Multiple minds. One answer.
AI Council is a local web application that queries multiple Large Language Models simultaneously, has them peer-review each other's responses, and synthesizes a final answer through a designated Chairman. Instead of asking one AI, ask a council.
- Reduce bias — Different models have different training data and perspectives
- Improve accuracy — Cross-validation catches errors individual models might make
- Get comprehensive answers — Synthesis combines the best of all responses
- Compare models — See how GPT-4, Claude, Gemini, and others approach the same question
┌─────────────────────────────────────────────────────────────────┐
│ YOUR QUESTION │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 1: INDEPENDENT RESPONSES │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ GPT-4 │ │ Claude │ │ Gemini │ │ Llama │ │
│ └────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘ │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ Response A Response B Response C Response D │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 2: PEER REVIEW │
│ Each model ranks the OTHER responses (anonymized) │
│ "Response 2 is most accurate because..." │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ STAGE 3: SYNTHESIS │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ CHAIRMAN MODEL │ │
│ │ Integrates all perspectives + rankings into one answer │ │
│ └─────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ FINAL ANSWER │
│ Comprehensive, cross-validated, synthesized response │
└─────────────────────────────────────────────────────────────────┘
| Requirement | Version | Notes |
|---|---|---|
| Python | 3.10+ | python3 --version |
| Node.js | 18+ | node --version |
| OpenRouter API Key | - | Get free key |
# 1. Clone the repository
git clone https://github.com/yourusername/ai-council.git
cd ai-council
# 2. Run setup (creates .env, installs dependencies, initializes database)
./scripts/setup.sh
# 3. Add your OpenRouter API key
nano .env
# Change: OPENROUTER_API_KEY=sk-or-v1-your-actual-key-here./start.shOpen http://localhost:3847 in your browser.
# Terminal 1: Backend
cd backend
pip install -r requirements.txt
python main.py
# Terminal 2: Frontend
cd frontend
npm install
npm run devPresets — Ready-made council configurations:
| Preset | Description | Best For |
|---|---|---|
| Simple Discussion | 3 models + chairman | Quick questions |
| Debate Council | Pro vs Con + moderator | Controversial topics |
| Expert Panel | Technical, Business, Legal, Creative | Complex decisions |
| Devil's Advocate | Challenge-response format | Stress-testing ideas |
| Research Council | 7 diverse perspectives | Deep analysis |
Custom — Build your own council from scratch:
- Click Edit → Models tab to see all 200+ available models
- Drag models onto the canvas to add them
- Click any model to configure its role, system prompt, and personality
- Designate one model as Chairman to synthesize the final answer
- Save your custom configuration for reuse
Type your question in the input bar at the bottom and press Enter or click Run.
- Stage 1: See each model's response stream in real-time
- Stage 2: View peer rankings and reasoning
- Stage 3: Get the synthesized final answer
Click the History icon to:
- View past council sessions
- Replay previous configurations
- See cost and token usage
# Required
OPENROUTER_API_KEY=sk-or-v1-your-key-here
# Optional
PORT=8347 # Backend port
HOST=0.0.0.0 # Bind address
CORS_ORIGINS=http://localhost:3847 # Frontend URL
MAX_BUDGET=10.0 # Max spend per hour (USD)AI Council uses OpenRouter which provides access to 200+ models. Models are fetched live from OpenRouter, so you always have the latest releases.
Models used in presets:
| Model ID | Used As |
|---|---|
anthropic/claude-opus-4.5 |
Chairman / Synthesizer |
anthropic/claude-sonnet-4.5 |
Primary Responder / Expert |
openai/gpt-5.2-chat |
Analyst / Advocate |
openai/gpt-5-pro |
Integrator |
google/gemini-3-pro-preview |
Critic / Legal Expert |
deepseek/deepseek-r1 |
Creative / Visionary |
nvidia/llama-3.3-nemotron-super-49b-v1.5 |
Pragmatist |
Browse all models: In Edit mode, click the Models tab to see 200+ models with live pricing from OpenRouter.
In Edit mode, you have full control over your council:
Adding Models:
- Click Models tab in the sidebar
- Browse or search 200+ available models
- Drag any model onto the canvas
- Each model shows estimated cost per query
Configuring Participants:
- Click any model node to open settings
- Set a Display Name and Role (Analyst, Devil's Advocate, etc.)
