The 5 GitHub Repos Rewriting How AI Trades Money

By Prahlad Menon 3 min read

GitHub is where the future gets built before anyone talks about it. This week, five repos in the AI finance space exploded in stars — and they tell a coherent story about where the industry is heading.

Here’s what they are, why they matter, and what the pattern means.

The Multi-Agent Trading Firm Pattern

1. TradingAgents — Your LLM-Powered Trading Desk

TauricResearch/TradingAgents — +9.3K stars this week

TradingAgents doesn’t just run an LLM against a stock ticker. It replicates the structure of an actual trading firm in software.

You get a Fundamentals Analyst, a Sentiment Analyst, a News Analyst, and a Technical Analyst — each an independent LLM agent focused on its domain. Above them sit Bull and Bear Researchers who debate the analysts’ findings. A Trader synthesizes the debate into a position. A Risk Manager evaluates volatility and liquidity. A Portfolio Manager makes the final call.

The key innovation: these agents argue before anything trades. No single agent’s blind spot kills the portfolio.

Built on LangGraph, it supports GPT-5, Claude, Gemini, Grok, and local models via Ollama. There’s a paper behind it (arXiv: 2412.20138) and an active research community at Tauric Research.

The +9.3K star week is real. They fully open-sourced the framework after significant community interest, which is what caused the spike.

2. ai-hedge-fund — The Legendary Investor Simulator

virattt/ai-hedge-fund — 49.6K total stars

This one has been around longer and it shows. At nearly 50K stars, it’s one of the most-starred finance AI repos on GitHub.

The premise: what if you could have Warren Buffett, Charlie Munger, Michael Burry, Cathie Wood, Stan Druckenmiller, Peter Lynch, and seven other legendary investors as LLM agents — each analyzing the same stock from their own lens?

That’s exactly what it does. Each investor-agent has been carefully prompted to reflect their known methodology — Burry hunts deep value and contrarian bets, Munger only buys wonderful businesses at fair prices, Cathie Wood bets on disruption. A Risk Manager and Portfolio Manager synthesize the debate.

It’s educational by design (no real trades), but the architecture maps cleanly to production systems. It has a full web UI now, not just a CLI.

The conceptual overlap with TradingAgents is real — both use a multi-agent debate structure before any decision. That’s not coincidence. It’s converging architecture.

Autonomous Execution

3. NOFX — The AI That Manages Itself

NoFxAiOS/nofx — 11.2K stars

NOFX is where the multi-agent concept meets actual execution — and removes the human from the loop entirely.

The pitch: Any market. Any model. Zero configuration. The AI selects which model to use, pulls its own market data, and decides when to trade. You define a strategy in the visual Strategy Studio. The AI handles everything else.

Two things stand out technically:

x402 micropayments: Instead of managing API keys across different AI providers, you fund a USDC wallet. Every model request triggers a micropayment. No accounts, no quotas, no key rotation — your wallet is your identity. This is a genuinely novel payment primitive for AI agents.

Safe Mode: If the AI fails at least 3 times consecutively, it auto-protects your positions. This kind of circuit-breaker logic is what separates a demo project from something you’d actually run with real money.

It connects to Binance, Bybit, OKX, Hyperliquid, and more. Built in Go (backend) + React (frontend). One-command install via curl.

The “AI Competition” feature — where multiple AI models compete in real-time on a leaderboard — is a clever way to surface which models actually perform on live data.

Prediction Market Infrastructure

4. prediction-market-analysis — 36GB of Ground Truth

Jon-Becker/prediction-market-analysis — 2.3K stars

Polymarket and Kalshi have become serious financial venues. Polymarket crossed $1B in volume on major events. Kalshi became federally regulated. But there’s been almost no public dataset to study how these markets actually behave at the microstructure level.

This repo fixes that. It’s the largest publicly available dataset of Polymarket and Kalshi trade history — 36GB compressed, covering market metadata, full trade histories pulled from APIs and blockchain, and an extensible framework to run your own analyses.

Two papers have already been published using this dataset:

  • “The Microstructure of Wealth Transfer in Prediction Markets” (Becker, 2026)
  • “Decomposing Crowd Wisdom: Domain-Specific Calibration Dynamics in Prediction Markets” (arXiv: 2602.19520)

For anyone building trading models, calibration research, or market microstructure analysis on prediction markets — this is the starting dataset.

5. pmxt — CCXT for Prediction Markets

pmxt-dev/pmxt — 1.2K stars

If you’ve ever integrated a crypto exchange, you know ccxt — the unified library that abstracts away exchange-specific APIs. pmxt is that, but for prediction markets.

Polymarket, Kalshi, Limitless, Myriad, Opinion — each has its own API, auth flow, data format, and event/market/outcome hierarchy. pmxt wraps all of them behind a single consistent interface.

The API design is clean. You query for an Event (broad topic), drill into a Market (specific question), and trade an Outcome (Yes/No share). Same code works across platforms.

Available in both Python (pip install pmxt) and JavaScript (npm install pmxtjs). It also ships a migration tool (dome-to-pmxt) for codebases that previously used the Dome API.

Active fixes shipping all week is a good sign for a young infrastructure library — it means the maintainers are responsive to real-world usage.

What the Pattern Tells You

Three distinct things are happening simultaneously:

Multi-agent debate is winning. Both TradingAgents and ai-hedge-fund independently converged on the same architecture: specialist agents → structured debate → risk review → final decision. The pattern works because markets are adversarial — having a Bull and Bear argue before you trade is how professional desks actually operate.

Full autonomy is the next frontier. NOFX represents the shift from “AI assists the trader” to “AI is the trader.” The x402 payment primitive is interesting beyond finance — it’s a generalization of how AI agents should handle resource access at runtime.

Prediction markets are going institutional. A 36GB research dataset and a unified trading library in the same week signals that prediction markets are graduating from crypto-curious experiments to serious venues for quantitative work. The academic citations on Jon Becker’s dataset are a leading indicator.

If you’re building in this space — trading systems, multi-agent architectures, or prediction market tooling — these are the repos worth watching. All five are open source and actively maintained.