Geode mod for Geometry Dash 2.2

Type it.
Watch it build.

An AI level designer that lives in your editor. Describe a level, watch it draw onto the grid in blue, then Accept, Edit, or Deny — your level, your call, always.

Install the .geode Free with Platinum
3,986 object types
12 providers   WindowsmacOSAndroidiOS
editor · blueprint previewAI · drafting
Preview
3,986
Object types in library
12
AI providers, 5 free
—
Platinum workers online
∞ free
With local or Platinum AI
§ 01
The loop

Describe it. It drafts in blue.
You decide.

Every generation is a conversation. The AI works on its own editor layer — the rest of your level fades back — so nothing changes until you say so.

01 — DESCRIBE
you › a fast ship sequence
over lava, cyan pulses on
the beat, gets harder
toward the end
Just say what you want
Open the panel with E (or the mobile bubble), start a chat, pick a target — the current editor, a new level, or any of your saved levels — set difficulty, style, and length, then Generate.
02 — DRAFT
ai › planning sections…
ai › placing 1,240 objects
ai › adding color triggers
ai › self-review: pass
Watch it build, live
Objects spawn in batches as a blue blueprint preview. The AI plans, self-critiques, and can even see a snapshot of what it built. Close the panel — or the whole editor — it keeps working in the background.
03 — DECIDE
Accept Edit Deny
Accept turns it gold
Accept commits the blueprint into real objects. Edit keeps talking — small tweaks or full reworks. Deny rolls it all back. Every edit the AI makes is previewed the same way and reversible until you accept.
Conversations survive restarts — pick up exactly where you left off.
§ 02
What it actually does

Not a toy. A designer
that does the work.

Editor AI plans, builds, decorates, critiques itself, and reworks existing levels by the thousand-edit — then hands you the result to approve.

A chat that remembers

One unified panel for everything. Each generation is a multi-turn conversation — keep talking for small edits, big reworks, plans, or questions. Sign in to OpenRouter or HuggingFace right in the browser, no key-copying. Conversations persist across game restarts.

make the drop section twice as long and add a wave portal
› extending drop by 240u
› inserting wave portal @ 90
› 612 objects added ▌
staged as preview · accept to commit
perfect, now make it harder

The blueprint preview

Generated objects appear on their own editor layer, tinted blue, while the rest of the level dims back. You see exactly what changed before it's real. Accept commits it, Edit refines it, Deny rolls it back — fully reversible.

AcceptedPreview

Real gameplay objects

3,986 types
Every block, spike, portal, orb, pad, and decoration in Geometry Dash — the full object library, bundled in and resolved by human-readable name.

Two-pass generation

gameplay → decoration
Builds the playable skeleton first; after you accept it, a second pass layers in decoration, color triggers, and atmosphere — without touching the gameplay you approved.

Vision

it sees what it built
On models that support images (Claude, Gemini, GPT-4o, LLaVA…), a rendered snapshot of the level is attached to review turns — so the AI corrects what it can actually see.

AI tools

web · levels · songs · subagent
When enabled, the AI can search the web, analyze any GD level by ID for style reference, look up Newgrounds songs, and consult a second model you choose — e.g. a big cloud model asking your local Ollama for object ideas, or vice versa — before it composes.

Self-critique & refinement

it checks its own work
After the draft, the AI runs review passes — scoring its own level, patching the weakest spot, and checking for death zones and unreachable gaps before it stages.

Heavy reworks, enforced

target objects · edit budget
Set a minimum object count and an edit-workload target — the AI keeps moving, deleting, adding, and restyling — up to thousands of edits, with a default target around 1,500 — until it hits the bar, instead of nudging a handful of blocks.

Copilot mode

watches as you edit
While you build by hand, Copilot watches for death zones and physics problems and quietly proposes a fix blueprint — you still accept or deny. At most one auto-suggestion every couple of minutes.

Runs in the background

close it and walk away
Generations keep going even if you close the panel — or the whole editor. Come back when it's done and your blueprint is waiting to Accept.

It learns your taste

rate → personalize
Rate generations with a thumbs or 1–10 and your best-rated levels become few-shot examples for next time. Opt in to Auto-Share and every generation (prompt + objects + ratings — never your identity or keys) is sent to the community collector; levels rated above 5 become training data for the free community models.
§ 03
Bring any brain

Twelve providers.
Five are free.

Pick a backend in Settings and paste a key (or sign in). Start with Platinum or local Ollama for a zero-cost setup — select a channel to see exactly how.

