Jev Town

The first Jev-based AI simulation system.

Jev Town is a game engine where the characters are driven by AI, not scripts. Every agent in the world decides what to do next through two kinds of models working together:

• Jev, a System One model, makes the fast, structured decisions: whether to seek someone out, who to go to, where to wander, how long to wait. The engine offers the options; Jev chooses within them, so an illegal move is never possible.
• An LLM does the language work: conversations between agents, memory, and what each agent believes about the world.

Drop a cast of agents into a pixel-art room and watch what emerges — nobody is playing. The whole simulation runs in a single browser tab; a small local proxy holds your model keys so the browser never sees them. It is an MVP and still taking shape.

Coming soon

We are building a game on top of this engine: a pixel-art narrative world where every character you meet is a Jev-driven agent with a life of its own. Star the repository to follow along.

Installation

The Jev demo currently lives on feat/jev-demo-solarium until it merges to main.

Requires Node.js 22 LTS (22.22.1 or a newer patch) and npm.

1. Install dependencies:

   npm ci

2. Create .env.local in the repository root with your model credentials:

   # LLM (any OpenAI-compatible endpoint)
   LLM_API_URL=https://your-openai-compatible-endpoint/v1
   LLM_API_KEY=...
   LLM_MODEL=...
   LLM_EMBEDDING_API_URL=https://your-embedding-endpoint/v1
   LLM_EMBEDDING_API_KEY=...
   LLM_EMBEDDING_MODEL=...
   
   # Jev (System One decider)
   JEV_API_KEY=...
   VITE_ACTION_DECIDER=jev

   OPENAI_API_KEY, TOGETHER_API_KEY, and a local Ollama host also work for the LLM. Leave out VITE_ACTION_DECIDER to let the LLM make every decision instead of Jev.

3. Run the demo:

   npm run play:demo

4. Open http://localhost:5173/ai-town/ and watch the agents go. Set VITE_DEMO_AGENTS=n in .env.local to change the cast size (five by default, up to fifty).

Every decision and every line of dialogue is a model call, so a large cast costs more. The proxy stops after 2000 model calls per run as a spend backstop; restart it to continue.

License and credits

This project started from a16z-infra/ai-town and is directly inspired by it: the agent loop, the simulation shape, and parts of the client all trace back to that work, and the project still uses assets from it. Upstream code is © 2023 a16z-infra and distributed under the MIT License; attribution for the initial snapshot belongs to the upstream authors.

Original code and modifications by NevaMind-AI contributors are likewise MIT licensed, with copyright retained by each contributor. Story, art, music, and third-party assets may carry their own licenses, which take precedence where they apply; the code license grants no rights over player game data.

Upstream assets and foundational work:

• Map tiles: George Bailey and hilau.
• Original art: ansimuz.
• UI assets: Mounir Tohami.
• The upstream proof of concept was based on phaser3-simple-rpg.
• Character rendering uses PixiJS.