AI Hedge Fund on GitHub: 9 Finance Repos

Table of Contents

1. September 2026 update: what changed in six months
2. The original five, re-ranked by stars (GitHub API, 26 Sep 2026)
3. Four new repos worth knowing (created in 2026)
4. What the new entries have in common
5. 1. TradingAgents: Multi-Agent LLM Trading Firm in a Box
6. Architecture Highlights
7. Practical Application
8. 2. ai-hedge-fund: The 49K-Star Giant
9. Architecture Highlights
10. Practical Application
11. 3. NoFx: The AI Trading Assistant with a Kill Switch
12. Architecture Highlights
13. Practical Application
14. 4. prediction-market-analysis: 36GB of Prediction Market Truth
15. Architecture Highlights
16. Practical Application
17. 5. pmxt: CCXT for Prediction Markets
18. Architecture Highlights
19. Practical Application
20. The Common Thread
21. If You Work in This Space

Updated 26 September 2026. The original five projects are below with their March star counts; this section re-checks every number against the GitHub API and adds four repos that did not exist, or were too small to notice, when this was first written.

The most interesting AI finance work is not happening behind closed doors at hedge funds or inside Bloomberg terminals. It is happening in public, on GitHub, where open-source developers are building systems that would have required a 50-person quant team five years ago. One data point: TradingAgents, the first project on this list, went from 44,674 stars at the end of March 2026 to 108,676 on 26 September 2026.

Disclosure: Ultra Lab, MindThread and Ultra Advisor are products of the same group, Ultra Creation (傲創實業).

#September 2026 update: what changed in six months

#The original five, re-ranked by stars (GitHub API, 26 Sep 2026)

Project	March 2026	September 2026	Last push
TauricResearch/TradingAgents	44.7K	108,676	2026-09-25
virattt/ai-hedge-fund	49.6K	63,752	2026-09-25
NoFxAiOS/nofx	11.2K	12,958	2026-09-05
Jon-Becker/prediction-market-analysis	2.3K	3,858	2026-09-21
pmxt-dev/pmxt	1.2K	2,169	2026-07-18

TradingAgents is the story: it more than doubled in six months and overtook ai-hedge-fund as the most-starred repo on this list. Everything else grew at a normal pace. Note the last-push dates: TradingAgents (v0.5.0 shipped in September), ai-hedge-fund and prediction-market-analysis all pushed within the past week, while pmxt has not pushed since 18 July, which matters if you plan to build on it.

A correction: earlier versions of this table put TradingAgents at about 9.3K stars in March, which overstated its growth. An archived copy of the repo page from 31 March 2026 shows 44,674.

#Four new repos worth knowing (created in 2026)

6. HKUDS/Vibe-Trading · 34,055 stars · MIT · Python · created 1 April 2026

"Your personal trading agent." From the University of Hong Kong's data science lab (the same group behind LightRAG). One pip install vibe-trading-ai gives an agent harness with bundled trading skills, a backtest engine, portfolio optimizer, market data fallbacks across US, HK, China A-shares, Korea and crypto, an MCP server, and a desktop shell. The changelog in the README is updated almost daily, and unusually honest: entries like "backtest reports now show the book that actually filled" and "the test suite stopped writing into your real config root" tell you exactly what was broken before. It is the most-starred repo on this list created in 2026, and it added about 2,500 stars between 24 August and 26 September. One caution straight from their README: there is a fake X account and a fake token using the project's name; the project has never launched a token.

7. TraderAlice/OpenAlice · 7,162 stars · AGPL-3.0 · TypeScript · created 18 February 2026

"Your one-person Wall Street." The idea is different from every other entry: instead of building a trading agent, OpenAlice gives coding agents (Claude Code, Codex, opencode, Pi) a trading-shaped workspace: per-task git repos, markdown issues, a memory graph of tickers and theses, an inbox for reports, and "trading as git" where account actions are staged, committed, reviewed and pushed through an approval gate rather than fired directly. Runs locally, state is plain files under ~/.openalice. The README calls the trading layer "especially beta" and says not to use it for live trading with money you cannot lose. Best read as a research workspace first.

