A developer going by anteloc has released LDraw Nova, agent tooling that turns a model idea into a buildable LEGO CAD file — by writing Python that generates the geometry rather than guessing part positions directly.
2026-10-03 · by Moe Ameen
A developer who goes by anteloc has open-sourced LDraw Nova, agent tooling that generates buildable LEGO models in the LDraw format. It surfaced on GitHub and Hacker News in early October 2026 as a first public release (tagged v0.6.0), and it takes an unusual approach to a hard problem: getting a language model to design a real, physical, buildable thing. You give an AI agent a model idea, guide it, and get back an LDraw source file — the plain-text "assembly language" that LEGO CAD tools like LDView and LeoCAD read — along with a 3D viewer, a VR view for Meta Quest 3, a Blender-editable glTF/GLB export, rendered images, and the full chat history of how the agent reasoned its way there.
The core insight is in how it avoids a known LLM weakness. Models are poor at the geometric math needed to place bricks correctly in 3D space, so LDraw Nova doesn't ask the model to place parts directly. Instead the agent reads the project's documentation, writes a detailed plan of the model's geometry as JSON, generates Python scripts that compute and emit the LDraw placement instructions, runs them, then renders, inspects, and adjusts in a loop until the build holds together. It is the same trick that makes coding agents useful — have the model write code that produces the answer rather than produce the answer itself.
Under the hood the tooling was built with Claude Opus 5.5 and GPT-6 Astra, and it uses a semantic search layer called Jev (TypeSafe's Jev System One model) via jev-rerank to find the right parts and reference models from LDraw's large parts library, falling back to full-text search when Jev isn't configured. It ships as a Dockerized web app: you clone the ldraw-nova and ldraw-nova-docker repositories at matching tags, build the image, and run it locally, reaching the UI over HTTPS on port 8443 (required for VR) or HTTP on 8765. The code is licensed AGPL-3.0.
anteloc is candid about the limits. There is no physics modeling, Technic mechanisms are a weak spot, and the agent still flips parts or gets signs wrong fairly often; large, correct models reliably come only from high-end models, and the whole process is currently slow and expensive to run. This is a research-grade first release, not a one-click toy — but it is a genuinely novel demonstration of an agent designing something you could actually build out of bricks.
This is a launch you can ride the same day you read about it. The novelty does the hook work; your job is to turn one build session into a week of posts, and that is where [Kompozy](/) comes in. Run LDraw Nova on a model idea, screen-record the agent planning, generating, and self-correcting in the 3D viewer, and grab a few clean renders and the final turntable. Then hand that raw footage and the render stills to Kompozy and let it fan the moment across platforms: [Clipped Shorts](/glossary/clipped-short) cut the most satisfying "it fixed itself" beats into vertical highlights, [Persona Shorts](/glossary/persona-shorts) or a HeyGen avatar narrate what happened over the renders, brand-exact [Carousel Posts](/glossary/hyperframes) walk through plan → Python → bricks, and Quote Graphics pull the sharpest line from the agent's reasoning log.
Because this is a timely subject, speed matters more than polish, and Kompozy's edge here is that one capture becomes a full drop without you editing each cut by hand. A [Persona Brief](/glossary/persona-brief) keeps your voice consistent across all of it, the transcript spins into a blog explainer and an email newsletter for your list, and [Autopilot](/glossary/autopilot) schedules and publishes the batch across the eight social platforms plus blog and email through a per-post review pass — so you are early on the trend on every channel at once instead of posting the same clip to one app and calling it a day.
LDraw Nova is open-source agent tooling, released by a developer called anteloc, that generates buildable LEGO models in the LDraw format. You give an AI agent a model idea and it produces LDraw source code (a .ldr/.mpd CAD file), plus a 3D viewer, a VR view, a Blender-editable glTF/GLB export, rendered images, and the agent's reasoning log.
Rather than asking the model to place bricks directly (which LLMs are bad at, geometrically), the agent writes a JSON plan of the model's geometry, generates Python scripts that compute and emit the LDraw placement instructions, runs them, then renders, inspects, and adjusts in a loop until the build is correct. It was built with Claude Opus 5.5 and GPT-6 Astra and uses a semantic search layer called Jev to find the right parts.
The code is open-source under AGPL-3.0 and runs as a self-hosted Docker web app, but you supply the models that power the agent, which the author notes is currently slow and expensive. The honest limits per anteloc: no physics modeling, weak Technic support, frequent part-placement flips, and reliable large builds only with high-end models. It is a research-grade v0.6.0 first release.
The build session itself is the content. Screen-record the agent planning, generating, and self-correcting in the 3D viewer and grab the final renders, then use a tool like Kompozy to turn that single capture into clipped shorts, an avatar-narrated video, a plan-to-bricks carousel, a blog explainer, and an email newsletter — scheduled across the eight social platforms plus blog and email.