF CLAUDE CODE USERS

Lets Claude read, write, and search files on disk. Without this MCP, Claude can't see your project, only whatever you paste into the chat by hand. With it, Claude works right inside your files: searching directories, reading configs.

When to get it: any project with files on disk. An Obsidian vault, a codebase, a folder of templates. It's the first MCP everyone installs.

_[How to install →]_ (expandable instructions)

Step 1: add it to .mcp.json

``json { "mcpServers": { "filesystem": { "command": "npx", "args": [ "-y", "@modelcontextprotocol/server-filesystem", "/Users/YOUR_USERNAME/Desktop/YOUR_PROJECT" ] } } } ``

Swap in your own path. You can list multiple folders in args, comma-separated.

Step 2: restart Claude Code

Close it and open it again. At the start of the session you'll see: Connected: filesystem. Done, Claude now sees all your files.

Requirements

Node.js 18+. To check: node --version in your terminal. If you're below 18, update via nodejs.org.

modelcontextprotocol/servers · github.com/modelcontextprotocol/servers ↗

GitHub

Repos, issues, pull requests, code search

Claude opens repositories, reads PRs, creates issues, and searches code across your entire org. Not just "view a file": full-blown GitHub work without a browser. It sees commit history, diffs, comments.

When to get it: code review, working with open repositories, monitoring issues, finding implementation examples in other people's projects.

_[How to install →]_ (expandable instructions)

Step 1: create a Personal Access Token

Go to github.com → Settings → Developer settings → Personal access tokens → Fine-grained tokens → Generate new token. Permissions you need: Contents (read), Issues (read/write), Pull requests (read/write). Copy the token; it's only shown once.

Step 2: add it to .mcp.json

``json { "mcpServers": { "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_YOUR_TOKEN_HERE" } } } } ``

The token lives only in your local .mcp.json; it stays out of git as long as the file is in .gitignore.

Step 3: restart Claude Code

After the restart, tell Claude: "look at the issues in OWNER/REPO" and it'll see them instantly, no browser needed.

modelcontextprotocol/servers · github.com/modelcontextprotocol/servers ↗

Memory

Persistent memory across sessions

Claude saves facts, preferences, and decisions to a knowledge graph. Next session, it remembers who you are, how you like to work, and what you've already discussed. Without it, every session = a blank slate.

When to get it: long-running projects (3+ months), personal preferences, building up knowledge of the architectural decisions you've made.

_[How to install →]_ (expandable instructions)

Step 1: add it to .mcp.json

``json { "mcpServers": { "memory": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] } } } ``

Memory is stored locally in a JSON file next to your project. Nothing gets sent anywhere.

Step 2: restart Claude Code

After the restart, Claude will automatically start saving important facts.

How to use it

Tell Claude: "remember that I prefer TypeScript and don't use Prisma" and it'll save that. Next session, it'll remember. It also works automatically: when Claude spots an important decision, it saves it on its own.

modelcontextprotocol/servers · github.com/modelcontextprotocol/servers ↗

Postgres

Direct SQL queries against your database (read-only)

Claude connects to Postgres, sees the schema, runs queries, analyzes the data. No need to copy SQL results into the chat; it figures out what's in the database on its own. Read-only by default, so it never modifies your data.

When to get it: debugging production data without exporting to CSV, checking migrations, analyzing anomalies right in the live database.

_[How to install →]_ (expandable instructions)

Step 1: find your connection string

For Supabase: open your project → Settings → Database → Connection string → URI. For local Postgres: postgresql://postgres:password@localhost:5432/mydb

Step 2: add it to .mcp.json

``json { "mcpServers": { "postgres": { "command": "npx", "args": [ "-y", "@modelcontextprotocol/server-postgres", "postgresql://user:password@host:5432/dbname" ] } } } ``

Swap in your own connection string. For Supabase, use the Transaction Pooler URL (port 6543) for better compatibility.

Step 3: restart Claude Code

Now you can say: "show me the structure of the users table" or "find all records where status = pending from the last 7 days" and Claude will write and run the query itself.