8. questflowai/investorskills · 1,862 stars · MIT · created 27 May 2026

Not code, judgment. A library of SKILL.md packages that encode how named investors (Buffett, Livermore and others) filter opportunities, size positions and define what would invalidate a thesis, in a YAML plus markdown schema any agent harness can read. Works in Claude Code, Codex, Cursor and friends. The open repo is the free layer for the Questflow product; the closed "exclusive" skills are not published. Useful if you want your agent to reason like a specific school of investing instead of generic "analyze this stock".

9. mnemox-ai/tradememory-protocol · 1,422 stars · MIT · Python · created 23 February 2026

A memory layer for trading agents, shipped as an MCP server (pip install tradememory-protocol, then claude mcp add). Before a trade the agent recalls what happened last time in similar conditions; after the trade one call records outcome, reasoning and emotional state into five memory layers; every decision is SHA-256 hashed for a tamper-evident audit trail, pitched at MiFID II style decision documentation. It does not execute trades. The README says the project is feature-complete and in maintenance mode as of August 2026, so take it as a finished tool rather than a growing platform.

#What the new entries have in common

All four are agent infrastructure, not trading strategies. Six months ago the interesting repos were "an AI that trades"; now they are the harness, the workspace, the judgment library and the memory layer that any agent plugs into. That matches what happened to coding agents a year earlier, and it is the same lesson: the strategy is the easy part to copy, the scaffolding around it is where the work is.

---

The original article follows, with the March 2026 star counts it was written with.

---

#1. TradingAgents: Multi-Agent LLM Trading Firm in a Box

Repo: TauricResearch/TradingAgents | 44.7K stars (31 March 2026 snapshot; first published here as 9.3K)

Corrected 27 September 2026. The March version of this section listed a fund manager agent and left out the researcher and trader agents. The README at the time described an analyst team, bull and bear researchers who debate the analysts' reports, a trader, a risk management team, and a portfolio manager who approves or rejects each trade, so the list below follows it.

One-liner: A multi-agent LLM framework that simulates an entire trading firm's decision-making process (analysts, researchers, a trader, risk managers and a portfolio manager), all as autonomous AI agents.

#Architecture Highlights

TradingAgents doesn't just ask an LLM "should I buy AAPL?" It models the organizational structure of a real trading desk. The system runs multiple specialized agents in parallel:

• Fundamental Analyst Agent: parses SEC filings, earnings transcripts, balance sheets
• Technical Analyst Agent: reads chart patterns, moving averages, volume signals
• Sentiment Agent: monitors news feeds, social media, analyst upgrades/downgrades
• Bull and Bear Researchers: debate the analysts' reports, weighing potential gains against risks
• Trader Agent: turns the analyst and researcher reports into a trade, deciding its timing and size
• Risk Manager Agent: evaluates position sizing, correlation risk, drawdown limits
• Portfolio Manager Agent: approves or rejects the trade; approved orders go to a simulated exchange

Each agent has its own system prompt, tool access, and memory. They debate. The portfolio manager agent receives conflicting recommendations and must weigh them, just like a real PM sitting in a morning meeting.

The framework is LLM-agnostic (GPT-4, Claude, Gemini, local models via Ollama) and uses LangGraph for agent orchestration. Backtesting is built in.

#Practical Application

This isn't production-ready for real capital (and they say so clearly in the README). But it's an exceptional research tool. If you're studying how multi-agent architectures handle conflicting signals under uncertainty, this is the best open-source implementation available. Quant researchers can fork it, plug in their own alpha signals, and test whether agent debate actually improves signal quality versus a single-model approach.

---

#2. ai-hedge-fund: The 49K-Star Giant

Repo: virattt/ai-hedge-fund | 49.6K stars

Corrected 26 September 2026. The March version of this section described bull and bear agents, a portfolio manager that re-weights agents by recent accuracy, and Yahoo Finance, Alpha Vantage and Polygon.io as data sources. The project's README at the time described none of that, so the architecture below is rewritten from it.

One-liner: An AI hedge fund proof of concept where agents modelled on famous investors analyze stocks and a portfolio manager agent makes the final trading decision, on paper only.

