Working with AI¶

The command pipeline the UI uses is open to an LLM, so an assistant builds and edits models with the exact operations a person would, and every change is logged with the actor that made it.

In-app chat¶

The chat panel is docked at the bottom of the viewport, above the bottom bar on mobile. It streams responses and updates the model live as commands are applied. Behind the scenes it calls the curated command catalog through the same dispatch() path as the UI, and can take a screenshot of the viewport (including stress results) to check its own work. On mobile the assistant is the only editing path, since the tree and tools are read-only there.

Every command shows up as a compact trace line. When the same command runs several times in a row (five new parameters, say) the lines collapse into one row like parameter_create ×5 that you can expand, with any failures marked in the header. Screenshots stay on their own line.

For anything beyond a couple of features the assistant writes itself a build plan first, which appears as a Plan · 2/5 done row you can click open to read the steps and see which are finished. It updates as it works.

It can inspect an assembly as well as build one: ask whether parts collide and it runs the same interference check you would, reading back which bodies overlap and by how much. It can also combine bodies with union, cut or intersect.

You can attach reference images to a message, by drag and drop onto the chat panel or the attach button, and ask it to work from them ("model this bracket as shown"). Up to four PNG, JPEG or WebP images (max 5 MB each) ride along with your prompt, are saved with the conversation, and are shown again when you reopen it.

Ask how to do something in the interface ("how do I extrude?") and the assistant looks the answer up in this documentation and links the page it used. It can also highlight the matching control with a pulsing ring, using the same mechanism as the guided tutorial.

The guided tutorial drives desktop toolbar and sidebar controls, so it is desktop only. Launching it on a phone shows a short card pointing you to a larger screen and leaves the chat untouched; start it from the Guided tutorial card on your dashboard.

Asking for variants¶

To get several versions of one part, just ask: "make a Small and a Large variant", or "add a version without the mounting tabs". The assistant creates a named configuration per variant and stores each difference as a scoped override (a changed parameter, or a feature forced on or off) instead of duplicating the part, then switches between them. Manage the results in the Configurations panel.

Migrating old projects¶

If you open a .n3xd file this version can no longer load, the failure toast offers Try to migrate. It opens a fresh project and hands the old file to the assistant, which reads its stored feature history and rebuilds the model command by command. This is best-effort: imported STEP/STL/IGES bodies are not reproduced, and edge-based picks (fillets, chamfers) may land differently on the rebuilt solid, so check the result against the original. The assistant reports anything it could not recreate. Migration needs the AI assistant, so it is available in cloud/SSO mode only.

This applies only to a file this version cannot read. A file whose project id already belongs to another account fails for a different reason, so that toast offers Import as copy instead.

Driving N3XD from outside¶

For automation and external assistants, N3XD ships an MCP server exposing the same catalog and an execute_command tool, plus read tools for the tree, topology, measurement and more. Point an MCP client at it to inspect and modify projects programmatically.

The chat, the REST API and the MCP server are three front doors onto one pipeline, so none of them can drift from the others.