Wizard Hackathon 2026 Entry
A self-improving AI agent game powered by:
- Nebius AI — LLM inference for agent reasoning & self-improvement
- InsForge — auth, database, and storage backend (agent-native cloud infra)
- Meshy.ai — 3D character generation, rigging & animation
- Claude Code Proxy — agent orchestration (docs): Claude Code routed through a Nebius bridge to build & operate the agents (dev/operator loop, not a runtime dependency of the shipped game)
Players summon AI agents (warrior / mage / rogue / healer) that face procedurally
generated challenges. Each agent reasons with a Nebius LLM, then periodically
analyzes its own performance and rewrites its strategy — a self-improving loop.
Every decision is persisted to InsForge as a memory, and completed runs land on
an InsForge-backed leaderboard. 3D characters are generated with Meshy and their
.glb models stored in InsForge Storage.
Game Client (React + Three.js / WebGL)
│
├── Nebius AI ............ LLM inference + self-improvement reasoning
├── InsForge ............. auth · database (agent_runs, agent_memories) · storage (characters)
└── Meshy.ai ............. text-to-3D character pipeline → InsForge Storage
Target: Best Use of InsForge + Nebius AI Integration
# 1. Link or create an InsForge project
npx @insforge/cli link # or: npx @insforge/cli create
# 2. Apply the schema (agent_runs, agent_memories + RLS)
npx @insforge/cli db query --file docs/insforge-schema.sql
# 3. Create the storage bucket for Meshy models
npx @insforge/cli storage create-bucket characters --public
# 4. Grab credentials for .env
npx @insforge/cli secrets get ANON_KEY # → VITE_INSFORGE_ANON_KEY
# URL = oss_host in .insforge/project.json → VITE_INSFORGE_URLThe game runs without InsForge configured — it falls back to
localStorageand shows "local mode" in the HUD. Add the env vars to light up the live backend.
cp .env.example .env # fill in your keys
npm install
npm run dev # http://localhost:3000
npm run build # production build