Godogen turns a game description into a working game — autonomously. “Describe a game. Godogen plans it, writes the code, generates assets, runs the engine, checks screenshots, and fixes what looks wrong.” Supports both Godot 4 (C#) and Bevy (Rust), works with Claude Code or Codex, and includes a unique frame-grounded self-repair loop that fixes visual bugs, not just compilation errors. ~3k GitHub stars.
Source: GitHub - htdt/godogen
How It Works
"Describe a game"
↓
Code generation (engine-specific: C# or Rust)
↓
Asset creation (Gemini → characters, Grok → textures, Tripo3D → 3D models)
↓
Engine execution (run the game)
↓
Screenshot capture + visual inspection
↓
Frame-grounded self-repair
└── Fixes: clipping, scaling, frozen animations, missing assets
↓
Iterate until visually correct
↓
Working game (+ optional Android APK)
The key innovation: frame-grounded self-repair. Instead of just checking if code compiles, Godogen runs the game, captures screenshots, and analyzes them for visible defects. This catches problems that pass compilation but look wrong.
Supported Engines & Agents
| Engine | Agent | Output |
|---|---|---|
| Godot 4 (.NET) | Claude Code | C# project with scenes, scripts, assets |
| Godot 4 (.NET) | Codex | Same, via Codex agent |
| Bevy | Claude Code | Rust/Bevy project with code-first scenes |
| Bevy | Codex | Same, via Codex agent |
Single publish.sh script — engine and agent choice is a publish-time flag, not separate source trees.
./publish.sh --engine godot --agent claude --out ~/my-game
./publish.sh --engine bevy --agent codex --out ~/my-game
Asset Generation Pipeline
| Asset Type | Provider |
|---|---|
| Characters / reference art | Google Gemini |
| Textures / objects | xAI Grok |
| 3D model conversion | Tripo3D |
Requirements
- Godot 4 (.NET build) or Rust/Cargo with Bevy docs
- Python 3, FFmpeg, ImageMagick
- API keys:
GOOGLE_API_KEY,XAI_API_KEY,TRIPO3D_API_KEY - Vulkan support + Xvfb (for headless rendering)
- Full runs can take hours — GPU instances recommended
How LearnAI Team Could Use This
- Game dev courses — students describe games, AI builds them, students analyze and modify the output
- Teaching autonomous agents — the frame-grounded repair loop is a concrete example of visual feedback in agent systems
- Rapid prototyping — quickly generate game prototypes for research or teaching demos
- Cross-engine comparison — same game description, Godot vs Bevy output — great for teaching engine tradeoffs
Real-World Use Cases
- Indie developers — rapid game prototyping from descriptions
- Game jams — accelerate initial development with AI-generated scaffolding
- Education — learn game engine patterns by reading AI-generated code
- Research — study autonomous code generation with visual verification loops