Open-source agent tooling that designs buildable LEGO models in the LDraw format — the agent writes a plan, generates Python that emits the brick placements, then renders, inspects, and fixes in a loop.
Last verified · 2026-10-03 · by Moe Ameen
LDraw Nova is open-source agent tooling, released by a developer who goes by anteloc, that generates buildable LEGO models in LDraw — the plain-text "assembly language" that LEGO CAD programs like LDView and LeoCAD read, where each line places a single part. You give an AI agent a model idea, guide it, and get back a real CAD file you could open, order parts for, and physically build. It surfaced on GitHub and Hacker News in early October 2026 as a v0.6.0 first release, licensed AGPL-3.0.
Its defining idea is how it sidesteps a known LLM weakness. Language models are poor at the geometric math required to place bricks correctly in 3D, so LDraw Nova never asks 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 — the same "have the model write code that produces the answer" pattern that makes coding agents work.
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) via jev-rerank to find suitable parts and reference models in LDraw's large library, falling back to full-text search when Jev is not configured. It runs as a self-hosted Docker web app — you clone the ldraw-nova and ldraw-nova-docker repositories at matching tags, build a roughly 5GB image, and open the UI over HTTPS on port 8443 (required for VR) or HTTP on 8765.
Be clear-eyed about the stage: the author is candid that there is no physics modeling, that Technic mechanisms are a weak spot, that the agent still flips parts and gets signs wrong fairly often, and that the process is currently slow and expensive, with reliable large builds coming only from high-end models. It is a research-grade tool that rewards tinkering, not a one-click consumer app.
LDraw Nova is unusually generous with artifacts — every session leaves you with LDraw source, still renders, a 3D turntable, a VR view, a Blender glTF/GLB, and a complete reasoning log. That is a content kit hiding in a project folder, and [Kompozy](/) is the engine that turns it into a repeatable publishing system rather than a one-off post. The trick is to treat each build as an episode in a series: one LDraw Nova run becomes a structured content drop, and the next run slots into the same template.
Map the assets to formats deliberately. The turntable render and screen-recorded build loop feed [Clipped Shorts](/glossary/clipped-short) and [Persona Shorts](/glossary/persona-shorts) or a HeyGen avatar that narrates how the agent designed it; the still renders and the parts list become brand-exact [Carousel Posts](/glossary/hyperframes) and Quote Graphics via HyperFrames; and the chat log — the plan, the Python, the fixes — is a ready-made outline for a Blog Article tutorial and an Email Newsletter. A [Persona Brief](/glossary/persona-brief) keeps one voice and your banned words across every asset so a whole build series reads as one brand, and [Autopilot](/glossary/autopilot) schedules and publishes the batch across the eight social platforms plus blog and email through a per-post review pass. LDraw Nova designs the bricks; Kompozy turns every build into finished, on-brand content on every channel.
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 a real CAD file (.ldr/.mpd), plus a 3D viewer, a VR view, a Blender-editable glTF/GLB, rendered images, and the agent's reasoning log. It is licensed AGPL-3.0 and self-hosted via Docker.
Rather than placing 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 placements, runs them, then renders, inspects, and adjusts in a loop. 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.
Buildable LEGO models as LDraw source you can open in LeoCAD or BrickLink Studio, plus rendered images, a 3D turntable, a VR walkaround for Quest 3, and a Blender-editable glTF/GLB export. The author is honest that it is a v0.6.0 first release — strong on vehicles and modular builds, weak on Technic, with no physics and frequent part-placement fixes needed.
Screen-record the agent designing the model and export the renders, then bring them into Kompozy. It turns one build session into clipped shorts, an avatar-narrated explainer, a plan-to-bricks carousel, a blog tutorial from the chat log, and an email newsletter — all in one brand voice and scheduled across eight social platforms plus blog and email.