# Wiring tools to your AI (a starter)

<!--
GUIDANCE: A constitution gives your AI a personality. Skills give it routines. Tools give it hands.
This page is the hands part, in plain language, with a config you can copy.
-->

## What MCP actually is

MCP (Model Context Protocol) is a standard plug. On one side is your AI client. On the other is a tool: your files, a browser, your GitHub, your notes. The AI client speaks MCP, the tool speaks MCP, so they connect without custom glue for each one.

Think of it like USB for AI tools. You don't rewire your laptop for every device. You plug in. MCP did the same thing for AI: one socket, many tools.

Before MCP, every integration was bespoke. Now a tool ships an MCP "server" once and any MCP client can use it. That's the whole idea.

## The minimal config

Most clients keep MCP servers in one JSON file (Claude Desktop calls it `claude_desktop_config.json`; other clients use a similar `mcpServers` block). Here's a starter with two servers:

```json
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/aios"]
    },
    "fetch": {
      "command": "uvx",
      "args": ["mcp-server-fetch"]
    }
  }
}
```

That `/Users/you/aios` path is the only folder the AI can touch. Point it at your project folder, not your whole drive. Restart the client and the tools show up.

## The first servers to add, and why

Add these in order. Each one earns its place before you add the next.

1. **Filesystem** (official reference server). Read, edit and create files in folders you choose, with the path locked down. This is the one that turns a chat window into a tool that works on your actual files. Start here. Daily use. https://github.com/modelcontextprotocol/servers/tree/main/src/filesystem

2. **Fetch** (official reference server). Pulls a web page and hands the AI clean markdown instead of raw HTML soup. Lightweight. You want this the first time you paste a link and the AI says it can't open it. Daily use. https://github.com/modelcontextprotocol/servers/tree/main/src/fetch

3. **Memory** (official reference server). A small knowledge graph so the AI remembers facts and preferences between sessions instead of starting cold every time. This is what makes a second brain feel like it knows you. https://github.com/modelcontextprotocol/servers/tree/main/src/memory

4. **GitHub** (official github-mcp-server). If you keep code or version your files, this reads repos, reviews pull requests and opens issues for you. Skip it if you don't code yet. Add it the day you do. https://github.com/github/github-mcp-server

## What comes next

Two worth knowing once the first four are steady:

- **Context7** (Upstash): feeds the AI current, version-correct docs so it stops guessing old API calls. The fix for "that function doesn't exist anymore." https://github.com/upstash/context7
- **Playwright** (Microsoft): drives a real browser to test, fill forms and pull data. Around 33,500 GitHub stars by mid-2026, usually the #1 or #2 server overall. Reach for it when you need the AI to actually click through a site. https://github.com/microsoft/playwright-mcp

<!--
GUIDANCE: Don't add ten servers on day one. Each one is a new door into your stuff. Add Filesystem,
live with it for a week, then add the next. A small wired-up setup you trust beats a big one you don't.
-->