#Architecture Highlights

In March this was the most-starred repo on this list, narrowly ahead of TradingAgents, and the architecture explains the appeal. The README at the time listed eighteen agents working together:

• Twelve investor personas: Warren Buffett, Charlie Munger, Ben Graham, Peter Lynch, Phil Fisher, Michael Burry, Cathie Wood, Bill Ackman, Stanley Druckenmiller, Aswath Damodaran, Mohnish Pabrai and Rakesh Jhunjhunwala, each analyzing a stock through that investor's philosophy
• Four analysis agents: valuation, sentiment, fundamentals and technicals, each generating its own trading signal
• Risk Manager: calculates risk metrics and sets position limits
• Portfolio Manager: makes the final trading decisions and generates orders, which the system does not actually send anywhere

Built on Python with LangChain, it pulls prices, fundamentals and earnings from the Financial Datasets API, so you need that key plus one model provider key. The September 2026 version keeps the idea but repackages it as a terminal app where you build a fund (stocks, strategies, rebalance cadence) and backtest it. The codebase is clean and well-documented, which partly explains the star count: it's genuinely accessible to intermediate Python developers.

#Practical Application

Two real use cases we've seen in the wild: (1) Financial educators using it to teach portfolio management concepts; the agent debates make abstract concepts concrete. (2) Solo traders building personal "investment committees": they run the system before every trade as a structured second opinion. Nobody should be auto-executing trades from this, but as a decision-support tool, it's surprisingly useful.

---

#3. NoFx: The AI Trading Assistant with a Kill Switch

Repo: NoFxAiOS/nofx | 11.2K stars

Corrected 26 September 2026. The March version of this section described an accuracy-triggered safety mode (3 consecutive misses), CCXT exchange support and an indefinite paper-trading mode. NoFx's README did not describe any of that in March and does not now, so this section is rewritten from the current README.

One-liner: An open-source trading terminal where a language model is the strategy and a Go runtime enforces the risk limits, outside the model's reach.

#Architecture Highlights

What makes NoFx stand out isn't the AI; it's the risk engineering. Each trader runs a loop (read the market, decide, execute, record the reasoning), and every order passes through limits written in code that the model cannot change:

• Position limits: a cap on concurrent positions, notional sized as a ratio of equity, one position per symbol
• Leverage clamps: hard caps applied when the order is sized, whatever leverage the model asks for
• Exchange-side stops: stop-loss and take-profit placed on the exchange right after every entry
• Drawdown auto-close: profitable positions that give back too much from their peak are closed
• Trade throttling: minimum hold times, re-entry cooldowns per symbol, and entry limits per cycle and per hour
• Safe mode: repeated model failures block new entries until the model recovers

It is written in Go, connects to nine exchanges (Binance, Bybit, OKX, Hyperliquid, Bitget, KuCoin, Gate, Aster and Lighter), and stores every decision with the model's full reasoning. It is built for live trading: the guided Autopilot launch walks through four steps (fund, connect, deposit, start), and a preflight check verifies model access, wallet funds, strategy and exchange balances before any trader may start. The README does not describe a paper-trading mode.

#Practical Application

The design choice worth studying is where the safety lives. In its README's words, "the model proposes, the runtime disposes": the limits sit in code the model cannot reach, not in the prompt. Even if you never use NoFx itself, that split is worth copying for any automated decision system. If you do run it, remember that it places real orders with real money; the README calls AI-driven strategies experimental, says they can lose money, and tells you never to trade funds you cannot afford to lose.

---

#4. prediction-market-analysis: 36GB of Prediction Market Truth

Repo: Jon-Becker/prediction-market-analysis | 2.3K stars

Corrected 27 September 2026. The March version of this section listed complete historical order books and analysis notebooks. The README and data schemas at the time describe market metadata and trade-by-trade history, and the analyses are Python scripts rather than notebooks, so those two lines are corrected below.

One-liner: The largest public prediction market dataset ever compiled: 36GB of historical data from Polymarket and Kalshi, cleaned and ready for analysis.

#Architecture Highlights

This isn't a trading system. It's a dataset, and it fills a gap that researchers have been complaining about for years. The repo contains:

• Trade-by-trade history from Polymarket and Kalshi
• Resolution data: what actually happened vs. what the market predicted
• Price time series at minute-level granularity for major markets
• Market metadata: categories, descriptions, resolution criteria, liquidity depth
• Pre-built analysis scripts showing calibration curves, Brier scores, and market efficiency tests

The data pipeline is documented end-to-end: scraping, cleaning, deduplication, normalization. Storage is in Parquet format (columnar, compressed) with a DuckDB interface for fast local querying. You can run complex analytical queries on the full 36GB dataset on a laptop without spinning up a database server.

