Analysis October 3, 2026 workflow 7 min read

ldraw-nova shows that an AI agent which writes an editable parts file, rather than generating a picture of bricks, gives designers something they can revise, render and price, but its own instructions admit a pretty render does not prove a model can be built, so you check the build yourself.

ldraw-nova LDraw BrickLink Studio Blender creative-workflows open-weights local-creative-ai prompt-craft

ldraw-nova: Design a LEGO Model by Describing It to an AI Agent, Then Open It in Blender

A new open-source tool has an AI agent place real bricks one by one and hands back a file you can edit, render in Blender or import into BrickLink Studio. Here is the workflow, what it costs, and the one warning built into its own rulebook.

Ask any image generator for "a red LEGO crane on a construction site" and you will get a convincing picture of one. Look closer and the bricks are mush. Studs melt into each other, a beam passes through a wall, a part exists that LEGO never made. It is a picture of the idea of LEGO.

A project posted to Hacker News on October 2, where its Show HN drew 113 points and 44 comments, goes about it the other way. ldraw-nova never draws a single pixel of the model. It has an AI agent write a script that places real, catalogued parts at exact coordinates, render the result, look at it, and fix what is wrong. What you get back is not an image. It is a parts file.

That difference is the whole story, and it reaches well past toys. When an AI hands you a file made of real components instead of a flat picture, you can change it. You can swap one piece, recolor a wall, move a door. A picture only lets you reroll.

Why a parts file beats a picture

LDraw, in the words of its own site, is "an open standard for LEGO CAD programs that allow the user to create virtual LEGO models and scenes." Its parts library holds 17,164 unique shapes and printed parts as of September 2026. Each line of an LDraw file says, in effect, "this part, this color, at this position, at this angle." It is closer to a recipe than a drawing.

That is what makes ldraw-nova useful to a designer. Because the output is a recipe of real parts, it flows into tools people already use:

Blender. ldraw-nova exports a .glb file that Blender opens directly, so a 3D artist can light, animate and render the model like any other asset.
BrickLink Studio. The free LEGO design app imports LDraw files in the .ldr and .mpd formats. ldraw-nova's agent is told to deliver an editable .mpd file, so the model can land in the program many builders already use to refine designs.
A 3D viewer and VR mode built into the app, including a Meta Quest 3 mode the author openly calls rough.

You also get the full chat history and the agent's reasoning, which turns out to be the most educational part.

How the agent actually builds

The README lays out the loop. The agent takes your prompt, reads a long instruction file about the LDraw format and building principles, plans the model, then works in cycles: render, inspect, adjust. It writes a plan describing the full geometry and a Python script that produces the final LDraw file.

The instruction file reads like a senior builder briefing a junior one. It fixes the measurements (a stud is 20 units apart, a brick is 24 units tall, a plate is 8). It tells the agent to bridge seams with real connecting parts, to reserve space for wheels before covering a vehicle's body, and to check each joint. It also tells the agent to build small and test often: one detail, then a module, then the whole scene.

The best line in the whole project is a warning the author wrote to the AI:

"Correct syntax, a plausible picture and a connected-component result are separate kinds of evidence; none alone proves physical buildability."

In plain terms: a file that loads, a render that looks right and a model where every part touches something are three different things, and none of them means the thing would survive being built with real bricks. The agent is told not to claim it reviewed images it never opened, and not to call a model buildable just because it rendered.

I trust a tool more when it tells its own AI to stop overclaiming. You should borrow that rule for yourself.

What it costs

The software is free and open source under the AGPL-3.0 license. That license matters only if you run a modified version as a service for other people; for personal and studio use it is not a concern.

The real cost is the AI. The README is blunt: "only expensive, high-end models, are currently capable of generating large-sized and correct models." It names Claude Opus 5.5 and GPT-6 Astra as the models that made the project work. A small model might give you a passable car. A crane with working geometry needs a top-tier model, many render-and-fix cycles, and the API bill that comes with them. The README does not publish a cost per build, so watch your usage on the first few runs before you queue anything ambitious.

