ryanfrigo/kalshi-ai-trading-bot

The honest edge layer for Kalshi prediction markets - an agent-native toolkit (CLI, MCP server, Claude skill) that measures whether you actually beat the book instead of claiming it.

What it solves

This project provides a rigorous toolkit for trading on the Kalshi prediction market. Instead of claiming guaranteed profits, it focuses on providing the tools to honestly measure whether a trading strategy has a statistical "edge" against the market book, preventing users from mistaking luck for skill through strict out-of-sample validation.

How it works

The toolkit integrates a signed Kalshi API client with a decision journal and telemetry system. It uses a "self-improvement loop" where trades are executed, settled, and then measured using metrics like Brier scores and log-loss. This data is used to automatically derive an "Edge Policy" that can block losing categories or apply "haircuts" to overconfident probability estimates to calibrate future trades.

It offers an agent-native surface via a CLI and an MCP (Model Context Protocol) server, allowing LLMs (like Claude) to drive the trading process—from researching catalysts to executing guarded orders—while keeping API keys local.

Who it’s for

Traders and developers interested in prediction markets who want a disciplined, data-driven approach to automated trading and a framework to prove their strategy's performance using real settlement data.

Highlights

  • Honest Edge Measurement: Uses Brier scores and log-loss and refuses to claim edge if the sample size is too small (under 10 forward-settled trades).
  • Self-Improvement Loop: Automatically updates trading policies based on settled P&L and calibration errors.
  • Agent-Native Design: Includes a first-mover MCP server and a dedicated Claude skill for LLM-driven account management.
  • Risk Governance: Built-in daily-loss circuit breakers, drawdown halts, and fractional Kelly sizing.
  • Pluggable LLM Integration: Connects to any model via OpenRouter with a configurable fallback chain.
  • Coded Example Strategies: Ships with AI-driven directional trading and a math-based "Safe Compounder" strategy.

Related

  • Project
    OctagonAI/kalshi-trading-bot-cli

    An AI-powered CLI for Kalshi prediction markets that identifies trading edges using AI probability estimates and executes trades with built-in risk management.

  • Project
    Oft3r/agentic-trading-desk

    A locally‑run personal trading system that lets a Claude (or Gemini) AI agent fetch market data, run deterministic Python indicator scripts, and place risk‑limited market orders on Robinhood. The core is three pure‑Python modules—indicators, macro‑sentiment, and scoring—that compute a three‑pillar score (trend, momentum, macro) and output a concrete trade decision. A pre‑tool shell hook enforces a strict Mandate (market‑only, regular‑hours, $≤1,200 per buy, circuit‑breaker, etc.). The project includes a Claude Code skill, a Gemini‑Spark sanitized copy, and full CLI usage instructions.

  • Project
    chrisworsey55/atlas-gic

    A framework for autonomous AI trading agents that self-improve their prompts through market feedback, autonomous agent spawning, and regime-specific evolutionary training.

  • Project
    openlit/openlit

    An open-source agent harness engineering platform that uses OpenTelemetry to trace, evaluate, and guard AI agents and coding agents.

  • Project
    YichengYang-Ethan/oracle3-prediction-market-agent

    An open-source trading engine and MCP server for prediction markets that identifies and trades probability-based arbitrage opportunities across Kalshi, Polymarket, and Solana.