From a ChatGPT brief to a local CAD model
Trace the conversation, MCP tools, isolated Python worker, native kernel and viewer through one controlled modeling request.
When you ask Intact for a part, several distinct systems participate. ChatGPT interprets your request. The plugin exposes engineering tools. A CAD engine constructs geometry. A local viewer displays the exported representation. Understanding those responsibilities makes it easier to write a useful brief and diagnose a surprising result.
The conversation defines the task
Begin with purpose, units, dimensions, interfaces and the deliverable. For an enclosure, specify outside size, wall thickness, lid arrangement and mounting features. Call out dimensions that can be assumed and dimensions that must be clarified. Ask for separate named components where you need separate parts.
ChatGPT uses the installed plugin’s workflow instructions and tool descriptions to choose operations. The assistant’s explanation is not a substitute for a measured result. Ask it to retain checkable requirements and report what was actually verified.
MCP connects intent to operations
The Model Context Protocol interface exposes typed engineering operations. Requests include settings and, for existing designs, the current model and revision. Input checks and revision guards help prevent a change based on stale geometry from overwriting a newer design.
For native CAD, a bounded Python program uses build123d and OpenCASCADE. For explicit sketch relationships, the structured route uses FreeCAD constraints and supported feature operations. Sampled TPMS geometry follows a mesh path. The representation determines which measurements and exports are meaningful.
Generated Python runs behind a boundary
The supported runtime uses Linux namespaces and bubblewrap, with filesystem/network isolation and resource limits. Program policy checks supplement this operating-system boundary. CAD generation is disabled when required isolation is unavailable; installation problems should be repaired without bypassing it.
The worker builds a candidate, inspects its geometry and generates artifacts. Audits and saved design requirements are evaluated before publishing an accepted revision. If the build fails, the last accepted design remains available. Source repair should address the failing feature, not change the brief merely to make a check pass.
The viewer closes the loop
Orbit, section and inspect the result. Select a current part, face or edge and ask a specific question or edit. The selected reference carries source identity so the system can detect staleness instead of guessing which later face has the same number.
A small follow-up can be precise: “Move this mounting boss 3 mm in +X. Preserve the bore diameter and the other bosses. Verify the resulting center positions.” For an uncertain idea, request a parameter preview before saving it.
Where your account and computation fit
Your own ChatGPT account provides the AI conversation; the installed local runtime performs CAD and supported numerical work. The current package is a personal plugin installation, not an automatic download triggered by opening ChatGPT. No separate model-service API key is required for the supported signed-in workflow, while account availability and usage limits still apply.
Projects remain local, but selected requests and relevant model context pass to the AI provider through the client. Review the data-flow explanation before using confidential material.
Read Architecture for the full system map and Get Started for setup.
Install once.
Design through conversation.
Download Intact v0.20.0 for Windows. Install privately in Codex with your own account; CAD runs on your computer.
Follow the installation procedure →