Skip to content
- 0 1
.LLM3D
© 2026 llm3d

= 01234

(The 3D engine for LLMs.)

llm3d is a modeling engine built for language models. Claude writes objects as code: named parts, booleans, fillets, on a Rust core where every solid is watertight. Show it a photo, describe a part, or just say what's wrong.

Get early access How it works
Photo of a red hair dryer with its cord and plug, lying on a table
The llm3d model of the same hair dryer, rendered from above→ Model
  1. barrel
  2. handle
  3. cord
  4. grille
  5. plug
Fig. 1 — 1 photo → 11 parts
.01 > Core
Rust SDF
.02 > Interface
MCP + Python
.03 > Booleans
Never fail
.04 > Mean error
1.34%
.02 > How it works

One loop, about a minute a round.

  1. .01 >

    Read the photo

    A vision model describes every part. SAM 3 outlines them, MoGe-2 measures depth.

  2. .02 >

    Model

    Claude writes the object as code: named boxes, revolves, lofts, booleans, fillets.

  3. .03 >

    Compare and fit

    The model is rendered from the photo's camera, checked against it, and its numbers fitted.

  4. .04 >

    Refine by talking

    Say what's wrong: “the handle is missing”. Each round takes about a minute.

// Export
  • STEP
  • GLB
  • OBJ
  • STL
  • USD
  • 3MF
  • .blend
  • .py
.03 > Accuracy

Measured against real 3D scans.

Mean error: 1.34%

= Mean error, 5 scanned objects

We modeled five real objects from one photo each and measured the models against true 3D scans. The average error is 1.34% of the object's size, closer than a dedicated photo-to-3D model.

llm3d, full run1.34%
SAM 3D (photo-to-3D model)1.44%
llm3d, quick run (40 tool calls)2.0%
Fig. 2 — mean error over 5 scanned objects. Lower is better.
.04 > Early access

Coming soon. Plugs into Claude Code.

A hosted service. Credits pay for the AI steps: image generation, vision reading, depth. The modeling itself is free.

  1. .01 > Buy a credit pack. No account.
  2. .02 > Get a key with the pack.
  3. .03 > Run one command.

One email when early access opens. No spam.

.03 > The command, once you have a key

claude mcp add --transport http llm3d https://mcp.llm3d.xyz/mcp --header "Authorization: Bearer <your key>"