MCP Servers Cheatsheet: Concepts, Configs, Useful Servers and Debugging

The Model Context Protocol (MCP) is the open standard that lets AI assistants call real tools — files, databases, browsers, APIs — through one plug. This page is the working reference: concepts in one table, copy-paste configs for the two clients most people use, the servers worth installing first, and how to debug the failures everyone hits. For the full story and your first custom server, read our MCP guide.

Concepts in 60 seconds

Config: Claude Desktop

File locations — macOS: ~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json. Fully quit and restart Claude Desktop after editing.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem",
               "/Users/you/projects", "/Users/you/notes"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_your_token" }
    },
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"]
    }
  }
}

Config: Cursor

Project scope: .cursor/mcp.json in the repo. Global scope: ~/.cursor/mcp.json. Same JSON shape; remote servers use url instead of command:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    },
    "my-remote-server": {
      "url": "https://mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

In Claude Code you can edit the same files or use claude mcp add name -- command… — see the Claude Code cheatsheet for the exact commands and scopes (user, project, local).

Useful servers to install first

Rule of three for beginners: start with Filesystem + GitHub + Playwright. Add a server only when you feel a specific missing capability — every server is more context, more permissions and more things that can break.

Security rules (non-negotiable)

• Whitelist narrowly: give Filesystem one project folder, not your home directory; give database servers read-only users for exploration.
• Secrets live in env, never in prompts or repos. If a config with real keys gets committed, rotate the keys — deleting the commit is not enough.
• Only install servers you can inspect. Prefer the official reference servers and well-known publishers; a random MCP server runs with your user's permissions.
• Watch tool calls: approve writes/deletes deliberately. A prompt-injected web page can try to talk your assistant into abusing its tools.

Debugging: the fixes that solve most problems

Workflow that saves hours: terminal first (does the command run?), Inspector second (do the tools respond?), client last (is the config right?). Ninety percent of "MCP is broken" is a JSON typo, a stale client that was never restarted, or an env var pasted into the wrong server block.

Building agents on top of all this? Continue with the AI agents tutorial and the rest of our cheatsheets.