how are you wiring MCP into Claude Code without it turning into a mess?

ive got maybe 5 MCP servers connected now and im starting to feel the downside. every session Claude has access to all of them and sometimes it reaches for the wrong tool, or burns tokens poking at one i didnt ask about. my current setup is basic. Claude Code in the terminal, a postgres MCP for my db, a filesystem one, and a couple api wrappers. works, but it feels like i bolted it together without a plan. what im trying to figure out: do you keep all servers on all the time, or enable per project? do you w