nendo is a browser app for building and refining 3D models with written instructions alone. Type “make a dog”, then “make the ears droop more”, then “longer tail, thicker at the tip”, and the 3D view updates each time. There are no numeric fields, no gizmos, no parameter sliders.
Shapes are defined as signed distance fields and turned into meshes by isosurface extraction. The language model outputs only a skeleton — joint positions and thicknesses — plus editing intent; it never touches vertices. The fleshy continuity where a leg meets the torso comes from the smooth minimum of two distance functions, a shape a boolean union can never produce.
Distance-field evaluation and isosurface extraction run inside a Web Worker. Evaluating every voxel would mean 16 million samples at 256³, so two coarse passes narrow things down to the band the surface can pass through, and only that band is refined — about a third of what a brute-force sweep of the same grid would cost.
There is a second way to build, Blocks mode, which combines primitives — boxes, spheres, cylinders, cones and more — into things like houses, cars and furniture. It skips the distance field and draws three.js's standard geometry directly.
Most of the cost of learning 3D modeling sits in translating a shape you already have in mind into tool operations. This app tests the hypothesis that if language alone can reach an organic shape, that translation step disappears. The key constraint is never letting the language model compute vertices or coordinates: it states intent, and the runtime resolves the numbers and builds the geometry.
glTF (.glb), OBJ, STL, scene JSON, PNG and SVG. All of it happens in the browser; nothing passes through the server. The SVG is a vector drawing of the model from the current view, made of filled polygons. Only the scene JSON can be loaded back for further editing.