ldraw-nova workflow
ldraw-nova gives an AI agent tools to build a LEGO model in the LDraw format from a prompt, with real parts and collision checks
- 544 ★ on the repository
- Last commit
- AGPL-3.0
Extended access
Why this level
- Runs Docker containers and generates and executes Python scripts written by the agent
- The web app requires no login and is reachable from other devices on the network by the computer's IP
Install
- Works in
Manual install
git clone --branch v0.6.0 https://github.com/anteloc/ldraw-nova.gitClone ldraw-nova-docker at the same tag beside it, then run docker compose build in its folder.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Runs Docker containers and generates and executes Python scripts written by the agent
- The web app requires no login and is reachable from other devices on the network by the computer's IP
This is third-party code. Review the repository files before installing.
What it does
ldraw-nova is a set of tools, examples and instructions an agent uses to design models from real LEGO parts in the LDraw format. The agent takes a prompt, reads instructions.md and related documents, plans the model with submodels, and then loops: renders images, inspects them and adjusts position and look. It almost never places parts directly. It builds a plan.json plan, writes a generate.py script from it, and the script produces an LDraw model.mpd file. The author's hypothesis is that agents write Python that produces geometry better than they do the rotation and position math themselves. The tools help find suitable parts, example models and submodels, detect collisions and gaps when placing parts, and render headlessly. Part search uses the jev-rerank semantic search with a TYPESAFE_API_KEY; without a key the agent falls back to full-text search and models may come out worse. The output is LDraw source, a 3D viewer, a player, VR for Meta Quest 3, images, a .glb file for editing in Blender and the chat history with the agent's reasoning. It runs as a web app in Docker built from two repositories.
Who it is for. LEGO fans and agent experimenters who want editable digital models from a description, and researchers of generated physical constructions.
Good fit when
- You need a LEGO model in LDraw from a description, with part collision checks
- The model has to open in Blender: the project outputs a .glb file with metadata in custom properties
- You want to see how the agent reasoned while building: chat history and thinking process are saved
- You study how agents handle geometry through Python generators
Not a fit when
- You want cheap generation: only expensive top-tier models can currently build large, correct models
- You need minifigs, animals or good spaceships: the author lists these families as weak spots
- The network is not trusted: the web app has no login and must not be exposed on untrusted networks
Example request
Build a tower crane from real LEGO parts, check collisions and give me a file for BlenderLimitations
- A first release. Building the Docker image takes long and needs about 5 GB of disk, with Git and Docker and the two repositories ldraw-nova and ldraw-nova-docker at the same tag.
- Without a TYPESAFE_API_KEY part search falls back to full-text and results are worse.
- VR on Meta Quest 3 has model handling and performance issues. Generation is slow.
- The web app has no login, so run it only on trusted networks.
How to disable. Stop the app with docker compose down in the ldraw-nova-docker folder. The Docker image and volumes are removed with Docker commands, and the cloned repository folders are deleted by hand. The TYPESAFE_API_KEY is stored in the web app's settings.
FAQ
What is LDraw and why does an agent need it?
LDraw is a text format for describing models made of LEGO parts, an assembly language for 3D CAD. It is simple, low-level and executable, so an agent can produce it with code and verify the result with renders.
How do I run ldraw-nova?
Clone ldraw-nova and ldraw-nova-docker at the same tag, run docker compose build and docker compose up -d in the ldraw-nova-docker folder. The app opens at https://localhost:8443 (needed for VR) or http://localhost:8765.
Do I need a key for part search?
Better search needs a TYPESAFE_API_KEY for the Jev System One model, entered in the web app's settings. Without a key the agent uses full-text search and models may come out worse.
What does the agent output?
LDraw source, a 3D viewer and player, a VR mode for Meta Quest 3, images, a .glb file for Blender editing with metadata in custom properties, plus the chat history and the agent's reasoning.
Why does the agent build a plan and a script instead of placing parts?
The author observed that agents write Python that produces geometry better than they compute part rotations and positions themselves. So the chain goes from a plan.json plan to a generate.py script and a model.mpd file.
Discussion
No questions or reviews yet.
Related
Workflows3D generation MCP server: from a description the agent gets a GLB model with named parts, joints and a Blender script
Plugin for Claude Code and Codex: the agent builds exact parts in Blender 5.2 via bpy, exports glTF and STL and audits meshes.
A local MCP server and CLI: the agent writes JavaScript for a 3D model, checks it on renders and exports a GLB with source
The scan-to-threejs skill refines a 3D scene in a loop, checking renders against room photos from matching camera angles.
The official Lovart skill: an agent draws images, makes video, music and 3D models and keeps projects with chat history
Gives text-only models vision: images go to GLM and Agnes, OCR to Baidu and Windows, PDFs to MinerU, the answer comes back as JSON
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| ldraw-nova (this card) | no | 544 | ||
| build123d-mcp | no | not required | 151 | parametric CAD code for engineering parts, while this builds models from ready-made LEGO parts in LDraw |
| Blender Agent Skills | no | 1,014 | skills for Blender, while this builds models from the LEGO part catalog with collision checks |
README in short
The README shows the output, explains the Docker install with a note on the part search key and lists open issues. A section on the idea tells the path from three earlier experiments to the current scheme, followed by the agent's steps and a table of guides.