The project ldraw-nova is an open‑source web application that enables large language model agents to design physical LEGO constructions and export them as LDraw source files. Agents receive a natural‑language prompt, consult provided documentation (instructions.md, visual design guides, workflow manuals), and formulate a build plan that lists required parts, submodels, and aesthetic considerations. Rather than placing bricks directly, the agent creates a JSON plan (e.g., atlas‑crane.plan.json) and a Python generator script (e.g., generate.py); executing the script yields one or more LDraw files (e.g., atlas‑crane.mpd) that describe the model in the low‑level LDraw assembly language. Supporting tooling includes jev‑rerank, a semantic search with re‑ranking powered by TypeSafe’s Jev System One model (requiring a TYPESAFE_API_KEY), which falls back to full‑text search when the key is absent. The agent also uses collision and gap detection, headless rendering, and example models to iteratively refine the design via rendered images. The application is deployed via Docker: cloning the ldraw‑nova and ldraw‑nova‑docker repositories at tag v0.6.0, building the image (≈5 GB disk space, noticeable first‑build time), and starting services with docker compose up -d. The service is reachable at https://localhost:8443 for VR on Meta Quest 3 (self‑signed certificate) or http://localhost:8765 for plain HTTP; no authentication is implemented, so it should run only on trusted networks. Current limitations include VR performance issues, high computational cost restricting successful generations to advanced agents, slow generation speed, and limited support for certain model classes such as minifigs, detailed Technic mechanisms, and high‑quality spaceships, with planned improvements for low‑end agents, efficiency, expanded model families, and finer‑grained inspection of submodel assembly steps.