A multi-agent investing simulator, for learning
ai-hedge-fund is a teaching project that wires up a team of LLM agents to analyze stocks and produce buy and sell signals. It models a roster of investor personas after famous investors, layers them over agents that handle valuation, sentiment, fundamentals, and technicals, and runs the output through a risk manager and a portfolio manager that produces a final decision. The appeal is watching multi-agent reasoning play out on real tickers. At 60,129 stars as of 2026-06, it became one of the most-starred AI-finance projects on GitHub.
The most important thing to understand is what it is not: it does not place trades, and it is explicitly built for education and research, not investing.
What it does
You give it tickers, and the agents each form a view from their angle, then the portfolio manager combines those into a position decision. It supports several LLM providers (OpenAI, Anthropic, Groq, DeepSeek, and local models via Ollama) and pulls market data from a financial data API. It can also backtest a strategy over a date range. The value is pedagogical: it is a readable, hackable example of orchestrating many LLM agents toward a structured decision.
Install
It uses Poetry and needs at least one LLM API key plus a financial data API key:
poetry install
poetry run python src/main.py --ticker AAPL,MSFT,NVDAFor a no-API-cost run you can use local models with the Ollama flag, and there is a backtester entry point for historical runs. Expect to pay for LLM tokens and data access if you use hosted providers, and note that a full backtest over many tickers and dates can run up real cost.
Read this before you connect real money
This is the part that matters more than any feature. The README is explicit that the project is for educational and research purposes only, is not intended for real trading or investment, provides no investment advice or guarantees, and that the creator assumes no liability for financial losses. The system does not actually make trades. Past performance in a backtest does not indicate future results, and the backtester has had known bugs affecting accuracy. If you wire the signals to a broker yourself, every consequence is on you. Treat this as a way to learn multi-agent LLM design, not as a source of investment decisions.
Where it fits, and where it doesn’t
Reach for ai-hedge-fund when you want to study how a multi-agent LLM system can be structured around a decision, when you want a readable codebase to fork, or when you want to experiment with investor-style prompting. As a learning artifact it is genuinely good.
It is the wrong thing to trust with money. There is no real trade execution, the data source has coverage gaps, agent signals are combined simply without a robust consensus mechanism, and backtest results have been unreliable. For a more complete agent trading framework or a serious research platform, the alternatives below fit better.
ai-hedge-fund versus the other finance projects
| ai-hedge-fund | TradingAgents | qlib | FinGPT | |
|---|---|---|---|---|
| Stars | 60,129 | 85,821 | 44,469 | 20,502 |
| Type | educational agent simulator | multi-agent trading framework | quant research platform | financial LLM resources |
| Real trading | no, simulation only | execution-capable framework | backtest and execution framework | no, models only |
| License | MIT | Apache-2.0 | MIT | MIT |
| Best for | learning multi-agent design | building agent strategies | quant ML research | financial NLP and fine-tuning |
Counts are from GitHub as of June 2026. TradingAgents is the larger multi-agent trading framework with execution paths. qlib is Microsoft’s research-grade quant platform. FinGPT provides financial language models rather than a trading system. ai-hedge-fund’s niche is being the most approachable teaching example of the four.
The star curve
The star curve spiked hard, which says more about the appeal of the concept (LLM agents playing famous investors) than about production readiness. It is a viral teaching repo, and the gap between its star count and its actual trading capability is the thing to keep in mind.
What the issue tracker warns you about
The tracker maps the limits clearly. There is no real broker execution, so signals stop at a recommendation. The single financial data source has coverage gaps, especially for ETFs and some markets. Backtesting has had bugs that undercut the reliability of historical results. Agent signals are aggregated without a real consensus or weighting mechanism, so conflicting views are not resolved well. And LLM calls without timeouts can hang long analyses. None of this is fatal for a learning tool, but all of it argues against trusting the output as advice.
Related
For a fuller multi-agent trading framework, see TradingAgents. For a quant research platform, see qlib. For financial language models, see FinGPT. For the agent frameworks underneath, see LangChain-style tooling. For what is trending, see the daily digest and the weekly report.
FAQ
Can I use ai-hedge-fund for real trading? No. The README states it is for educational and research purposes only, does not place trades, provides no investment advice, and the creator assumes no liability for losses. If you connect it to a broker yourself, all risk is yours.
ai-hedge-fund vs TradingAgents, which should I use? ai-hedge-fund is the more approachable teaching example, built around investor personas, with no real execution. TradingAgents is a larger framework with execution paths. Choose ai-hedge-fund to learn multi-agent design, TradingAgents to build something closer to a real strategy.
How much does it cost to run? You pay for LLM tokens from your provider and for the financial data API. A single analysis is cheap, but a full backtest over many tickers and dates stacks up cost. Running local models via Ollama avoids token cost at the expense of quality.
Can ai-hedge-fund use local models? Yes, through Ollama, which removes LLM API cost. Web UI support for local models has been incomplete, but the CLI path works.
Are the agents trained on financial data? No. They are zero-shot LLMs driven by system prompts that encode each persona and analysis style. There is no custom model training; the behavior comes from prompting, not fine-tuning.