Our MCP (Model Context Protocol) server lets Claude Desktop and Claude Code use Gydmation AI as a set of tools — search nodes, generate, validate, deploy, self-heal — directly from a natural-language conversation. The interesting design decision was where the actual language model comes from for the "generate" and "fix" steps.
Rather than requiring our own hosted model or a separate API key, we use MCP's sampling capability: the tool asks the connected client to run the completion using the model the user is already signed into, then falls back to a configured API key only if the client doesn't support that.
The practical effect: generation can run on infrastructure the user already has, with nothing extra to configure and nothing routed through us — the same local-first principle behind self-hosting the product itself, applied to the model layer too.
