I developed a small MCP server that exposes an FME workbench (.fmw) as a service that can be invoked via natural language, using an AI client (e.g., Claude).
The interesting aspect is the architecture and the simplicity of "assembling" the various components:
1) I build the FME flow that executes the workflow, managing logic, transformations, and formats (we leave to FME what belongs to FME),
2) the MCP server exposes operations and parameters in a controlled manner,
3) the AI client interprets the request in natural language and translates it into a structured call.