AI CAD: edit 3D models by voice

For a designer, the value is simple: keep your eyes on the model and describe the next edit out loud.

The problem: the edit is clear before the clicks are

The change is already clear in the designer's head. Put a port here. Center the vents. Make the walls taller. A conventional CAD interface turns each idea into a sequence of tools, constraints, and numeric fields before the model can change.

Those controls aren't going away. A hole still needs a diameter, and a wall still needs a height. But if you've already selected the right surface, why describe its location again in a chat box? AI CAD lets that selection do its job while voice carries the change.

What Vaibhav built

AI CAD connects a live voice agent to a visual editor. A click tells the app where the edit belongs; a spoken request tells it what to change. The designer sees the next revision in the same workspace, rather than switching between a model and a separate conversation.

The useful bit is that neither input has to do the whole job. You don't have to describe a face that's already selected, or use speech to identify a small feature the pointer can reach. The build keeps direct selection, a visible model, and revision history, then adds conversation for the changes between revisions. It's a voice layer inside CAD, not a voice-only replacement for it.

How AI CAD uses Vocal Bridge

This maps to Vocal Bridge's Voice for your agent surface. Vocal Bridge's Client Actions let the CAD workspace send selection state to the voice agent and receive requests that trigger an update in the editor.

A React integration for this pattern can send the active selection and handle the edit request:

import { useEffect } from "react";
import { useAgentActions } from "@vocalbridgeai/react";

export function CadVoiceBridge({ selectedSurfaceId, applyPatch }) {
  const { onAction, sendAction } = useAgentActions();

  useEffect(() => {
    return onAction("apply_cad_patch", ({ patch }) => applyPatch(patch));
  }, [applyPatch, onAction]);

  useEffect(() => {
    if (selectedSurfaceId) {
      void sendAction("surface_selected", { surfaceId: selectedSurfaceId });
    }
  }, [selectedSurfaceId, sendAction]);

  return null;
}

Illustrative pattern, not the demo's source code. Use inside a VocalBridgeProvider; action names and CAD payloads depend on your app.

The CAD app remains responsible for geometry, dimensions, selection, and revisions. Vocal Bridge handles the real-time voice-to-voice conversation, live transcript, and Client Actions connection. That leaves the builder working on CAD behavior, rather than implementing a voice transport alongside it.

What's hard about voice here

The hard part isn't capturing a sentence. It's grounding that sentence in the geometry on screen, showing when a revision is actually ready, and keeping every measurement easy to verify.

• "Right here" is ambiguous without visual context. A spoken request can name the change without identifying the surface. "Put a hole here" only works when the agent and CAD app agree on what's selected, even as the designer moves between surfaces. The CAD app resolves the selection to valid geometry; Vocal Bridge Client Actions pass that context between the interface and agent.
• A fast acknowledgment isn't the same as a finished edit. In the demo, the agent replies before the model changes. The interface needs clear pending and completed states so the designer knows when the new geometry is ready to inspect. The CAD app owns revision status; Vocal Bridge's real-time voice channel carries the spoken acknowledgment.
• A spoken number needs visible proof. A six-millimeter hole or another five millimeters of height has to remain checkable after the edit. The CAD app validates the dimensions; Vocal Bridge's live transcript keeps the request beside the model so the designer can compare what they said with what changed.