From chat to CAD: the 3-layer architecture behind a design assistant you can trust

You cannot just plug ChatGPT into Inventor and hope for the best.

I know because I built a conversational assistant that creates parametric parts and assemblies inside Autodesk Inventor 2026, from plain natural language. And the most important lesson was not about AI. It was about architecture.

The 3 layers

Layer 1. The LLM: understands intent. "Create a base plate, 200 x 200, with 4 mounting holes". The model interprets what the engineer wants: dimensions, components, relationships. It decides WHAT. Never HOW.

Layer 2. MCP: the contract. The Model Context Protocol exposes a fixed catalog of tools with typed parameters. The model can only call what exists, with the parameters that exist. No hallucinated API calls, no improvising on your file.

Layer 3. The Inventor API: executes. Real, parametric, editable geometry. Components resolved against the Content Center, assembly templates, constraints. What Inventor knows how to do, done by Inventor.

The key design decision

The LLM never touches the CAD API directly. It talks to a contract. If it asks for something that does not exist, the call does not happen. If it passes an invalid parameter, the tool layer rejects it before it ever reaches the 3D model.

That is the difference between a demo that impresses on Tuesday and breaks on Wednesday, and a tool an engineer trusts on a deadline.

The output matters as much as the path

The result is not a mesh or a dumb solid. It is native parametric Inventor geometry you can keep editing like you modeled it yourself. Because you did. You just described it first.

That same recipe (LLM for intent, MCP as contract, native API for execution) is what I use in VaultPilot to operate Vault by conversation, and it applies to virtually any engineering software with an API. The pattern travels. Improvisation does not.