An AI-native narrative RPG where an AI Game Master generates the entire game in real time — narrative, dialogue, world changes, and cutscenes — from whatever you type.
Infinit is a text-driven RPG where an AI Game Master — like a D&D Dungeon Master — generates all narrative, dialogue, and world reactions as you play. There are no dialogue trees and no scripted paths: you type what you want to do, and the world responds.
The AI doesn't just write text — it runs the game. It changes the music, moves NPCs between locations, gives and removes items, and triggers illustrated cutscenes based on your actions. Supporting agents maintain a quest journal and write a summary of your story each time you save, so you can pick up where you left off.
It's set in Solhai, a hand-authored Himalayan fantasy world of mountain villages, ancient monasteries, and smoldering secrets — inspired by months traveling through Nepal and Bhutan. The world is written by hand; the AI's job is to bring it to life, not to invent it. Every player explores the same Solhai, but every playthrough is different because the Game Master responds to your choices.
Infinit vs. AI Dungeon: AI Dungeon generates text in a generated world. Infinit generates a game inside an authored world — the AI issues structured commands that mechanically change game state, so it feels like a tabletop session, not a chatbot.
 |
 |
| AI-generated cutscenes reflect your actual story | NPCs have their own knowledge, secrets, and trust levels |
 |
 |
| The world reacts to whatever you try | Items the GM gives and takes in real time |
 |
 |
| A separate agent tracks your discoveries | Another agent summarizes your story on save |
- A team of AI agents runs the world. A PydanticAI Game Master narrates and emits structured commands; supporting agents handle the quest journal, save summaries, cutscene/image generation, and emotive voice — all streaming in real time.
- The AI advises; code owns the rules. Narrative and taste are the model's job. The deterministic half — inventory, the rules/parser, cost tracking — is plain, tested code. The two are deliberately kept on opposite sides of a trust boundary.
- A real testing pyramid around a fuzzy product. Fast unit tests at the base; behavioral evals in the middle that run the real LLM and deterministically assert what it did (valid commands, items that actually exist, multi-turn consistency, prompt-injection / "god-mode" red-teaming); a human rubric on top for the genuinely subjective stuff — which never gates shipping.
| Layer | Tech |
|---|---|
| Client | Godot 4.5 (GDScript), HTML5 / macOS / Windows builds |
| Backend | Python 3.14, FastAPI, PydanticAI |
| Models | Claude (Anthropic) + Gemini (Google) |
| State / queues | PostgreSQL, Redis |
| Assets | Cloudflare R2 (S3-compatible), Flux image gen, ElevenLabs voice |
| Evals | pydantic-evals, Promptfoo, Claude Code LLM-as-judge, Logfire tracing |
client/ Godot game client (GDScript)
server/ FastAPI backend — agents, game services, evals, scripts
design/ Canon: hand-authored NPCs, locations, lore (JSON)
docs/ Architecture & design notes
itch/ itch.io page config and build settings
Requires Python 3.14, Godot 4.5, Docker (for Postgres + Redis), and API keys in
server/.env.
make install # Set up the server venv
make dev # DB, Redis, server, worker, Godot editor — everything
make server # FastAPI server only
make play-game # Run the game clientcd server && pytest # Fast unit tests (no DB)
cd server && pytest tests/integration # Integration tests (needs DB)
make evals # GM behavioral evals (real LLM calls)
make cc-eval AGENT=journal # Claude Code LLM-as-judge evalI've written about the architecture and the AI-testing approach behind Infinit:
- The Architecture of an AI-Native RPG — how a team of agents runs the game world in real time.
- Automating Exploratory Testing for LLM Apps — vibe-checking an LLM app at scale with a rubric, a CLI, and a coding agent.
- Turn Your Dev Workflows Into CLIs Your Agent Can Use
- How to Add Emotion to AI Voices (ElevenLabs)
- How to Make an LLM Choose From a Very Long List Using RAG
- How to Architect LLM-Powered Web Apps
Built by neon — a solo developer. Infinit is an early, free alpha: things will break, the AI will occasionally hallucinate, and every turn costs real money in inference. Play it free and come say hi in Discord.
MIT © 2026 neon