MCP Servers
Connective tissue between assistants and systems that were never built for them.
Tooling for AI
ongoing
MCP, TypeScript, Python, REST, JSON Schema
Most of what I build now is tooling for AI rather than with it. MCP servers are the layer that decides what a model can see and what it can do, and almost all of the difficulty lives there rather than in the model.
I have built servers over my own university course material, over investment and market data, and a conversion layer that takes an existing REST API and turns it into a working MCP server, so a backend nobody designed for models can be used by one without a rewrite.
The recurring questions are the same every time. What is the smallest tool surface that still does the job. How precise does a schema have to be before a model stops guessing. Where does the rate limit go. Which operations should never be exposed at all.