Early development

CAD in a language
AI actually understands

Parametric CAD breaks when LLMs touch it. TokenField doesn't.

TokenField is a geometry platform built for AI-generated CAD. Instead of fragile topology and constraint systems, it provides a functional representation that LLMs can reliably generate—and engineers can trust.

Get early updatesNo spam. Only real progress.

Why current CAD formats fail with AI

Today's AI-CAD pipelines look like this: prompt → LLM → legacy CAD code or STEP files.

The problem isn't the model. It's the target format:

  • B-Rep is brittle. Explicit edges and faces create a combinatorial explosion of ways to produce invalid geometry. LLMs hit these constantly.
  • Constraints break silently. Parametric systems rely on implicit relationships that collapse when edited programmatically.
  • There's no separation of concerns. "What the design should be" is tangled with "how it's represented internally."

TokenField takes a different approach: geometry as composable functions, with constraints and metadata in a separate layer. The result isa format that's native to how LLMs reason—and far harder to break.

What TokenField is

Directional. Interfaces and APIs are still evolving.

ƒ

TokenField Language

A functional geometry IR designed for LLM generation:

  • F-Rep/SDF foundation
  • Constrained output space
  • Matches LLM reasoning

Builder

The interpreter and compiler for the TokenField language:

  • Parses and validates code
  • Compiles to renderable forms
  • Clear error reporting

Studio

A native desktop application for hands-on design work:

  • Real-time GPU rendering
  • Visual code editing
  • Local file workflows

Platform

A web-based environment for TokenField designs:

  • WebGPU ray-marching
  • AI pipeline integration
  • Export to STL, STEP

Who we're building for

Primary

AI tool developers

Developers building AI-powered design tools, CAD copilots, or text-to-3D systems. TokenField provides the geometry backend that eliminates topology errors.

"Stop fighting topology errors. Generate geometry LLMs can't break."

Secondary

3D printing and makers

Engineers and hobbyists who want to describe parts in natural language and get printable, inspectable geometry.

"Describe it. Generate it. Print it."

Tertiary

Programmatic CAD developers

OpenSCAD, CadQuery, and code-first CAD users who appreciate the F-Rep approach and want AI-native generation.

"Code-first geometry that composes like functions, not topologies."

Why F-Rep

Boundary representation (B-Rep) stores geometry as explicit edges, faces, and vertices. This works well for manual modeling but creates problems for AI generation:

AspectB-RepF-Rep (TokenField)
Invalid statesMany (non-manifold, self-intersecting, etc.)Few (functions compose cleanly)
Edit stabilityFragile (topology changes cascade)Robust (parameter changes, not topology)
LLM generationError-proneReliable
RenderingMesh conversion requiredDirect GPU evaluation

F-Rep isn't new—libfive, nTopology, and others use it. What's new is designing a language and IR specifically for LLM generation, not human authoring.

Follow the project

We'll send updates when there's something real: demos, benchmarks, or concrete results.