#Practical Application

Three immediate uses: (1) Calibration research: how accurate are prediction markets, really? The data shows Polymarket is well-calibrated on high-liquidity markets (events priced at 70% happen roughly 70% of the time) but significantly miscalibrated on thin markets. (2) Feature engineering for trading models: prediction market prices are a leading indicator for traditional assets. Election markets move before polls. Crypto event markets move before spot. (3) Building your own prediction market analytics tool: the dataset is the hard part, and it's done for you.

---

#5. pmxt: CCXT for Prediction Markets

Repo: pmxt-dev/pmxt | 1.2K stars

Corrected 27 September 2026. The March version of this section listed webhook notifications for market resolution. pmxt's README and feature table at the time list no webhooks; the real-time feature is streaming order books and trades, which the feature list below now says.

One-liner: A unified API for prediction markets: trade on Polymarket, Kalshi, Limitless, and Myriad through a single interface, just like CCXT unified crypto exchanges.

#Architecture Highlights

If you've used CCXT (the universal crypto exchange connector), you know the value proposition instantly. pmxt does the same thing for prediction markets:

// Same code, any platform
const market = pmxt.exchange('polymarket')
const positions = await market.getPositions()
const order = await market.createOrder('US_ELECTION_2028', 'buy', 'yes', 100)

The abstraction layer handles the gnarly differences between platforms: Polymarket runs on Polygon (blockchain-based, requires wallet signing), Kalshi is a CFTC-regulated exchange (traditional API auth), Limitless uses a different order book model entirely. pmxt normalizes all of this into a consistent interface.

Key features: unified order types, standardized market discovery (search across all platforms simultaneously), portfolio aggregation across platforms, and real-time order book and trade streams.

#Practical Application

Prediction markets are fragmented. The same event might be listed on Polymarket at 62% and Kalshi at 58%. pmxt makes cross-platform arbitrage trivially easy to implement. Beyond arbitrage, the unified API is essential for anyone building prediction market analytics dashboards, aggregators, or research tools. Writing platform-specific code for four different exchanges is a maintenance nightmare; pmxt eliminates it.

---

#The Common Thread

All five projects share a pattern: they democratize capabilities that were previously locked behind institutional walls. Multi-agent trading systems, prediction market data pipelines, unified exchange APIs: these used to require dedicated engineering teams and six-figure data budgets. Now they're a git clone away.

The risk is obvious. Easier access to sophisticated tools doesn't make markets easier to beat. These systems reduce the barrier to entry, which means any edge they provide gets arbitraged away faster. The real value isn't in running them out-of-the-box; it's in understanding the architectures, adapting the patterns, and combining them with domain expertise that can't be cloned from a repo.

That last part, domain expertise, is where the human advantage still holds.

---

#If You Work in This Space

Two pointers from our own shop, in case they fit. If you advise clients on money for a living, Ultra Advisor is our financial advisory SaaS with 18 visualization tools that turn portfolios, policies, and retirement numbers into pictures clients actually follow. And if you are building a finance product yourself and want it to show up when people ask Google or ChatGPT about your niche, UltraGrowth, our AI visibility software on a subscription, automates the SEO and AI-visibility content work, with every article published on your own site rather than rented space.

---

#AIFinance #OpenSource #GitHub #UltraAdvisor

FAQ

What is the AI hedge fund repo on GitHub?

It is virattt/ai-hedge-fund, an MIT-licensed Python proof of concept that uses AI investor-persona agents to make trading decisions for a fund you configure (stocks, strategies, rebalance cadence), with a built-in backtest. It had 63,752 stars on 26 September 2026. It installs with pipx install aihf and needs a Financial Datasets API key plus one model provider key.

Can ai-hedge-fund trade with real money?

No. Its README says it is for educational and research purposes only and that the system does not actually make any trades. TradingAgents is also positioned as a research framework, not financial or trading advice. If you want a repo that places real orders, NoFx does: it connects to nine exchanges, and its execution layer, written in Go, enforces position, leverage and stop-loss limits in code the model cannot change. Its README still warns that AI-driven strategies are experimental and can lose money.

Which of these AI finance repos has the most GitHub stars?

TradingAgents, with 108,676 stars on 26 September 2026, followed by ai-hedge-fund (63,752) and Vibe-Trading (34,055). All counts come from the GitHub API on that date.