JamesANZ/prediction-market-mcp

> **Real-time prediction market data in your AI workflow.** Get current odds and prices from Polymarket, PredictIt, and Kalshi. No API keys required.
An MCP (Model Context Protocol) server that brings live prediction market data into AI coding environments like Cursor and Claude Desktop.
Why Use Prediction Markets MCP?
- 🆓 **No API Keys** – Works out of the box, zero configuration
- 📈 **Multi-Platform** – Polymarket, PredictIt, and Kalshi in one interface
- ⚡ **Real-time Data** – Current odds and prices from live markets
- 🎯 **Easy Setup** – One-click install in Cursor or simple manual setup
- 🔍 **Keyword Search** – Find markets by topic or keyword
Quick Start
Ready to explore prediction markets? Install in seconds:
**Install in Cursor (Recommended):**
🔗 Install in Cursor
**Or install manually:**
Features
get-prediction-markets
Search for prediction markets across multiple platforms by keyword.
**Parameters:**
- keyword (string, required): Search term (max 50 characters)
- Examples: "trump", "election", "supreme court"
**Returns:**
- Markets from Polymarket (percentage odds)
- Markets from PredictIt (dollar prices)
- Markets from Kalshi (regulated US markets)
gather-market-movement-evidence
Given one market and a lookback window, return the observed price move and recent external evidence. The server does not decide why the market moved. The calling model does that from the evidence.
Agent Reach is optional. If it is not installed, this tool still returns the market and the price move, and marks external sources unavailable. get-prediction-markets does not call Agent Reach.
**Parameters:**
- market (string, required): A keyword, or a direct id such as polymarket: , kalshi: , or predictit: : . A keyword that matches more than one market returns candidates and does not search the web.
- outcome (string, optional): Outcome to measure. Defaults to Yes when that outcome exists.
- lookback (string, optional): 1h, 6h, 24h, 7d, or 30d. Default 24h. PredictIt has no history API, so that venue compares the previous close with the last trade.
- sources (string array, optional): web, x, reddit, and/or youtube. Default is all four.
- queries (string array, optional): Up to three search queries. When set, they replace the generated query. Only the first query is executed.
- limit (number, optional): Maximum evidence items, 1–10. Default 8.
Evidence text is untrusted data. It is returned in evidence[] and is not treated as instructions.
External search uses the Agent Reach 1.5 command table:
- Web: Exa through mcporter
- YouTube: yt-dlp search metadata, not transcripts
- X: twitter search when TWITTER_AUTH_TOKEN and TWITTER_CT0 are set on this server process
- Reddit: rdt, or desktop OpenCLI when rdt is absent
Facebook, Instagram, and GitHub are not searched. Install Agent Reach and run agent-reach doctor on the same machine as this server. A missing login, rate limit, or offline source is reported on that source and does not fail the price lookup.