Important

The server is read-only. Claude can't modify, delete, or insert data, which protects you from accidental changes in production.

modelcontextprotocol/servers · github.com/modelcontextprotocol/servers ↗

Playwright

Browser automation and screenshots

Claude opens a browser, clicks around, fills out forms, takes screenshots, checks the UI. We use it for E2E verification of every feature right during development: Claude looks at the result itself and fixes what's off.

When to get it: checking UI changes after edits, web scraping, testing forms and flows without opening a browser by hand.

_[How to install →]_ (expandable instructions)

Step 1: install the Playwright browsers

`` npx playwright install ``

Run this once in your terminal. It downloads Chromium, Firefox, and WebKit (~300 MB). One-time setup for the whole machine.

Step 2: add it to .mcp.json

``json { "mcpServers": { "playwright": { "command": "npx", "args": ["-y", "@playwright/mcp"] } } } ``

Step 3: restart Claude Code

Now say: "open localhost:3000 and take a screenshot of the home page" and Claude will open the browser and drop the screenshot right into the chat. Or: "check that the Log In button works on /login."

How we use it at AGINE

After every UI change we tell Claude: "check that /dashboard looks right" and it opens the page, takes a look, and if something's broken, fixes it on the spot. Saves 10-15 minutes of manual checks per session.

microsoft/playwright-mcp · github.com/microsoft/playwright-mcp ↗

Finding and connecting MCPs

Didn't find yours? Type in the service name.

There are already hundreds of MCP servers: for Notion, Slack, Linear, Figma, and thousands of other services. Tell us what you work with, and the AI will find the right server in real time, hand you a ready-made config with keys, and give you step-by-step connection instructions.

[McpFinder component: interactive search]

• Input field with placeholder: Service name: Notion, Slack, Linear…
• Button: Find MCP (while running: "Searching…")
• Examples: Notion · Slack · Linear · Google Drive · Figma
• Loading state: "The AI is searching for an up-to-date MCP server in real time; this takes a few seconds."
• Errors: "Enter a service name (at least 2 characters)." / "The search couldn't be completed." / "Network unavailable. Try again."
• Result (found): an "official" or "community" badge, an .mcp.json block with a "copy" button, an "Access keys" section, a "How to connect" section (numbered steps), a "Verify the package at the source" link.
• Result (not found): "No ready-made MCP server turned up for this service. Check the official MCP registry." + an "Open the MCP registry" link.
• Result footer: "This result was assembled by a real-time search. Before pasting in any keys, open the source and make sure the package is the one you actually need."

Our actual config

What's running at AGINE right now.

• → filesystem: pointed at the project folder + the agine-knowledge Obsidian vault. Claude reads the full context of past sessions and decisions made.
• → github: for working with repositories. Claude Code creates PRs, reads issues, runs code search.
• → playwright: every UI change gets verified with a screenshot right in the session, before git push. Claude checks the result itself.
• → postgres: debugging Supabase without hand-copying data out of the dashboard.
• → memory: Claude Code's built-in auto-memory saves important facts and decisions across sessions automatically.

All the servers live in a single .mcp.json in the project root. None of them needs extra infrastructure: just npx and access keys.

Once you're connected

Ask the AI what to automate.

Your tools are connected; don't guess what comes next. Give the AI this prompt: it will study your real data, ask clarifying questions about your goal and metrics, suggest what to optimize and automate, and immediately execute the one task with the biggest impact.

[CopyPrompt component: a prompt to copy] (label: "prompt: copy and paste into Claude")

``` You're connected to my work services. Follow these steps exactly:

1. Look at which services I have connected and skim what real data is in them. Don't assume anything; rely only on what you actually see.

2. Before recommending anything, ask me clarifying questions and wait for my answers: • what I do (my own business, freelance, a project) and what my role is; • what my main goal is right now (grow revenue, save time, scale, cut costs); • which 2-3 key metrics I track (revenue, profit, leads, customers, conversion, costs); • which routine eats up most of my time. Phrase the questions based on what you've already seen in my data.

1. Only after I answer, analyze the data in depth against my goal and tell me: 1) what can be optimized to reach the goal faster? 2) what can be automated to free up my time? Back every point with specifics from my services.

4. Pick the one task with the biggest impact on my goal and do it right now. ```