Describe a Geometry Dash level in plain words and watch it get built for you. EditorAI talks to the AI of your choice, then places the result in the editor on its own preview layer you can accept, edit, or deny — your level, your call, always.
Consider supporting the project by starring it on GitHub.
- Download the latest release (
.geodefile) - In GD, open the Geode page → press the manual install button → pick the file
- Restart the game
EditorAI is distributed here rather than on the Geode index, whose guidelines don't allow AI-placed objects.
Press E anywhere in the game (desktop), or tap the floating AI bubble (mobile). Everything lives in that panel: starting generations, watching them run, chatting with the AI, and every setting.
Open the panel → Settings → choose a provider. Two are free with zero accounts:
- Platinum (free, community-run — servers by VLT GG) — set provider to
ollama, enable Use Platinum. Your prompts run on volunteer machines, so keep personal info out of them. Want to donate compute? See the Platinum repository. - Ollama (free, fully local — recommended with 6+ GB VRAM) — install Ollama, pull a model, set provider to
ollama, pick the model. Nothing leaves your machine.
For hosted providers, pick one and either press Sign in with browser (OpenRouter, HuggingFace — no key-copying needed) or paste an API key:
| Provider | Get a key |
|---|---|
| OpenRouter | sign-in button, or openrouter.ai/keys |
| HuggingFace | sign-in button, or huggingface.co/settings/tokens |
| Claude | console.anthropic.com |
| OpenAI | platform.openai.com/api-keys |
| Mistral | console.mistral.ai |
| DeepSeek | platform.deepseek.com |
| Groq | console.groq.com/keys (free tier, fastest inference) |
| Gemini | aistudio.google.com (tight rate limits) |
LM Studio, llama.cpp, and any OpenAI-compatible endpoint work too — point the URL at your server.
- Panel → Generate
- Pick a target: the current editor, a brand-new level, or any of your created levels
- Describe what you want, set difficulty/style/length (or type your own), press Generate
- Watch it think and work live in Sessions — you can close the panel, or even the editor, while it runs
- With live block placement (default on) blocks appear in the editor while the AI is still building; the result lands on its own editor layer (the rest of the level fades back): Accept, Edit, or Deny
- Keep talking to it in Sessions to refine, plan, or just ask questions
- Rate generations when asked — your ratings teach the AI your taste over time
Every option in the panel has a hover tooltip explaining what it does.
While it works, the AI can analyze your level's song (BPM + waveform) to sync gameplay to the music, keep scratch memory between generations, and set itself a goal with a task checklist that it grinds through until done — watch the checklist tick off in the status line and session log.
Join the HuggingFace org to publish models trained on EditorAI datasets.
Found a bug or have an idea? Open an issue.