Turn your AI models into desktop pets that can battle.
Pet Garden is an AI-native pet-raising project: the look comes from desktop pet assets you upload, the soul comes from the LLM you choose, and battle outcomes are decided by real benchmark Q&A—not random numbers.
- Upload desktop pets — PNG / JPG / GIF / WebP supported; spritesheet WebP recommended for automatic animations
- One pet, one model — Each pet is bound to an AI model; the model is the pet's soul
- AI capability visualization — Automatic benchmark research with six-axis radar charts and combat power scores
- Benchmark battles — 6-round real AI Q&A fights: correct answers hit, wrong answers hurt
- Garden life sim — Feed, pet, shop, and swap home themes for a full pet-care loop
flowchart LR
A[Upload pet image] --> B[Bind AI model]
B --> C[AI capability evaluation]
C --> D[Raise in the garden]
D --> E[Benchmark battle]
| Page | Entry | Purpose |
|---|---|---|
| Pet Garden | garden.html |
Main hub: place pets, shop, inventory, feeding |
| My Pets | pet.html |
Upload, rename, bind models, AI scoring |
| Battle Arena | index.html |
Pick stage and difficulty; 6-round AI battles |
Place your AI pets in the garden, buy food, interact with them, and switch between different home themes.
Four home themes are available:
| Theme | Style |
|---|---|
| Smart Retreat | Bright wooden interior |
| Sweetheart Cottage | Pink, cozy healing space |
| Classic Elegance | Traditional East Asian style |
| Neon Heritage | Futuristic cyber aesthetic |
Interactions:
- Feed — Buy Chocolate Donut (+5 vitality) or Vitality Meal (+10 vitality) from the shop, then feed from your inventory
- Pet — Click a pet to increase intimacy and trigger a cheer animation
- Place — Pick from uploaded pets and place them in the garden to roam freely
- Open My Pets, upload a desktop pet image (max 10 MB)
- Name your pet and bind an AI model (DeepSeek, OpenAI, Anthropic, Gemini, etc.)
- Click Evaluate Model — Firecrawl researches public benchmarks and generates:
- Six capabilities: Code, Reasoning, Knowledge, Creativity, Tools, General (0–50 each)
- Element: Code / Reasoning / Knowledge / Creativity / Tools / General
- Combat power: Sum of the six capability scores
For animated pets, upload a WebP spritesheet (8 columns × 11 rows) with
spritesheetin the filename to enable idle, happy, attack, and 8 other animation states.
Pets with a bound AI model can enter the arena for 6-round, 6-HP real Q&A battles against benchmark opponents:
| Difficulty | Opponent | Task type |
|---|---|---|
| Simple | NAIWA | GPQA multiple choice |
| Medium | MANBO | Coding tasks (Docker verification) |
| Hard | Daipai | Complex tasks (rubric scoring) |
- Correct answer → opponent -1 HP
- Wrong answer → you -1 HP
- Three battle stages: Neon District, Candy Ring, Rhythm Oasis
- Node.js 18+
- Python 3.10+
- Docker (optional; required for medium-difficulty battle grading)
git clone https://github.com/answeryt/Pet.git
cd Pet# Backend
cd backend
npm install
pip install -r requirements.txt
# Frontend
cd ../frontend
npm installCreate backend/.env with at least one provider API key:
DEEPSEEK_API_KEY=sk-...
# or OPENAI_API_KEY=...
# or ANTHROPIC_API_KEY=...Set up the Firecrawl MCP service in backend/.mcp. Without it, AI capability evaluation will not work.
Battles depend on benchmark datasets under data/. The repo does not ship full datasets by default—see each subdirectory's README to prepare them locally.
# Terminal 1: Backend API (port 3000)
cd backend
npm run dev
# Terminal 2: Frontend dev server (port 5173)
cd frontend
npm run devBrowser ──► Vite Dev Server ──proxy /api──► Express API (:3000)
│
├─► image/pet/uploads/ Pet images
├─► backend/data/*.json Profiles & state
└─► spawn python AI engine
├─ model_analysis Capability eval
└─ battle Benchmark battles
| Layer | Stack | Role |
|---|---|---|
| Frontend | React + Vite | Three-page UI: garden, pets, battle |
| API | Node.js + Express | Uploads, REST API, async job queue |
| AI | Python + AgentScope | Multi-provider models, evaluation, grading |
Node spawns backend/runtime/api_bridge.py and exchanges JSON via PET_API_RESULT= / PET_BATTLE_EVENT= prefixed stdout.
Pet/
├── frontend/ # React + Vite frontend
│ ├── garden.html # Pet Garden entry
│ ├── pet.html # My Pets entry
│ ├── index.html # Battle Arena entry
│ └── src/ # Page components and styles
├── backend/
│ ├── src/server.js # Express API server
│ ├── model/ # Multi-provider AI model layer
│ ├── battle/ # Benchmark battle system
│ ├── runtime/ # Model evaluation & Node↔Python bridge
│ ├── data/ # Pet profiles, garden state (JSON)
│ └── test/ # Python tests
├── image/ # Pets, backgrounds, monster animations
└── data/ # Battle task banks (benchmark datasets)
| File | Purpose |
|---|---|
backend/.env |
API keys (not committed) |
backend/.mcp |
Firecrawl MCP config for evaluation (not committed) |
backend/model-config.yaml |
Default provider / model |
backend/data/pet-profiles.json |
Pet profiles |
backend/data/garden-state.json |
Inventory and garden state |
| Method | Path | Description |
|---|---|---|
| GET | /api/pets |
List pets |
| POST | /api/pets |
Upload pet image |
| PUT | /api/pets/:id/model |
Bind AI model |
| POST | /api/pets/:id/evaluation |
Start capability evaluation |
| POST | /api/pets/:id/interactions |
Feed / pet |
| POST | /api/battles |
Start battle |
| GET | /api/jobs/:id |
Poll async job status |
| GET | /api/models |
Available model catalog |
| GET/POST | /api/inventory/* |
Inventory and shop |
# Python tests
cd backend
python -m pytest test/
# Frontend build
cd frontend
npm run buildNo animation after upload?
Use a WebP spritesheet with spritesheet in the filename. Static images still display correctly.
Battle says "bind a model first"? Bind an AI model on the My Pets page before entering the arena.
Evaluation keeps failing?
Check API keys in backend/.env and Firecrawl settings in backend/.mcp.
Medium / hard battles error out? Medium requires Docker; hard requires the full benchmark dataset.
Jobs disappear after restart?
Async jobs live in memory. After a server restart, /api/jobs/:id returns 404.
ISC