MCP Servers Implementation cookbook

Another toy demo. Still not one real agent actually calls.

You wire your server straight into Claude Desktop, a tool call fails, and there's no way to tell if the bug is your code, your schema, or the client itself.

Why does a tool that works fine called directly get ignored the moment I ask normally?

One expected tool failure, and the whole session dies right there.

My server has a lookup the model calls before almost every real request, one extra round trip apiece, for a value that barely changes from one conversation to the next.

Four MCP servers connected, and the agent suddenly fumbles tool choices it used to get right.

Your coding agent guesses the database type or the auth method, gets it almost right, and you fix it after the fact every time.

My MCP server only works because a client launches it as a subprocess on my own machine, and now three teammates need to reach the same one.

Moved behind a load balancer, and a long tool call vanished mid-run.

Your tool takes two minutes to finish, and the client gives up long before it's done.

My MCP server holds one login to a real API on behalf of every user, with no check stopping it handing that access to the wrong one.

A teammate wants to add someone else's MCP server, and I have no checklist for whether it's safe.