Recipe 016

Don't Install 20 MCP Servers. Install These 3.

The beginner mistake with MCP isn't installing too little — it's installing a dozen servers and wondering why the agent got slower and dumber. Here's the starter pantry: the three servers this kitchen actually runs every day, what each one earns, and what we deliberately left out.

After last recipe — where we untangled MCP from API — the obvious next question showed up in our own kitchen: fine, MCP hands my agent appliances. Which appliances first?

And here's the trap waiting for that question. You open an MCP directory, find thousands of servers, and the collector instinct kicks in. Weather! Databases! Spreadsheets! Calendars! A dozen installs later your agent is… worse. Slower. Oddly distracted.

That's not bad luck. Every installed server's tool descriptions ride along in every conversation you have. Twenty servers means your agent reads a wall of appliance manuals before it reads your question. A pantry stuffed with unopened jars isn't abundance — it's clutter that slows down dinner.

If you've been running MCP servers for months, this recipe isn't for you. This is the starter pantry: the three servers we actually run in this kitchen, every day, operating this very blog. Each one has earned its shelf space.

1. Playwright MCP — eyes and hands

The single biggest upgrade for a new agent. Playwright MCP is Microsoft's official server that gives your agent a real browser: it opens pages, reads their structure directly, clicks, types, takes screenshots.

Why it's first: without it, anything involving a website means you describing the page to the agent. With it, the agent looks for itself. Every layout fix on this blog is verified by the agent opening the live site — we wrote a whole recipe about it.

claude mcp add playwright npx @playwright/mcp@latest

First test-drive: "Open example.com and tell me what you see." Honest note: an agent with a browser is an agent with hands — keep approval mode on until you trust the routine.

2. Context7 — documentation that isn't a year old

Here's the beginner moment nobody warns you about. The agent writes code confidently, the code fails, and you can't tell whether you broke it. Surprisingly often, neither of you did — the library changed after the model's training data was collected, and the agent cooked from an old edition of the cookbook.

Context7 (by Upstash, free tier, no key required to start) fixes exactly this: it pulls current, version-specific documentation straight into the conversation when the agent needs it. Fewer invented function names, fewer "that option was renamed two versions ago" mysteries.

claude mcp add context7 -- npx -y @upstash/context7-mcp

First test-drive: ask about whatever you're building with — "look up the current way to define routes in Astro, and tell me if it differs from what you remembered." Watching the agent correct its own stale memory is worth the install alone.

3. claude-video-vision — YouTube into project knowledge

The wildcard of the three, and honestly it arrives as a plugin that ships MCP tools rather than a bare server — install it from the marketplace inside Claude Code (/plugin → search "video vision"). It needs ffmpeg, plus yt-dlp for YouTube.

What it earns: your agent can watch video — pull captions, check metadata, extract frames when visuals matter. We use it to turn a marketing lecture series into timestamped study notes that then update this project's standing instructions. A YouTube video literally changes how our agent writes posts. Full pipeline here.

If your learning diet includes YouTube — and whose doesn't — this is the difference between watching and owning.

What we deliberately left out

The starter pantry is defined by what's not in it:

  • Filesystem servers — Claude Code already reads and writes your files natively. Redundant jar.
  • GitHub servers — the gh command line does everything the agent needs, without spending context on tool descriptions.
  • Memory servers — at this stage, a good CLAUDE.md beats them: simpler, visible, yours.

One rule going forward: when in doubt, don't install. Add a server the day a task actually demands it — not the day a listicle mentions it. (Yes, we see the irony. This is the listicle. Install three.)

Order up

Paste this into Claude Code:

Set up my starter MCP pantry, step by step.

1. Run `claude mcp list` and tell me what I already have installed.
2. Give me the exact commands to add the Playwright browser server and
   Context7, and tell me when to restart you.
3. After I restart you: prove each works — open example.com and describe
   what you see, then look up current docs for [a tool I'm using] and
   note anything that differs from your memory.
4. Standing rule: before I install any other MCP server, tell me what it
   adds and what it costs in context — then let me decide.

Step 4 is the real recipe. The pantry stays useful exactly as long as everything in it earns its shelf.

What's in your pantry — and what did you uninstall? Tell us at @AgentKitchenHQ.

Hungry for more?

New recipes land in your inbox as they leave the kitchen. No spam, unsubscribe anytime.