OwlBrain pits five specialised AI agents against each other in structured, multi-round debates about any business case you throw at them. Each agent can run on a different LLM — Claude argues against GPT argues against Gemini — and an independent judge scores consensus after every round. When the dust settles, a synthesizer produces a final verdict backed by the full transcript.
No opinions. No single-model bias. Just structured adversarial reasoning that converges on the strongest answer.
You ask ChatGPT a question — you get one model's opinion. You ask Claude — you get another. You ask Gemini — you get a third. None of them challenge each other. None of them evolve their reasoning. None of them tell you where they disagree.
OwlBrain fixes that.
Five agents with distinct roles, each assignable to any LLM:
| Agent | Role | What It Does | |
|---|---|---|---|
| 🎯 | The Strategist | Strategy Lead | Builds the core recommendation and execution plan |
| 📊 | The Analyst | Quant Expert | Stress-tests claims with data, numbers, and evidence |
| Risk Officer | Risk & Mitigation | Identifies failure modes, legal exposure, and downside scenarios | |
| 🚀 | The Innovator | Creative Lead | Proposes unconventional approaches and reframes the problem |
| 😈 | Devil's Advocate | Challenger | Attacks the strongest position — specifically to find its weaknesses |
Each agent has a distinct persona, configurable model, and independent memory. They reference each other's arguments, track stance shifts, and call out sycophantic agreement.
You submit a business case
│
▼
┌─────────────────────────────────────────────┐
│ ROUND 1 → ROUND N │
│ │
│ 🎯 Strategist ──► 📊 Analyst ──► ⚠️ Risk │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ 🚀 Innovator ──────► 😈 Devil's Advocate │
│ │
│ ┌──────────────────────┐ │
│ │ Consensus Judge │ │
│ │ scores 5 dimensions │ │
│ │ detects sycophancy │ │
│ └──────────┬───────────┘ │
│ │ │
│ threshold met? ──► yes ──► STOP │
│ no ──► compress memory ──► ↑ │
└─────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────┐
│ FINAL VERDICT │
│ Synthesizer produces a structured report │
│ backed by the full debate transcript │
└─────────────────────────────────────────────┘
Consensus is scored across five dimensions:
| Dimension | What the Judge Evaluates |
|---|---|
| Core Recommendation | Is there a clear, unified recommendation? |
| Reasoning | Is the reasoning well-supported and logically sound? |
| Execution Plan | Are there concrete, actionable steps? |
| Risk Assessment | Have risks been identified and addressed? |
| Stability | Have positions converged or are agents still shifting? |
🔀 Multi-LLM Orchestration — Assign different models to different agents. Run Claude as your strategist, GPT as your analyst, and Gemini as your devil's advocate in the same debate.
🧠 18 Models Across 3 Providers — Claude Opus 4.6, Sonnet 4.6, Haiku 4.5 · GPT-5.4, GPT-5, o4-mini, o3, GPT-4o · Gemini 3.1 Pro, 3 Flash, 2.5 Pro, 2.5 Flash — and more. Adding a new model is one catalog entry.
⚖️ Consensus Scoring — An independent judge evaluates the panel after each round. Scores range 0–100 with per-dimension breakdowns and written rationale.
🔍 Sycophancy Detection — The judge flags when agents agree too readily without substantive reasoning. Shallow consensus gets called out, not rewarded.
🔄 Two Debate Modes — Sequential: agents see previous responses and build on them. Independent: agents argue in isolation to eliminate anchoring bias.
📡 Real-Time Streaming — Watch the debate unfold live via Server-Sent Events. Responses stream as they're generated.
🧾 Memory & Stance Tracking — Tracks every agent's position across rounds, detects stance shifts, compresses history, and feeds structured memory bundles into later rounds.
🌐 Web Search — Any agent can ground arguments in real-time web data. Supported across Anthropic, OpenAI, and Google.
🎛️ Full Control Room — Configure every parameter: models, temperature, reasoning effort, thinking budgets, web search, frequency penalties, word limits, system prompts, and presets.
📝 Synthesis & Verdict — A dedicated synthesizer produces a structured final report that references the full transcript, acknowledges dissent, and delivers an actionable recommendation.
💾 SQLite Persistence — Every debate is saved with full round data, scores, event logs, config snapshots, and token usage. Browse or replay any past debate.
🔒 Demo Mode — Host a public demo with built-in rate limiting, token caps, model locking, and showcase debates. One env var activates everything.
