MCP Servers
MCP is the interface between a model and a system that holds real data. A server exposes a set of typed tools; the model chooses among them and acts through them. I build these for systems I use daily, which means the design questions — authentication, tool granularity, and how much surface a model can hold before it starts choosing badly — are ones I have had to answer rather than read about.
How I use it
Work history as queryable tools
A server exposing career data as structured tools rather than a document. The difference is that a specific question gets a specific answer from the underlying records, instead of a model paraphrasing a PDF and inventing the parts it cannot find.
A read-and-write tool surface over live data
A tracker exposing sessions, bankroll and opponent records through tools that both read and mutate state. Write access changes the design problem — every mutating tool needs its parameters constrained so a plausible-looking call cannot corrupt real records.
Tool granularity
The decision that determines whether a server is usable. Too many narrow tools and the model cannot reliably choose between them; too few and every call needs a paragraph of arguments it will get wrong. It is API design where the consumer re-reads the documentation on every single call.
Scoping what a server can reach
A server runs with real credentials against real systems. What it is permitted to touch is a security boundary, and it belongs in the server rather than in an instruction telling the model to behave.
What I use
- Model Context Protocol
- TypeScript
- Claude API
Related skills
Want the detail behind any of this?
I can walk through any of it in as much depth as you want.