Hosts, clients, and servers
MCP defines three roles. The host is the actual AI application the user interacts with — Claude Desktop, Claude Code, or a custom agent app. The host contains one or more MCP clients, and each client maintains a dedicated, one-to-one connection to a single MCP server. If a host wants to talk to three different servers — say a filesystem server, a database server, and a project-management server — it runs three separate client connections internally, one per server, rather than one client juggling three servers at once.
The server is the thing exposing capabilities: it might be a thin wrapper around an existing API, a local process reading files on disk, or a hosted service fronting a company's internal systems. Servers don't know or care which host they're talking to, and hosts don't need server-specific code — both sides only need to speak MCP correctly. This is what makes the one-to-one client-server pairing, multiplied across however many servers a host wants, still add up to constant integration work per side rather than growing with every new combination.