MCP (Model Context Protocol) lets AI tools connect to real services — GitHub, databases, browsers, Slack. Instead of copy-pasting context, the AI can read, write, and interact with your actual tools.
Most developers haven't set this up. That's a mistake. Here are the 10 servers that matter — what each one does, how to install it, and when it actually pulls its weight.
1. GitHub MCP — The Essential
51 tools. Create PRs, review code, manage issues — from your AI chat. If you install one server, this is it.
Install: npx -y @modelcontextprotocol/server-github
Auth: Set GITHUB_PERSONAL_ACCESS_TOKEN in env.
Pulls its weight when: You're in a chat saying "update the README and open a PR" and the AI actually does it instead of telling you the steps.
2. Brave Search — Real-Time Web
Your AI's knowledge has a cutoff. This gives it live web access. Check docs, find solutions, look up current APIs.
Install: npx -y @modelcontextprotocol/server-brave-search
Auth: Free Brave Search API key, 2,000 queries/month on the free tier.
Pulls its weight when: You need the AI to verify a library's current release, a changed API signature, or an error message that only showed up in the last few months.
3. Playwright — Browser Automation
AI controls a real browser. Click, type, navigate, screenshot. Perfect for testing and scraping.
Install: npx -y @playwright/mcp
Pulls its weight when: You're writing E2E tests and want the AI to explore the app interactively first, then generate the test based on what it actually clicked. Much less hallucinated selector nonsense.
4. XcodeBuildMCP — iOS Development
Build, run, test iOS apps without touching Xcode. Essential for iOS developers using Claude Code.
Install: npx -y xcodebuildmcp@latest
Pulls its weight when: You're running xcodebuild + xcrun simctl loops to reproduce a bug. The AI can build, launch the simulator, tap through the app, and read logs without you switching windows.
5. Sentry — Crash Reports
"What's crashing in production?" becomes a question you can ask your AI instead of opening a dashboard.
Install: npx -y @sentry/mcp-server
Auth: Sentry auth token with project:read and event:read scopes.
Pulls its weight when: A support ticket drops in and you want the AI to correlate "user X reported Y" with the actual stack trace in Sentry before suggesting a fix.
6. PostgreSQL — Database Access
Query your database conversationally. "Show me users who signed up this week" — the AI writes and runs the query.
Install: npx -y @modelcontextprotocol/server-postgres postgresql://user:pass@host/db
Pulls its weight when: You're debugging data issues. The AI can SELECT its own way to the answer rather than asking you to paste query results. Pair with a read-only DB user — see the security note below.
7. Filesystem — File Access
Controlled read/write to your local files with guardrails and audit logging.
Install: npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/dir
Pulls its weight when: You want the AI to work across multiple repos or folders (notes, designs, configs) that live outside the current project. Pass explicit allowed paths — the server refuses anything else.
8. Memory — Persistent Context
The AI remembers your preferences, decisions, and project architecture across sessions. A lot of the "why did I decide X?" friction goes away.
Install: npx -y @modelcontextprotocol/server-memory
Pulls its weight when: You're returning to a project after two weeks. Instead of re-explaining the stack and past decisions, the AI already knows them because the memory server persisted the conversation's facts.
9. Slack — Team Context
Search conversations, read channels. "What did the team decide about the migration?" — answered instantly.
Install: npx -y @modelcontextprotocol/server-slack
Auth: Slack bot token with channels:history, channels:read scopes.
Pulls its weight when: You need to find "that decision we made three months ago in #engineering" without scrolling through Slack yourself.
10. Chrome DevTools — Browser Debugging
AI inspects DOM, network requests, console logs. "Why is this page slow?" gets a real, data-backed answer.
Install: npx -y chrome-devtools-mcp
Pulls its weight when: You're debugging web perf, Core Web Vitals, or a flaky network request. The AI can pull a Performance trace or a Network waterfall and tell you which specific requests blew the LCP budget.
Quick Setup
Add to your Claude Code config (~/.claude/config.json or the Cursor MCP settings):
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
},
"brave-search": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-brave-search"],
"env": { "BRAVE_API_KEY": "BSA..." }
},
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/code"]
},
"memory": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-memory"]
}
}
}
Start with GitHub + Brave Search + Filesystem + Memory. Those four cover ~80% of what most developers will use MCP for. Add the rest as specific workflows demand them.
A Security Note Nobody Tells You
MCP gives the AI actual capabilities on your systems. Treat it like any other service integration:
-
Use read-only credentials where possible. A PostgreSQL MCP connection should use a user with
SELECTonly unless you genuinely need the AI to write. -
Scope filesystem paths tightly. Pass only the directories you want the AI reading — not
/or$HOME. - Rotate tokens like you would CI tokens. Your GitHub PAT, Slack bot token, and Sentry auth token are all sitting in a config file now.
- Review what the AI did, not just what it said. MCP actions go through the normal tool-use approval flow in most clients — keep that on. The moment you auto-approve everything, a hallucinated action can run against a real system.
None of this is scary in practice, but it's worth calibrating before you wire up 10 servers with god-mode credentials.
Keep reading:
Top comments (0)