Trading Agent Harness for Prediction Markets

Coinjure is a trading agent harness for prediction markets. It empowers LLM agents to drive the entire strategy lifecycle purely through CLI commands — autonomously discovering cross-market relations, building executable strategies, running large-scale backtests, and deploying to live execution.

Using Coinjure, LLM agents can discover over 100 backtest-positive strategies in a single hour — a capability validated by deploying to live trading and generating real profit on prediction market exchanges.

Demos

• 1min Introduction of Coinjure
• Basic Functionality Demo of Coinjure
• Claude Code + Coinjure Demo

Installation

pip install coinjure

From source:

git clone https://github.com/ulab-uiuc/coinjure.git
cd coinjure
pip install poetry
poetry install

Quick Start

1) Discover — find markets and detect tradable relations

coinjure market discover -q "election" --exchange both --limit 20

Searches Polymarket and Kalshi, then auto-detects structural relations (exclusivity, complementary, implication) within each event. Discovered relations are persisted to ~/.coinjure/relations.json.

2) Backtest — validate strategies against historical data

coinjure engine backtest --all-relations

Each relation type auto-selects its matching strategy (see table above). Only relations with positive PnL advance to the next stage.

3) Paper trade — simulate live execution

coinjure engine paper-run --all-relations --detach

Deploys all backtest-passed strategies as independent background processes. Use coinjure engine monitor to attach a live TUI dashboard.

4) Live trade — execute with real funds

coinjure engine live-run --exchange polymarket --all-relations --detach

Requires exchange credentials (see Environment Variables).

5) Monitor and control

coinjure engine monitor         # live TUI across all engines
coinjure engine pause  --all    # pause all strategies
coinjure engine resume --all    # resume
coinjure engine killswitch --on # emergency halt

Architecture Overview

The pipeline has four stages:

LLM Agent + market data
    → [1. Discovery]  — discover cross-market relations via CLI
    → [2. Backtest]   — validate each relation against historical data
    → [3. Execution]  — paper-trade or live-trade validated strategies
    → [4. Monitoring]  — human operator monitors and intervenes when needed

Core runtime components:

TradingEngine
  ├── DataSource          (Live / Historical / Hub)
  ├── Strategy            (Relation-based / LLM-powered / Custom)
  └── Trader              (PaperTrader / PolymarketTrader / KalshiTrader)
       ├── RiskManager    (Standard / Conservative / Aggressive)
       └── PositionManager

Each engine instance runs as an independent OS process, allowing hundreds of strategies to execute in parallel without shared-state contention. A ControlServer (Unix socket) provides per-instance pause/resume/status/stop, while the StrategyRegistry persists portfolio state across sessions.

Human-in-the-Loop Model

The operator is intentionally lightweight:

• Use coinjure engine monitor for live TUI visibility across all running engines.
• Use coinjure engine pause|resume|stop for intervention.
• Use coinjure engine killswitch --on for emergency halt.
• Use coinjure engine swap for hot-swapping strategy logic without restarting.

The operator should not need to manually place/cancel orders in normal operation.

CLI Reference

coinjure market — Market discovery and analysis

Command	Description
market info	Fetch detailed info for a market (by ID or slug)
market discover	Multi-keyword search + auto structural relation discovery
market news	Fetch news headlines (Google, RSS, TheNewsAPI)
market relations list	List stored market relations (filter by type/status)
market relations add	Create a relation between 2+ markets
market relations remove	Remove a relation by ID

coinjure engine — Trading engine and portfolio management

Command	Description
engine paper-run	Start paper trading (single strategy or --all-relations batch)
engine live-run	Start live trading with real funds
engine backtest	Backtest relations against historical order book data
engine list	Show all strategies in the portfolio registry
engine status	Show engine status: positions, PnL, decisions, trades
engine pause	Pause decision-making (--all for all instances)
engine resume	Resume after pause
engine stop	Graceful shutdown (--all for all instances)
engine swap	Hot-swap strategy without restarting
engine retire	Stop and mark as retired (--all for all instances)
engine promote	Promote relation(s) from paper to deployed
engine monitor	Attach live TUI dashboard to running engines
engine killswitch	Toggle the global emergency kill-switch

Environment Variables

export POLYMARKET_PRIVATE_KEY="your_private_key"      # Polymarket live trading
export KALSHI_API_KEY_ID="your_kalshi_key_id"         # Kalshi live trading
export KALSHI_PRIVATE_KEY_PATH="/path/key.pem"        # Kalshi live trading

Citation

If you find Coinjure useful in your research or work, please cite:

@software{coinjure2026,
  title   = {Coinjure: A Trading Agent Harness for Prediction Markets},
  author  = {Yu, Haofei and Yang, Yicheng and Liu, Yuxiang and You, Jiaxuan},
  year    = {2026},
  url     = {https://github.com/ulab-uiuc/coinjure},
  note    = {University of Illinois Urbana-Champaign}
}

License

MIT

Disclaimer

This software is for educational and research use. Live trading carries financial risk.