Anthropic
| Model | Tier | Web Search | Structured Output |
|---|---|---|---|
| Claude Opus 4.6 | Flagship | ✅ | ✅ |
| Claude Sonnet 4.6 | Balanced | ✅ | ✅ |
| Claude Sonnet 4.5 | Balanced | ✅ | ✅ |
| Claude Haiku 4.5 | Fast | ✅ | ✅ |
OpenAI
| Model | Tier | Reasoning | Web Search | Structured Output |
|---|---|---|---|---|
| GPT-5.4 | Flagship | ✅ | ✅ | — |
| GPT-5.4 Pro | Flagship | ✅ | ✅ | — |
| GPT-5 | Reasoning | ✅ | ✅ | — |
| GPT-5 Mini | Budget | ✅ | ✅ | — |
| o4-mini | Reasoning | ✅ | ✅ | — |
| o3 | Reasoning | ✅ | ✅ | — |
| GPT-4.1 | Chat | — | ✅ | ✅ |
| GPT-4o | Chat | — | ✅ | ✅ |
| Model | Tier | Thinking | Web Search | Structured Output |
|---|---|---|---|---|
| Gemini 3.1 Pro | Flagship | ✅ | ✅ | ✅ |
| Gemini 3 Flash | Fast | ✅ | ✅ | ✅ |
| Gemini 3.1 Flash-Lite | Budget | ✅ | ✅ | — |
| Gemini 2.5 Pro | Balanced | ✅ | ✅ | ✅ |
| Gemini 2.5 Flash | Balanced | ✅ | ✅ | ✅ |
| Gemini 2.5 Flash-Lite | Budget | ✅ | ✅ | ✅ |
💡 Adding a new model? One entry in
server/modelCatalog.ts. That's it.
Prerequisites: Node.js 20+ and at least one API key from Anthropic, OpenAI, or Google AI.
git clone https://github.com/nasserDev/OwlBrain.git
cd owlbrain
npm install
cp .env.example .env # Add your API keys
npm run dev # Open http://localhost:5000That's it. Five agents, ready to argue.
npm install -g @railway/cli
railway login
railway init
railway upSet your API keys in the Railway dashboard. The app builds and deploys automatically.
Everything is configurable from the Control Room in the UI. For environment-level defaults:
# Provider API Keys (set at least one)
ANTHROPIC_API_KEY=your-key
OPENAI_API_KEY=your-key
GEMINI_API_KEY=your-key
# Server
PORT=5000
ADMIN_API_KEY=your-admin-password
# Demo Mode (for public deployments)
DEMO_MODE=false
DEMO_DAILY_TOKEN_CAP=5000000
DEMO_MAX_DEBATES_PER_IP_PER_DAY=3
DEMO_MAX_ROUNDS=3See .env.example for every option.
owlbrain/
├── client/src/
│ ├── pages/ DebatePage · AdminPage · ShowcasePage
│ ├── components/ AuthGate · DemoBanner · UI library
│ ├── context/ DemoStatusContext
│ └── lib/ Query client · debate state · utilities
│
├── server/
│ ├── orchestrator.ts Debate execution engine
│ ├── llmRouter.ts Multi-provider LLM routing
│ ├── consensus.ts Consensus judge & 5-dimension scoring
│ ├── synthesizer.ts Final verdict generation
│ ├── memory.ts Stance tracking & round compression
│ ├── promptBuilder.ts Prompt construction per round & mode
│ ├── configStore.ts SQLite persistence layer
│ ├── modelCatalog.ts Supported model registry
│ ├── demoConfig.ts Demo mode restrictions
│ ├── demoGuards.ts Rate limiting & budget protection
│ └── routes.ts API endpoints & SSE streaming
│
├── shared/ Shared types & validation schemas
├── .env.example Environment template
├── railway.toml Railway deployment config
└── nixpacks.toml Build environment config
Tech Stack
| Layer | Technology |
|---|---|
| Frontend | React 18 · TanStack Query · Tailwind CSS · shadcn/ui |
| Backend | Node.js 20 · Express 5 · TypeScript |
| Database | SQLite via better-sqlite3 |
| LLM SDKs | @anthropic-ai/sdk · openai · @google/genai |
| Build | Vite · esbuild · tsx |
| Streaming | Server-Sent Events |
Hosting OwlBrain for the public? One environment variable protects your budget:
DEMO_MODE=trueWhat activates:
| Protection | Default |
|---|---|
| Per-IP debate limit | 3 per day |
| System-wide daily token cap | 5M tokens |
| Per-debate token cap | 150K tokens |
| Model selection | Locked to cheapest viable combo |
| Thinking / reasoning tokens | Disabled |
| Admin panel | Read-only (operator password for changes) |
| Showcase debates | Pre-seeded as instant fallback |
| Capacity messaging | Graceful redirect to showcase |
Every restriction is configurable. Every guard is server-enforced. The same code runs in demo and full mode — the flag just activates the restriction profile.
Contributions are welcome. Open an issue first to discuss what you'd like to change.
npm run dev # Dev server with hot reload
npm run check # TypeScript type checking
npm test # Run test suite
npm run build # Production build
npm start # Start production serverOwlBrain is source-available under the Business Source License 1.1.
Free for personal use, research, education, evaluation, and non-commercial projects.
Commercial use requires a separate license. Contact me for details.
The license converts to Apache 2.0 on the change date specified in the LICENSE file.
Built with the same models that power the debates: Claude, GPT, and Gemini.
Built by nasserDev
