Multi-agent AI ensemble for Kalshi prediction markets. Five specialized LLM agents debate every market, reach consensus with Kelly-sized positions, and execute in paper or live mode — with full SQLite audit trail.

Most "AI trading bots" are a single prompt and a prayer. This one runs a structured pipeline:

1. Ingest — Kalshi markets, news, and context
2. Debate — five agents analyze in parallel, then synthesize
3. Consensus — weighted ensemble with disagreement penalty
4. Size — Kelly criterion with hard risk limits
5. Execute — paper or live via Kalshi REST + WebSocket
6. Track & evaluate — SQLite audit, dashboard, agent scorecards

You're not betting on one model's hunch — you're betting on a committee that must agree.

Consensus rules: minimum 3 models, disagreement penalty applied, min confidence 0.45.

If the ensemble is disabled, falls back to primaryModel + fallbackModel via ModelRouter.

• Paper trading — full pipeline without risking capital
• Live trading — gated by LIVE_TRADING_ENABLED
• Daily AI cost cap — DAILY_AI_COST_LIMIT prevents runaway LLM bills
• Model health routing — unhealthy models auto-demoted
• Multiple strategies — Safe Compounder, Market Making, Quick-Flip, category scorer
• SQLite audit — every decision logged to trading.db
• CLI dashboard — status, scores, history at a glance

• Node.js 22.5+ (native SQLite support)
• Kalshi API credentials
• OpenRouter API key

git clone https://github.com/openfi-dao/kalshi-trading-bot.git
cd kalshi-trading-bot
npm install
cp env.template .env

KALSHI_API_KEY=your_key
KALSHI_PRIVATE_KEY_PATH=./kalshi.pem
OPENROUTER_API_KEY=sk-or-...

# Safety defaults
LIVE_TRADING_ENABLED=false
DAILY_AI_COST_LIMIT=10
LOG_LEVEL=info

npm run bot
# or
npm run dev

kalshi-bot run [--live] [--daily-limit n] [--iterations n]
kalshi-bot dashboard
kalshi-bot status
kalshi-bot scores
kalshi-bot history [--limit n]
kalshi-bot safe-compounder [--live]
kalshi-bot health

flowchart LR
    Ingest[Market Ingest] --> Agents[5-Agent Parallel Analysis]
    Agents --> Debate[Debate Round]
    Debate --> Consensus[Weighted Consensus]
    Consensus --> Kelly[Kelly Sizing]
    Kelly --> Limits[Risk Limits]
    Limits --> Execute[Paper / Live]
    Execute --> Track[SQLite Audit]
    Track --> Eval[Agent Evaluation]

Configured in src/config/settings.ts:

src/
├── cli.ts              # Command-line interface
├── beastModeBot.ts     # Main orchestrator
├── agents/             # Ensemble + debate logic
├── clients/
│   ├── KalshiClient.ts # REST API
│   ├── KalshiWS.ts     # WebSocket
│   └── ModelRouter.ts  # LLM routing + failover
├── jobs/
│   ├── ingest.ts
│   ├── decide.ts
│   ├── trade.ts
│   ├── track.ts
│   └── evaluate.ts
├── strategies/         # Safe-compounder, MM, quick-flip
├── paper/              # Paper tracker + dashboard
└── utils/database.ts   # SQLite layer
trading.db              # Audit log (created at runtime)

Run with: kalshi-bot safe-compounder [--live]

npm test          # Vitest: ensemble, Kelly, DB, JSON repair
npm run typecheck
npm run lint
npm run build

• ModelRouter demotes unhealthy LLM providers automatically
• ensemble.enabled=false → single primary + fallback model
• All decisions persisted — replay and debug from trading.db

• Start in paper mode. LLM outputs are probabilistic, not oracle predictions.
• Set a daily AI cost limit before enabling live trading.
• Kalshi markets involve real money and regulatory constraints — know your jurisdiction.
• Authors not liable for trading or API costs.

Issues and PRs welcome — especially agent prompt improvements, backtesting hooks, and safer default limits.