Model Context Protocol

MCP, explained for busy people.

MCP is a standard plug between AI apps and the rest of your world. Out of the box an AI can only talk. Plug in an MCP server and it can open your browser, read your GitHub, write to Notion or query your database.

Get the config for my appHow it works
The idea

Three pieces, one plug.

  1. The app (host)

    What you already use: Claude, ChatGPT, Cursor, VS Code, Codex. It speaks MCP.

  2. The server

    A small program that exposes tools ("click this", "create issue") and data ("this page", "that table"). Someone like GitHub or Notion writes it once.

  3. You just ask

    "Open our staging site and screenshot the checkout." The AI picks the right tool, asks your permission, and uses it.

Before MCP, every AI app needed its own custom integration with every tool. Now one server works in every app, like USB for AI.

AI app Claude · ChatGPT Cursor · Codex MCP PlaywrightMCP server GitHubMCP server NotionMCP server Browser Repos Pages
One app, many servers. Each server is a separate plug you add or remove.
The one distinction that matters

Local vs remote servers

Local (stdio)

runs on your computer

Your app launches it as a small program, usually with npx (needs Node.js) or uvx (Python). Good for files, your browser, local databases.

Works in: Claude Code, Claude Desktop, Codex, Cursor, VS Code, Gemini CLI. Not in claude.ai or ChatGPT on the web.

Remote (Streamable HTTP)

just a URL

Hosted by the company (GitHub, Notion, Sentry…). You paste a URL and sign in with your account. Nothing to install, always up to date.

Works in: everything, including claude.ai and ChatGPT. If a service offers a remote server, prefer it.

Config generator

Pick a server, pick your app, copy.

Every app wants MCP configured slightly differently: a command here, a JSON file there, TOML in Codex. Choose both and get the exact thing to paste.

1 · Server

2 · Your app

Traps

Things that trip everyone up

"servers" vs "mcpServers"

VS Code's .vscode/mcp.json uses "servers". Claude, Cursor and Gemini use "mcpServers". Wrong key = silently nothing.

Restart the app

Claude Desktop only reads its config on start. Fully quit it (tray icon too), then reopen.

Codex speaks TOML

~/.codex/config.toml uses [mcp_servers.name] blocks, not JSON. Or skip the file: codex mcp add.

Web apps = remote only

claude.ai and ChatGPT can't start programs on your PC. They need a server with a public URL.

Gemini's "url" is SSE

In settings.json, "url" means the old SSE transport; use "httpUrl" for modern HTTP servers.

SSE is deprecated

If a guide says --transport sse, it's outdated. The current spec (2026-07-28) uses stdio and Streamable HTTP.

Windows paths

Inside JSON, write C:\\Users\\me with double backslashes, or use forward slashes.

Too many servers

Each server adds tools the AI has to read. Ten servers slow it down and confuse it. Keep what you use.

Safety

An MCP server can do what you can do.

It runs with your permissions and your logins. Treat installing one like installing an app.

Prefer official servers from the company behind the product (github/, microsoft/, notion…). They're marked "official" below.
Give the smallest access. Read-only tokens, one folder for the filesystem server, a test database first.
Watch for prompt injection. A web page or GitHub issue can contain text aimed at your AI ("ignore your instructions and…"). Keep "ask before running tools" on for anything that writes or sends.
Never paste API keys into chats. Put them in the config's env or let the server do OAuth sign-in.
Popular connectors

Start with these

Every connector we list, with what it's for and the setup steps, is in the tool finder.

Browse connectors

Find more

Official MCP Registry

preview

The protocol's own index of public servers. Mostly used by marketplaces, but browsable.

awesome-mcp-servers

list

The biggest community list, sorted by category. Quality varies: check the repo first.

modelcontextprotocol.io

docs

The spec and official guides, if you want to build your own server.