openfi-dao / kalshi-trading-bot

What is kalshi-trading-bot?

This is a TypeScript trading bot that connects to Kalshi prediction markets and uses five AI agents working in committee to decide when to buy and sell. Each market gets analyzed by a forecaster, news analyst, bull researcher, bear researcher, and risk manager that debate before reaching consensus. The system sizes positions using Kelly criterion, executes trades via REST API or WebSocket, and logs everything to SQLite. You can run it in paper mode to test without risking real money, or flip a flag to enable live trading. The CLI lets you check your dashboard, view history, inspect category scores, and run different strategies like Safe Compounder or market making.

Why is it gaining traction?

The multi-agent debate architecture is more sophisticated than most "AI trading bot" repos that just call one model and hope. By requiring multiple agents to agree and penalizing disagreement, it reduces the chaos of individual model hallucinations. The daily AI cost limit is a practical safety guard—LLM calls can get expensive fast, and the system stops itself when you hit your budget. The SQLite audit trail means every decision is replayable, which matters when you're debugging why the bot bought something stupid. The category scorer that tracks win rates by market type (sports, economics, politics) is genuinely useful for understanding what the system is actually good at.

Who should use this?

Prediction market traders who want a framework for systematic decision-making rather than gut feelings. Developers building automated trading systems will find the audit trail and ensemble architecture useful for iteration. Researchers interested in whether LLM committees can outperform single-model predictions will get value from the calibration tracking. You should have API credentials for Kalshi and OpenRouter, and you need to be comfortable with Node.js and environment configuration. If you want to plug this in and forget about it, look elsewhere—this requires tuning and monitoring.

Verdict

The architecture is sound and the multi-agent approach is thoughtful, but this project has only 46 stars and that 0.699% credibility score signals it's not battle-tested. The repeated-description red flag suggests hasty setup or abandoned development. Start exclusively in paper mode, set conservative position limits, and expect to spend time tuning prompts and thresholds before trusting it with real capital.