This session is a developer diary of my own hard lesssons learned trying to create an effective MCP server. I've gone through many wrong turns, trying to naively wrap my API, exploring different techniques that eventually evolved my MCP from a context polluter to an effective set of tools. On my journey, I've experimented with a Hypermedia approach to the tools API, some prompt injection, and learned some important lessons the hard way.

I've written a couple of blog posts relating to the topics covered in this talk:
https://towardsdatascience.com/how-not-to-write-an-mcp-server/
https://towardsdatascience.com/talk-to-my-agent/