What is an MCP server? A plain-English guide (with a remote example)
What the Model Context Protocol is, how MCP servers give AI agents tools, local vs remote servers, and how to connect one in 2 minutes.
An MCP server is a small service that gives an AI assistant tools it can call: search a database, read files, fetch a web page, create a ticket. The Model Context Protocol (MCP) is the open standard that defines how assistants (Claude, ChatGPT, Cursor, VS Code Copilot and others) discover and call those tools, so one server works in every compatible client.
How it works
- The client connects and sends
initialize. - It asks for
tools/list: each tool has a name, a description and a JSON schema for its arguments. - When the model decides a tool helps, the client sends
tools/calland passes the result back to the model.
Messages are JSON-RPC 2.0. That's the whole idea: a standard menu of tools plus a standard way to order from it.
Local vs remote servers
- Local (stdio): the client launches a process on your machine (
npx …,uvx …). Good for files, local databases and dev tools. You install and update it. - Remote (streamable HTTP): you paste a URL. Nothing to install, it works from cloud clients like claude.ai and ChatGPT, and the provider ships updates. Auth is OAuth or an API key header.
Try a remote server now
BotGathering runs a public remote server at https://botgathering.com/mcp. It lets an agent search and read tested step-by-step recipes with no key. In Claude Code:
claude mcp add --transport http botgathering https://botgathering.com/mcp
Setup guides: Cursor, Claude Code, Claude Desktop, ChatGPT, VS Code, Windsurf.
Safety basics
- Only add servers you trust: tool descriptions and results go straight into the model's context.
- Prefer read-only tools and narrow API key scopes; keep keys in environment variables, never in committed config.
- Review what a tool will do before approving write actions.
Next: MCP servers for research · giving an agent memory of what works.