MCP Server Skill — Optimize Existing or Architect New

This skill is the entry point for any MCP-server work that is bigger than one line of code. It helps you either tune up a server that already exists or shape a brand-new one before a single file is scaffolded. Both paths share the same reference library — one folder of decision trees, one folder of pattern files — so you only learn the routing once.

Two things to remember before anything else:

1. Do not apply patterns mechanically. Explore the repo (or the user's intent) first, present findings with options, let the user choose.
2. You are not expected to do it all alone. If a question needs fresh research or a big chunk of file work, dispatch a subagent — see "Dispatching Subagents" below.

Two Modes — Pick One Before You Do Anything

Mode A — Optimize an existing MCP server

Pick Mode A when the user already has an MCP server (their repo, a fixture, a package) and wants it improved, audited, hardened, or reviewed. The rhythm is Explore → Diagnose → Present → Optimize, and it never skips straight to code.

1. Explore the codebase. Before you ask a single question, read the repo. Look for an MCP entry point, tool registrations, schemas, error handling, and transport config. Useful starting searches: