HKU's Vibe-Trading Hits 11K Stars Turning Plain English Into Live Trades
Vibe-Trading is an open-source multi-agent framework that turns plain-English prompts into backtested trading strategies, now with live broker execution across 10 platforms
Takeaways
- What it is: Open-source multi-agent AI framework from HKUDS that converts plain-English prompts into backtested trading strategies across global markets. GitHub
- Live trading added: 10 broker connectors (IBKR, Robinhood, Alpaca, OKX, Binance, Futu, Tiger, Longbridge, Dhan, Shoonya) with mandate-gating and kill-switch safety guards.
- Scale: 74 finance skills, 29 swarm team presets, 7 backtest engines, 22 MCP tools, 13 LLM providers including free local Ollama.
- Shadow Account: Upload broker export, agent reverse-engineers your trading rules and backtests the "perfect you" to quantify behavioral alpha leakage.
- Free to use: MIT license, pip-installable; market data for HK/US/crypto/A-shares available with zero API keys via yfinance, OKX, and AKShare.
- Caveats: Live trading is experimental; LLM lookahead bias affects backtests; small models fabricate results instead of running actual tools.
Vibe-Trading, an open-source project from the Hong Kong University Data Science Laboratory, has quietly become one of the fastest-growing repositories in the AI-finance space. The pitch is simple: describe a trading idea in plain English, and a coordinated swarm of AI agents writes the strategy code, backtests it across global markets, and can now route live orders to your broker. It crossed 11,000 GitHub stars in under two months, with nearly 1,850 new stars in a single week.
Vibe-Trading is an AI-driven multi-agent financial workspace open-sourced by the Hong Kong University Data Science Laboratory. It transforms natural language instructions into executable trading strategies, research insights, and portfolio analytics, covering global markets including A-shares, Hong Kong/US stocks, cryptocurrencies, futures, and forex.
From prompt to backtest in one command
The core loop is a ReAct agent (short for Reason + Act: a pattern where the model alternates between thinking through a problem and calling external tools) that routes your natural-language request through a library of 74 specialist finance skills. Ask it to backtest a momentum strategy on the S&P 500, and it will select the right skill, pull data from the appropriate source, generate the signal engine code, run it through the backtest engine, and return a full performance report with Sharpe ratio, max drawdown, and benchmark comparison.
The system includes 74 professional financial skills, 29 pre-configured DAG-orchestrated expert agent teams, and 7 major backtesting engines including cross-market portfolio backtesting. It supports one-click export to TradingView, Tongdaxin/THS, and MetaTrader 5.
A quick example of what this looks like in practice:
pip install vibe-trading-ai
vibe-trading init # interactive .env setup
# Run a backtest from the command line
vibe-trading run -p "Backtest a 20/50-day moving average crossover on AAPL for the past year, show Sharpe ratio and max drawdown"
# Or kick off a multi-agent swarm debate
vibe-trading --swarm-run investment_committee '{"topic": "Is TSLA a buy at current levels?"}'The swarm is the product
The most distinctive feature is the Swarm system: pre-built DAG (directed acyclic graph) workflows where multiple specialized agents collaborate in sequence. A DAG here means agents are wired in a dependency graph, so a risk-review agent only runs after a bull/bear debate finishes. The result is a structured, reproducible research pipeline rather than a single LLM call.