Member-only story
Stop installing MCP servers for everything. After shipping all three in production, here’s the decision framework that saved 60k context tokens.
9 min read
Just now
--
Press enter or click to view image in full size
I had 12 MCP servers installed. My Claude Code session was eating 67k tokens before I typed a single prompt. That’s 33% of my context window gone, just to register tools I touched maybe twice a week.
So I rebuilt my setup from scratch. Most of what I had as MCP servers should have been skills. A few should have been subagents. And only three actually needed to be MCPs. The result: 14% startup overhead, sharper tool selection, and zero “Claude forgot what we were doing” moments mid-task.
This post is the decision framework I wish someone had handed me six months ago.
Table of Contents
· Why this matters now
· What does each one actually do?
· Skills are markdown runbooks
· Subagents are isolated Claude sessions
· MCP servers are live sockets to other systems
· How do you decide which to reach for?
· When should you build an MCP server instead?
· How do subagents actually save context?
· What does a production setup look like?
· Reality check: what still breaks
· FAQ
∘ When should I build an MCP server instead of a…