An X Article by @carm1nee, "The 10 Most Dangerous GitHub Repos in Trading (Full Course)," lays out 10 free GitHub projects that together form an "AI hedge fund": a research desk, an investment committee and an execution desk. Read the original X Article. Credit to @carm1nee for a useful map.
What "tested" means here: we read each repo, its license, its recent commit activity, and for the two committee tools we read the code. We did not live-trade any of them and we did not run backtests. This is a review, not a performance report.
The list (stars and dates from GitHub, Oct 26, 2026)
Research desk
- OpenBB - data platform. Apache-2.0 per its license file. Active. Its old GitHub address now redirects to a different, recently created org that the OpenBB website does not link to, so start from the official site and its OpenBB-finance GitHub org.
- Ccxt - one API for 100+ crypto exchanges. MIT. Very active. It can place orders, so treat it as execution code.
- Kronos - a forecasting model trained on price candles. MIT. Last commit April 2026, so it looks quiet.
- Microsoft Qlib + RD-Agent - quant research platform plus an agent that writes and runs its own code. MIT. Active. Run RD-Agent only on a machine with no real credentials.
- Daily_stock_analysis - a daily AI stock dashboard sent to Telegram. MIT. Very active.
Investment committee
- TradingAgents - analysts, a bull vs bear debate, a trader, a risk team and a portfolio manager, all LLM agents. Apache-2.0. About 115k stars. Active.
- Ai-hedge-fund - investor-style agents vote on a ticker, with backtest and paper modules. MIT. Active. In the code we read, the broker layer is paper and simulation only.
Execution desk
- Freqtrade - crypto bot. GPL-3.0, which matters if you plan to build on it. Active.
- Hummingbot - crypto market making. Apache-2.0. Active.
- NautilusTrader - event-driven trading engine. LGPL-3.0. Active.
What we think
The execution desk is the dangerous part, and the least useful to a paper desk like ours. The interesting piece is the committee: forcing a bear case and a risk review before a call is made is a good habit, with or without AI.
Two caveats the author also flags: LLM agents can give different answers on the same ticker and date, so judge them over many calls, not one. And grading them on past dates can mix forecasting with memory, because the models may have seen that period in training.
What we are doing next
We plan a paper-only shadow test of the two committee tools next to our existing paper desks, with rules written down before the test starts and a fixed end date. We will report the results either way.
More from this series: Market Calls, Investigated.
Disclaimer: This post is for information and education only. It is not investment advice, and nothing here is a recommendation to buy or sell any security or to use any software for trading. All of our desks are paper portfolios. Trading involves risk of loss. Do your own research.