Platinum

Free · no account
Editor AI's community compute network. Your requests run on volunteer GPUs — no key, no account, no payment. Prompts run on other people's machines, so keep personal info out of them.
Settings → AI Provider → ollama
Platinum speaks the Ollama protocol.
Enable Use Platinum
Routes you to the community network automatically — no URL to type.
Pick a model
e.g. entity12208/editorai:deepseek — the default Platinum model.
Generate
Press E, describe a level, hit Generate. Works only while volunteers are online — check the live count below.

Local Ollama

Free · private
Runs entirely on your machine — nothing leaves it. A GPU with 6+ GB VRAM is ideal; CPU works but slower.
Install Ollama
From ollama.com.
Pull a model
The fine-tuned EditorAI models are built for this:
ollama pull entity12208/editorai:deepseek
Settings → AI Provider → ollama
Leave Use Platinum off (defaults to http://localhost:11434) and pick your model tag.
Make sure Ollama is running
If it isn't already a background service: ollama serve.

Claude

Highest quality
Anthropic API access needs a funded account, but Claude produces the cleanest, most structured levels of any cloud option.
Modelsclaude-sonnet-4-6 — balanced, default · claude-opus-4-6 — highest quality
Get a key
From console.anthropic.com — add credits, create a key (sk-ant-…).
Settings → AI Provider → claude
Pick a model.
Paste the key
Into Claude API Key. Stored on this device in Geode save data.

DeepSeek

Best value
Excellent quality at low prices. deepseek-chat is great for levels; deepseek-reasoner excels at complex layouts.
Modelsdeepseek-chat · deepseek-reasoner · deepseek-coder
Get a key
From platform.deepseek.com.
Settings → AI Provider → deepseek
Choose a model.
Paste into DeepSeek API Key
Done.

OpenRouter

One key, many models
A single key for many hosted models — Claude, GPT, Llama, Gemini and more. Sign in straight from the browser, no key-copying.
Default model sluggoogle/gemini-2.5-flash — or any slug, e.g. anthropic/claude-sonnet-4
Settings → AI Provider → openrouter
Then either sign in with the browser button, or…
Paste a key
From openrouter.ai/keys into OpenRouter API Key.
Set a model slug
In OpenRouter Model.

OpenAI

Bring a key
API access needs a funded account.
Modelsgpt-4o — default · gpt-4.1-mini — faster, cheaper
Get a key
From platform.openai.com (sk-…).
Settings → AI Provider → openai
Pick a model.
Paste into OpenAI API Key
Done.

Mistral

Bring a key
Ministral models are fast and cost-effective. Needs a funded account at La Plateforme.
Modelsministral-8b-latest — default · ministral-3b-latest · mistral-small / medium / large-latest
Get a key
From console.mistral.ai.
Settings → AI Provider → ministral
Pick a model.
Paste into Mistral API Key
Done.

Hugging Face

Free (limited)
The Inference API has a free tier for many models. Sign in from the browser, or paste a token. Quality varies by model.
Default modelmeta-llama/Llama-3.1-8B-Instruct — any chat model with Inference API support works
Get a token
From huggingface.co/settings/tokens (hf_…).
Settings → AI Provider → huggingface
Set a model id.
Paste into HF Token
Or use the browser sign-in.

Gemini

Not recommended
Gemini's free tier has tight rate limits and tends to produce weaker layouts here. Prefer Platinum, local Ollama, Claude, or DeepSeek. If you still want it:
Modelsgemini-3-flash — default · gemini-3-pro — stronger, hits limits faster
Get a key
From aistudio.google.com (free tier).
Settings → AI Provider → gemini
Pick a model.
Paste into Gemini API Key
Expect to hit quota quickly.

LM Studio

Free · local GUI
A friendly desktop app for local models with a built-in server. No command line required.
Install & load a model
Get LM Studio, download a model (Qwen, DeepSeek, Llama…), and load it.
Start the server
In the Server tab → Start Server (default http://localhost:1234).
Settings → AI Provider → lm-studio
Keep the default URL if you didn't change it. Set the loaded model name.

llama.cpp

Free · lightweight
Fast, minimal C++ inference for GGUF models — great for older hardware.
Get llama.cpp + a GGUF model
From GitHub. Smaller 3B–8B models are fastest.
Start the server
./llama-server -m model.gguf --port 8080
Settings → AI Provider → llama-cpp
Default URL http://localhost:8080 — change the port if needed.

Custom (BYOPAK)

Your endpoint
Bring your own OpenAI-compatible API. Point Editor AI at any endpoint that speaks /chat/completions — a self-hosted gateway, a company proxy, anything.
Settings → AI Provider → custom
Set a display name (cosmetic).
Endpoint URL + model
Base URL or full /chat/completions URL, plus the model name your server expects.
Auth header template
Default Authorization: Bearer ${KEY} — ${KEY} is replaced with your key. Clear it to send no header.

