GitHub - kstonekuan/splatter-mcp-app: MCP app for generating and viewing 3D models using 3D Gaussian Splatting

MCP app for generating and viewing 3D models using 3D Gaussian Splatting

Built with mcp-use and ml-sharp.

🏆 Finalist at the YC MCP Apps Hackathon by Manufact.

Introduction

This is great for teams to move from source media to interactive 3D quickly

• Advertising
• Media and film
• Creative design
• Architecture
• Real estate

Features

• Upload .ply and view instantly in an interactive widget.
• Upload an image and use 3DGS via Modal (async job flow).

ChatGPT image generation to 3DGS example:

Tools Exposed

• open-ply-upload: opens upload widget (recommended entrypoint).
• view-ply-splat: render an existing .ply.
• generate-splat-from-image: direct image-to-splat tool path.

Local Development

pnpm install
pnpm run dev

For ChatGPT connector testing:

Environment Variables

• SHARP_MODAL_ENDPOINT_URL (required): deployed Modal image-to-splat endpoint.
• SHARP_MODAL_TIMEOUT_MS (optional, default 300000).
• SHARP_MODAL_TIMEOUT_SECONDS (optional fallback).
• MCP_URL (recommended for tunnel/deployed absolute URLs).

Modal Backend

Deploy from services/sharp-inference:

cd services/sharp-inference
uv run modal token new
uv run modal deploy modal_app.py::modal_app

Quality Checks

pnpm check
cd services/sharp-inference
uv run ruff check --fix
uv run ruff format
uv run ty check