Hardware is modest by comparison. Everything runs in Docker on your own machine and needs about 5 GB of disk to build.

LEGO is a trademark of the LEGO Group, and this project is not affiliated with it. Selling designs is a separate question you should research before you try.

Put this into practice

Here is a first session, start to finish.

1. Install it. You need Git and Docker. Clone both repositories at the same version:

git clone --branch v0.6.0 https://github.com/anteloc/ldraw-nova.git
git clone --branch v0.6.0 https://github.com/anteloc/ldraw-nova-docker.git
cd ldraw-nova-docker
docker compose build
docker compose up -d

Open http://localhost:8765 in your browser. Use https://localhost:8443 only if you want the VR mode; it uses a self-signed certificate, so your browser will warn you.

2. Look at the examples before you spend anything. The repo ships finished builds, including examples/atlas-crane/, plus vehicles, spaceships, mechanisms and a modular street. Open one, read its chat log, and see how many correction rounds it took. That tells you more about what a build will cost than any guess.

3. Write a builder's brief, not a mood. The agent works from geometry, so give it dimensions and function. Compare these:

Weak: "a cool retro spaceship."
Strong: "a small two-seat spaceship about 16 studs long, swept wings, a cockpit canopy, landing gear that sits flat, mostly light gray with dark red accents."

The second gives the agent sizes, parts of the model to plan, and a test it can check (does it sit flat?).

4. Ask for modules. The instructions already push the agent to test small pieces first. Lean into it. Ask for the cockpit, check the render, then the wings, then the full ship. A mistake caught in a module is cheaper than one found in the finished model.

5. Take it where you need it. For a render or an animation, import the .glb into Blender and treat it like any other model. For a physical build, import the .mpd file into BrickLink Studio, where you can inspect it part by part and fix what the agent got wrong.

6. Check it yourself. In Studio, rotate the model and look at every joint. Floating parts and pieces that only touch at a corner are the classic failures. Remember the agent's own rule: a nice render is not proof.

Where it breaks

Big models need big models. The author's own warning: large, correct builds currently require expensive, high-end AI. If you are hoping for a 2,000-piece castle on a budget model, you will be disappointed.

Buildability is on you. The project's rules admit that none of its automatic checks proves a model can be built. Treat every output as a strong draft that needs a human builder's eye.

Setup is developer-flavored. Docker, two repositories at matching versions, and a self-signed certificate for VR. The Docker repo's own README is marked as being rewritten, with several sections listed as coming soon. If you have never used a terminal, ask a friend for the first install.

VR is rough. The README says model handling on Meta Quest 3 has "many issues" and performance problems.

It is early. Version 0.6.0, 136 GitHub stars at the time of writing, one main author. Pin the version you install and expect things to move.

The useful lesson beyond bricks

The reason to care about ldraw-nova, even if you never build a LEGO model, is what it shows about how AI fits into design work. The image generators of the last few years gave us pictures we could admire and could not edit. This gives a file made of real parts that you can open, question and change.

The same pattern is showing up elsewhere: AI that writes the recipe instead of serving the finished dish. Recipes are what designers actually work with.

So pick one small model you have always wanted to see, write it a proper brief with sizes and a test, and watch the agent build it piece by piece. Then do the part no AI can do for you yet: turn it over in your hands, virtually or for real, and see if it holds.

Medium metadata

Title: ldraw-nova: Design a LEGO Model by Describing It to an AI Agent, Then Open It in Blender

Subtitle: A new open-source tool has an AI agent place real bricks and hands back an editable parts file for Blender or BrickLink Studio. The workflow, the cost, and the warning in its own rulebook.

Tags: LEGO, AI Tools, 3D Design, Blender, Generative AI

Estimated read time: 8 minutes