part-builder

Describe a Roblox model in words and get it back built from native Roblox parts (blocks, wedges, corner wedges), rigged and ready for Studio.

How it works

1. AI draws your idea, or you bring your own picture. You pick the one you like.

2. It makes front, side and back views of it (or you upload them). You approve them.

3. Claude builds the model from parts, compares it with the views and fixes it until it matches. You watch it happen in the editor.

4. You get a rigged .rbxmx file, or the model is pushed straight into Roblox Studio.

The editor: the model, its parts, the reference pictures and buttons to upload your own.

The review screen: reference, model, overlay and line map for each view, with the score.

Related MCP server: Roblox Bridge MCP

You need

• Claude Code

• Node.js (24 tested)

• An image source: ComfyUI on your own GPU (free), or an API key from Gemini, OpenAI, fal.ai or Replicate (paid, a few cents per image). Or none: use your own pictures.

• Roblox Studio (optional)

Setup

npm install
npm run build:web
cp .env.example .env

1. Set your image source in .env (the options are listed there). For ComfyUI, see docs/local-gpu.md.

2. Start Claude Code in this folder and allow the part-builder MCP server.

3. Optional: in Roblox Studio, open Assistant Settings > MCP Servers > Quick connect and turn on Claude Code (Roblox docs).

Use

Ask Claude Code something like "make a blocky fox, about 3 studs tall". Open the editor link it gives you to pick the picture, approve the views and watch the build.

• Your own pictures: give Claude the file path, or use the Upload buttons in the editor's References panel. A transparent PNG, or a plain background in a color the model doesn't use, works best.

• Results: projects/<name>/. The Roblox file is exports/<name>.rbxmx (Studio: right-click Workspace > Insert from File).

• Look: presets in styles/. Drop up to 6 example screenshots into a preset's folder to steer the pictures.

• No image source yet? Set PART_BUILDER_FAKE_IMAGES=1 to try it with a demo knight.

My setup

• A Windows laptop running Claude Code and Roblox Studio.

• A second PC with an RTX 3090 running Comfy Desktop, reached over Tailscale. The laptop starts ComfyUI on it over SSH when it needs pictures (COMFYUI_SSH in .env), and it shuts down when idle.

• FLUX.2 klein 4B for the first pictures, Qwen-Image-2.1 with the Multiple-Angles LoRA for the views.

• Studio connected with Quick connect, so finished models go straight in.

License

MIT. The AI models have their own licenses: Qwen-Image, used by the default ComfyUI setup, is non-commercial (details). Stud textures: THIRD_PARTY_NOTICES.md.

Not affiliated with Roblox.

Development

npm test and npm run typecheck. Code layout and conventions: CLAUDE.md.