A local web app — and the start of a fuller AI Game Studio — for generating game assets from text prompts. Today: 2D reference sprites and animation frames composed into a 1×N spritesheet with a looping animated preview. Backgrounds are chroma-keyed to transparency automatically, so frames drop straight into a game engine. Projects can be saved and loaded by name.

The app talks to OpenRouter as the single boundary to the model providers. One key gives access to 300+ image / video / audio / text models, which is the runway for everything on the TO-DO list (backgrounds, tilemaps, SFX, music, voice, …).

Pick the video model at generation time — currently Grok Imagine Video (xAI) or Seedance 2.0 (ByteDance), both routed through OpenRouter.

Full Demo: https://www.youtube.com/watch?v=MijheSPXnDo

• Node 20+
• ffmpeg on PATH
• An OpenRouter API key

npm install

cp .env.example .env
# then open .env and paste your key:
# OPENROUTER_API_KEY=sk-or-v1-...

npm run dev

Open http://localhost:5173.

This starts Vite (frontend, :5173) and an Express server (backend, :8787) together. Stop with Ctrl+C.

1. Type a sprite prompt in column 1 → Generate Reference Sprite.
2. Pick a video model and type a motion prompt in column 2 → Generate Frames (calls image-to-video via OpenRouter, polls until done, extracts transparent PNGs).
3. Click frame tiles to toggle which ones to include.
4. Generate Spritesheet → composes a 1×N PNG client-side, builds a looping GIF preview server-side.
5. Export PNG to download the spritesheet.
6. Header: New to start fresh, Save to name and snapshot the current project, Load to switch to a saved one.

Generated artifacts live under projects/ (gitignored). The current working state is always in projects/latest/.

• A pixel-art knight in silver armor with a longsword, side-view, full body, simple flat colors, standing pose
• Female ninja with red scarf, dynamic side-view, 2D sprite, anime style
• Cute green slime monster, side-view, big eyes, soft shading
• Cyberpunk hacker in a hoodie, glowing visor, side-view full body, gritty style

• Smooth walk cycle, side-view, no head tilting, no camera movement
• Sword slash attack, side-view, fast, no shadows
• Idle breathing animation, subtle, looping
• Jump arc — crouch, leap, mid-air, land

Tips:

• Keep motion prompts focused on the action. Phrases like "no camera movement", "side-view", and "no head tilting" help keep frames game-ready.
• Per-model default durations: Grok Imagine Video = 2 s, Seedance 2.0 = 4 s. ~24–30 fps on the source clip, so trim with the frame selector before composing.
• Switching the model is one entry in server/video.ts — see VIDEO_MODELS.
• Recommend sticking to Grok Imagine Video since it's much cheaper than Seedance 2

• Background generation
• Tilemap generation
• Aseprite format export
• Tiled format export
• SFX generation
• Music generation
• Voice generation
• Full asset scaffolding export

See AGENTS.md for the full spec, architecture, endpoint list, model-registry pattern, and chroma-key tuning notes.