Manual (copy-paste)

Free · any AI
No key, no account, no network. Use any chatbot you can open — free ChatGPT, Claude.ai, Gemini, even a phone app — to make levels by copy-paste. (Pairs perfectly with the tiny optimized prompt.)
Settings → AI Provider → manual
Press Generate — it copies a prompt
The full prompt lands on your clipboard.
Paste into your AI, copy its reply
Any assistant works — paste the prompt, copy the whole answer back.
Press Build
The mod reads your clipboard and builds it — same blue Accept / Edit / Deny preview.
§ 04
Getting started

Install in one .geode file.

Editor AI lives outside the Geode index, so you install it manually from GitHub Releases — about a minute, start to finish.

Why not the index? Geode's guidelines prohibit mods that use AI to place or modify objects, so Editor AI is distributed via GitHub Releases and editorai.pages.dev instead. Manual install is the only difference.
PASS 01
Download the .geode
Grab entity12208.edit-ai.geode from the latest release.
Download ↓
PASS 02
Open GD, then the Geode menu
Launch the game with Geode running and click the Geode icon on the main menu.
PASS 03
Manual Installation → pick the file → restart
Hit Manual Installation (bottom-left of the Geode menu), select the .geode you downloaded, and restart when prompted. The AI button appears in the editor — or press E.
Needs  Geometry Dash 2.2081 + Geode 5.7.1+ + geode.node-ids (auto-installed) incompatible with Editor Collab
§ 05
EditorAI Platinum

Free AI, run by the community.

Platinum is a distributed network: people with spare GPUs run a worker and donate compute; everyone else generates for free. No accounts, no payment, no keys.

Checking the network…
contacting coordinator at sn-1.vltgg.net
—
Workers
—
Queued
—
Processing
Coordinator & servers generously hosted by VLT GG — thank you.

Use Platinum

Generate for free on the community network.
Settings → AI Provider → ollama
Enable Use Platinum
Points to the community network automatically.
Pick a model & generate
e.g. entity12208/editorai:deepseek. No key needed.

Donate compute

Share a GPU or CPU and power free generations for others.
Install Ollama + Python 3.9+
Pull any model(s) you want to serve.
Clone the Platinum repo
git clone https://github.com/entity12208/EditorAI-Platinum
Install deps & run the worker
pip install -r requirements.txt && python worker/client.py
It registers automatically and starts accepting jobs.
Platinum on GitHub →
§ 06
Questions

Good to know.

Five ways to use it for free: Platinum (the community network), local Ollama, LM Studio, llama.cpp, and Manual (copy-paste with any chatbot you already have). Paid cloud: Claude, DeepSeek, OpenAI, OpenRouter, and Mistral need a funded API key. Hugging Face has a small free tier; Gemini has one too but we don't recommend it. The Custom (BYOPAK) option costs whatever your own endpoint does.

For zero cost and zero setup, Platinum. For private local generation, Ollama (6+ GB VRAM helps), LM Studio, or llama.cpp. For the best hosted quality, Claude; for the best value, DeepSeek; for one key across many models, OpenRouter. We don't recommend Gemini due to tight rate limits.

Yes. Start a chat and target the current editor or any of your saved levels. The AI moves, deletes, adds, and restyles objects, all shown as a blue preview first. With the edit-workload target it'll do a heavy rework — a default of around 1,500 edits — instead of nudging a few blocks, and everything is reversible until you accept.

Yes. Each generation is a conversation that survives game restarts — reopen the panel and pick up where you left off. Generations also keep running in the background, so you can close the panel (or the whole editor) while the AI works.

Platinum only works while volunteers have their workers online — if nobody's donating, requests have nowhere to run. Check the live count in the Platinum section above. If it's quiet, switch to local Ollama or a cloud provider, or come back later.

In mod settings — open the Geode menu, tap the gear next to Editor AI, and use the dedicated key field for your provider. For OpenRouter and Hugging Face you can skip the key entirely and use Sign in with browser. Keys stay on your device in Geode save data.

Tune Ground Y in settings (default 105) — lower it if blocks float, raise it if they bury. Editor AI also clamps Y to the ground line and shifts whole levels up when needed.

Max Objects defaults to 500 and goes up to 1,000,000. Set Target Objects to force a minimum count. Objects spawn in batches (the spawn speed is configurable) so the game stays smooth — very large levels just take longer to lay down.

Geometry Dash 2.2081, Geode 5.7.1+, and the geode.node-ids dependency (auto-installed). Runs on Windows, macOS, Android, and iOS. It's incompatible with the Editor Collab mod (alk.editor-collab).

Please do — it keeps Platinum free for everyone. You need Ollama, Python 3.9+, and the worker from the EditorAI-Platinum repo. Run python worker/client.py and your machine registers and starts taking jobs. See Donate compute above.