finds.dev← search

// the find

htdt/godogen

★ 3,458 · Python · MIT · updated Jun 2026

Autonomous game development for Godot and Bevy with Claude Code and Codex

Godogen is a scaffolding layer that turns Claude Code or Codex into an autonomous game developer for Godot 4, Bevy, and Babylon.js. You describe a game, it plans, codes, generates assets via Gemini/Grok/Tripo3D, runs the engine headlessly, and iterates on screenshots until things look right. It's for developers who want to experiment with agentic game prototyping, not for shipping production games.

The frame-grounded self-repair loop is the genuinely interesting idea here — the agent judges progress from captured screenshots rather than from 'it compiled', so visual bugs (wrong scale, missing assets, frozen animation) actually drive the next iteration instead of being silently accepted. Supporting three engines (Godot C#, Bevy Rust, Babylon.js TypeScript) with separate skill trees and a single publish.sh that renders the right layout is a clean abstraction. The optional Telegram stop hook for long server runs is a practical touch that shows someone actually used this on real hardware. Bevy docs are configured locally rather than scraped live, which avoids hallucinated APIs.

The dependency surface is brutal: you need Godot, Rust/Cargo, Node 22+, Chrome/Chromium with hardware WebGL2, Python 3, xvfb, ffmpeg, imagemagick, Vulkan tools, and four separate paid API keys (Google, xAI, Tripo3D, plus Claude or Codex) before you generate a single frame — setup.md is doing a lot of heavy lifting here. The whole pipeline is described as taking 'hours', and there's no checkpoint/resume mechanism documented, so a mid-run crash on a slow machine likely means starting over. With 3 stars, no documented test suite, and a changelog that presumably tracks a fast-moving project, the skill files are probably brittle — prompt engineering that worked last month may break when Claude Code updates. The indirection (godogen → game repo → game) is also conceptually clean but means debugging failures requires understanding three layers of abstraction.

View on GitHub → Homepage ↗

// want more like this?

We dig through GitHub every week and send a few repos picked for what you actually care about — each with an honest take like this one.

Get finds in your inbox → Search again →