🏛️ AI Council
April 15, 2026 · View on GitHub
Pit Claude, GPT-4, Gemini, and local Llama models against each other — on any topic.
AI Council is a blazing-fast, open-source platform for running structured multi-LLM debates. Build a diverse cast of AI personas, back them with different state-of-the-art models, inject unique system prompts, and watch them argue, challenge each other, and converge on a verdict—all streaming live in your browser.
🔒 Privacy First: API keys never leave your browser. Zero backend, zero database. All data is persisted locally.
📸 What it looks like
Creating a meeting — 3 steps
Step 1 — Set the topic
Step 2 — Pick your personas and set the speaking order
Step 3 — Configure rounds, synthesis mode, Autopilot, and Note Taker
Hit Create Meeting and watch the council debate live.
🚀 Why use AI Council?
A single LLM gives you one perspective. AI Council gives you a think tank.
Different models have genuinely different priors. A cautious Claude persona hedges, an aggressive GPT-4 pushes forward, and a local Llama provides a contrarian angle. By forcing them into structured rounds with domain discipline, you surface deep insights you'd never get from a single prompt.
🔥 Killer Use Cases:
- 🏗️ Architecture Reviews: Stress-test system decisions.
- 🕵️ Code Bug Hunts: Pit a Coder vs. a Security Auditor.
- 😈 Product Planning: A "Devil's Advocate" attacking a "Product Manager".
- 🤝 Hiring Rubrics: Run candidate profiles through a tough council panel.
- 🏛️ Philosophy/Ethics: Historical archetypes debating modern dilemmas.
✨ Features
- 🧠 Multi-Model Mastery: Mix Claude, GPT-4, Gemini, and local Ollama models in the same debate seamlessly.
- 🔄 Structured Rounds: Perfect order, full context history, no interruptions.
- 🗣️ @Mentions: Tag a persona mid-meeting for an immediate, targeted response.
- 📝 Synthesis & Note Taker: Choose between individual closing statements, a unified council verdict, or deploy an independent AI observer to log key insights.
- 🪄 Auto-Prompting: Describe a persona, and Claude automatically generates the perfect, highly-tuned system prompt.
- 🛡️ Domain Discipline: Personas are algorithmically forced to stay entirely within their designated lane.
- 🏎️ Autopilot: Let the council debate overnight with zero clicks required.
- 💾 Local-First & Exportable: Everything runs in
localStorage. Export debates perfectly to Markdown or JSON.
📦 Quick Start
Get your council running locally in under 60 seconds:
git clone https://github.com/meisamsharahi/aiCouncil.git
cd aiCouncil
npm install
npm run dev
⚙️ Setup: Go to Settings, paste your API keys (Anthropic, OpenAI, or Google) or link your local Ollama instance, and start your first debate!
🤖 Supported Models
| Provider | Example Models |
|---|---|
| 🟠 Anthropic (Claude) | claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5 |
| 🔵 OpenAI | gpt-4o, gpt-4o-mini, o3-mini |
| 🟡 Google (Gemini) | gemini-2.5-pro, gemini-2.0-flash |
| 🦙 Ollama (Local) | llama3.3, mistral, deepseek-r1, qwen2.5-coder |
Mix and match any of these inside the same meeting!
🛠️ Tech Stack
Built for speed, scale, and Developer Experience (DX):
- ⚡ Next.js 15 (App Router)
- 💎 React 19
- 🛡️ TypeScript (Strict Mode)
- 🐻 Zustand (Local storage persistence)
- 🎨 Tailwind CSS (Premium Dark Mode UI)
- 📡 Server-Sent Events (SSE) (Real-time token streaming)
🤝 Contributing
We love contributions! Before jumping in:
- Check IMPROVEMENTS.md for the active backlog (Context window pruning is our #1 priority!).
- Open an issue to discuss non-trivial changes.
- Keep PRs focused and atomic.
🗺️ Roadmap
- ✂️ Context window pruning for long meetings
- 🔗 Shareable read-only meeting permalinks
- 📚 Persona library (Import/Export)
- 🔀 Branching (Re-run a round with different personas from the same history)
- 🎙️ Voice output (ElevenLabs / Web Speech API Integration)
📜 License
Released under the MIT License.