- Write a System Prompt to define behavior
- Adjust Temperature (0.0 = focused, 1.0 = creative)
Removing Models:
- Select a node and press Delete, or
- Right-click and choose Remove
Designating Chairman:
- Click a model node → Set as Chairman
- The Chairman synthesizes all responses into the final answer
| Feature | Description |
|---|---|
| Visual Council Builder | Drag-and-drop canvas to design custom councils |
| Real-time Streaming | Watch responses arrive live via WebSocket |
| Peer Review Rankings | See how models rate each other's responses |
| Conversation History | Review and replay past council sessions |
| Cost Tracking | Token usage and estimated costs per query |
| Rate Limiting | Built-in protection against API overuse |
| Prompt Injection Defense | Security against malicious inputs |
| PII Redaction | Sensitive data automatically redacted from logs |
ai-council/
├── backend/ # Python FastAPI server
│ ├── main.py # API endpoints & WebSocket
│ ├── council.py # Council execution orchestration
│ ├── openrouter.py # OpenRouter API client
│ ├── database.py # SQLite operations
│ ├── security.py # Prompt injection defense, PII redaction
│ ├── rate_limiter.py # Request & cost limiting
│ ├── requirements.txt # Python dependencies
│ └── data/ # SQLite database (gitignored)
│
├── frontend/ # React 19 application
│ ├── src/
│ │ ├── components/ # UI components
│ │ │ ├── canvas/ # ReactFlow council builder
│ │ │ ├── panels/ # Side panels (config, history, results)
│ │ │ └── layout/ # Header, navigation
│ │ ├── stores/ # Zustand state management
│ │ └── utils/ # Helpers & preset definitions
│ └── package.json
│
├── scripts/
│ ├── setup.sh # Initial setup script
│ ├── backup-data.sh # Backup local database
│ └── reset-data.sh # Reset to fresh state
│
├── docs/
│ └── SECURITY_REPORT.md # OWASP security audit
│
├── start.sh # Launch both servers
├── .env.example # Environment template
└── .gitignore # Excludes data, logs, .env
| Layer | Technology | Purpose |
|---|---|---|
| Backend | FastAPI | Async API server |
| Frontend | React 19 + Vite | Modern UI framework |
| State | Zustand | Lightweight state management |
| Canvas | ReactFlow | Visual node-based builder |
| Animation | Framer Motion | Smooth UI transitions |
| Styling | Tailwind CSS | Utility-first CSS |
| API | OpenRouter | Unified LLM access |
| Database | SQLite | Local conversation storage |
| WebSocket | FastAPI WS | Real-time streaming |
# Check if .env exists and has valid key
cat .env | grep OPENROUTER
# Key should look like: sk-or-v1-abc123...
# NOT: sk-or-v1-your-key-here# Kill process on port 8347 (backend)
lsof -ti:8347 | xargs kill -9
# Kill process on port 3847 (frontend)
lsof -ti:3847 | xargs kill -9# Reinstall backend dependencies
cd backend && pip install -r requirements.txt
# Reinstall frontend dependencies
cd frontend && rm -rf node_modules && npm install- Ensure backend is running on port 8347
- Check CORS_ORIGINS in .env matches frontend URL
- Look for errors in backend terminal
- OpenRouter routes to the fastest available provider
- Some models (GPT-4, Claude Opus) are naturally slower
- Check your internet connection
- Try budget-tier models for faster responses
# All tests
cd frontend && npm test
# With coverage
npm run test:coverage
# Watch mode
npm run test -- --watchTest Coverage: 90/90 tests passing
AI Council includes several security features:
- Prompt Injection Defense — Detects and blocks injection attempts
- PII Redaction — Emails, phones, SSNs automatically redacted from logs
- Rate Limiting — Prevents API abuse (requests/minute, cost/hour)
- Input Validation — Max length, sanitization
- Parameterized SQL — No SQL injection vulnerabilities
- Security Headers — X-Frame-Options, X-Content-Type-Options, etc.
See docs/SECURITY_REPORT.md for the full OWASP audit.
OpenRouter pricing varies by model. Preset estimated costs:
| Preset | Models | Est. Cost |
|---|---|---|
| Simple Discussion | 3 models + chairman | ~$0.05 |
| Debate Council | 4 models | ~$0.12 |
| Expert Panel | 5 models + chairman | ~$0.20 |
| Devil's Advocate | 4 models | ~$0.15 |
| Research Council | 7 models | ~$0.40 |
Costs are shown in real-time in the Edit sidebar. Set MAX_BUDGET in .env to limit hourly spending. Check OpenRouter pricing for current model rates.
Contributions welcome! Please:
- Fork the repository
- Create a feature branch (
git checkout -b feature/amazing) - Commit changes (
git commit -m 'Add amazing feature') - Push to branch (
git push origin feature/amazing) - Open a Pull Request
MIT License — see LICENSE for details.
- OpenRouter for unified LLM API access
- ReactFlow for the visual builder
- Framer Motion for animations
- Built with Claude, GPT-4, and Gemini collaboration
Questions? Open an issue or start a discussion.
Like this project? Give it a star!