Build a 3D game by describing it
Type what you want. The assistant adds objects, materials, lights and Lua scripts to your scene, runs the physics, takes screenshots to check its work, and publishes to a link. Everything it makes is yours to open and edit.
Two ways to build with AI
1. The built-in assistant
Open the chat panel in the editor and describe what you want: a racing track with three hairpins, a sword-wielding enemy that telegraphs its attacks, a HUD timer. The assistant creates the objects and writes the Lua, and asks a question when a brief is ambiguous. Every plan includes an AI allowance for it: a little each day on Free, a monthly amount on Creator and Studio, then your credit balance. See the AI assistant guide.
2. Your own AI tool, over MCP
The editor exposes a Model Context Protocol server. Connect Claude, ChatGPT, Codex or any MCP client with OAuth, pick the scopes you want to grant, and that tool can do everything the built-in assistant can: inspect the project, create and edit objects, write scripts, generate meshes, textures and audio, run the simulation, take screenshots, write tests, configure multiplayer and publish. Connecting your own tool does not use your Airogel AI allowance. The MCP tool guide maps every tool.
Proof: every demo was built this way
None of the sixteen demos was built by hand. Each started as a written brief, from a paragraph to a short design document, and an AI agent connected over MCP did the rest: it built the scene, wrote the scripts, imported or generated assets, configured rooms where the brief asked for multiplayer, tested what it made, and published. Every later change was made the same way. Each demo page shows the brief it started from and what the agent did with it, for example The Drowned Valley, an open-world slice with rising water, companion AI and co-op rooms, or CityScape, a city block of more than 3,000 objects.
What makes it work
- Real physics, no setup. The agent places a car and it drives; it places a ragdoll and it falls. It does not have to configure colliders or write a physics loop.
- One language. Everything is Lua, so the agent (and you) can read any script in any project.
- It can see and test. The MCP tools include running the simulation and taking screenshots, so the agent checks its own work, and a test-rule API so it can write regression tests for scripts.
- It can ask for help. When an agent hits an engine limit it can file a support ticket from inside the editor. Several of those tickets became engine features.
- Publishing is a tool call. A publish check, then a public link; republishing keeps the same address.
You stay in control
Nothing the assistant makes is locked. Objects appear in the scene tree, scripts open in the editor, and you can change anything by hand or by asking again. If you prefer to write every line yourself, the assistant is optional; see Lua scripting.
Frequently asked
Do I need to know how to code?
Not to start. Describe the game and the assistant writes the scripts. When you want more control, every script is Lua you can open and edit.
Which AI models can I connect?
Any MCP client. The docs cover Claude, ChatGPT and Codex, and the same OAuth flow works for other clients that implement MCP.
What does the AI allowance cover?
Only the built-in assistant. External tools connected over MCP use their own provider's billing. See pricing.
Who owns what the AI builds?
You do. It is your project, your objects and your scripts, published under your game's own address.
Free to start. Nothing to install.