By Keel Research Team
Updated October 18, 2026Open-source AI trading agents
Open-source AI trading agents: AI Hedge Fund, TradingAgents
Open-source AI trading agents are free projects on GitHub in which several language-model agents, such as analysts, researchers and a risk manager, study a market and reach a trading decision together. The best known, virattt's AI Hedge Fund and TauricResearch's TradingAgents, analyse stocks and produce decisions, and their READMEs describe them as research projects. Keel covers the next step from your own assistant: turning an idea into a rule, backtesting it on Hyperliquid data and paper trading it.
The multi-agent pattern
- 01Analysts read prices, news and filings
- 02Researchers argue the bull and bear case
- 03Trader weighs the case and proposes a trade
- 04Risk manager approves, resizes or vetoes
The projects, as of October 2026
Open-source AI trading agents compared
What each project is, what it needs to run and whether it places trades, taken from each project's own README. They change often, so check the repository before you build on one.
| Project | What it is | Language, data and keys | Does it trade? | Licence |
|---|---|---|---|---|
| AI Hedge Fund (virattt) | A proof of concept for exploring AI trading decisions. Agents modelled on well-known investors analyse stocks, and a portfolio manager agent makes the call. | Python. Needs a Financial Datasets API key for prices and fundamentals, plus a key for a language model provider. | No. Its README says the system does not actually make any trades and is not intended for real trading. It includes a backtester. | MIT |
| TradingAgents (TauricResearch) | A multi-agent framework where analyst, researcher, trader and risk manager agents debate before a decision. It comes with a research paper. | Python. One language model provider key; market data from sources such as Yahoo Finance and Alpha Vantage. | It produces buy, sell or hold decisions and can backtest them. Its README says it is designed for research and is not trading advice. | Apache 2.0 |
| FinRobot (AI4Finance) | An agent platform for financial analysis: equity research, valuation and reports, built from multi-agent workflows. | Python. Financial data APIs plus a language model key. | Analysis and reports. Its README says its outputs are not recommendations for live trading. | Apache 2.0 |
| FinGPT (AI4Finance) | Open-source financial language models and training datasets, for tasks such as reading financial text. A building block, not an agent. | Python. | No. Its README says nothing in it is a recommendation to trade real money. | MIT |
| Vibe-Trading (HKUDS) | A personal trading agent you run yourself, with research, backtesting and broker connectors. | Python. Language model keys, with market data from several providers. | Yes, through connectors: it backtests, simulates orders locally for paper trading, and can trade live with each mandate bound to one account. | MIT |
How they work
How multi-agent trading systems work
Most of these projects copy the structure of a trading desk. Several agents, each a language model with its own instructions, take one job apiece: one reads prices and indicators, one reads news or filings, one argues the bull case and another the bear case. A trader agent weighs what they found, and a risk agent can veto or resize the trade.
The appeal is that each step is written down. You can read why the system leaned long, which you can't with a black box. The cost is that every decision means many model calls, and the same question can get a different answer on a different day, because the agents reason afresh each time rather than following a fixed rule.
That second point matters most for trading. A decision you can't reproduce is hard to backtest honestly. The projects that backtest do it by replaying the agents over history, which is slow and can give a different answer on each run. A rule distilled from the agents' reasoning can be tested quickly and repeatably.
Agent frameworks
Frameworks for building your own trading agent
If you want to build an agent rather than use one, general agent frameworks do the plumbing. LangChain provides the building blocks for models, tools and agent loops, and LangGraph, from the same team, orchestrates long-running agents that keep state. CrewAI is an open-source framework for teams of agents working together. Microsoft's AutoGen was a popular choice, but its repository is now archived and points new projects to Microsoft Agent Framework. None of them is trading-specific: the market data, the cost model and the backtest are yours to add.
Where Keel fits
From an agent's idea to a tested rule
Most of these projects are built for research. You run them on your own machine, bring your own data keys and model keys, and most of them stop at a decision or an analysis. That is a good way to explore ideas. Turning one into something you would trust with money takes three more steps: a fixed rule, a backtest on real market history with costs, and a stretch of paper trading.
Keel does those steps without a server of your own. Connect Claude, ChatGPT or any assistant that can add a remote MCP server, describe the idea an agent came up with, and Keel builds it as a rule from blocks you can read, backtests it on Hyperliquid history with fees, slippage and funding charged, and gives you a link to the result. You paper trade it in the Keel app on live prices before anything runs with real money.
Using them with Keel
Using an open-source agent with Keel over MCP
Hand off the idea
Run the open-source agent as you normally would, then give its reasoning to an assistant connected to Keel and ask for it as a rule that can be backtested.
Call Keel from your agent
If your agent framework can call MCP servers, point it at https://mcp.usekeel.io/mcp. Your agent can then search Keel's blocks, compose a strategy and run a backtest as one of its tools.
Keep the agent as a reviewer
Agents are good at critique. Paste a Keel backtest into one and ask what could be misleading about it, then test the doubts it raises.
FAQ
Questions about open-source AI trading agents
What is the AI hedge fund on GitHub?
It is virattt's AI Hedge Fund, an open-source Python proof of concept. Agents modelled on well-known investors analyse stocks and a portfolio manager agent decides. Its README says it does not make real trades and is for education and research.
Is TradingAgents free?
The code is free and open source under the Apache 2.0 licence. It calls a language model through your own provider key, and your provider may charge for that usage.
Can these agents trade real money?
Most of the well-known ones don't. AI Hedge Fund and TradingAgents produce decisions for research, and FinRobot produces analysis. Some newer projects, such as Vibe-Trading, connect to brokers for paper and live trading. Read each README before connecting an account.
Do I need an API key?
Usually more than one. Most need a key for a language model provider, and some need a market data key, such as AI Hedge Fund's Financial Datasets key. Keel needs no key: you sign in to Keel when your assistant connects.
How do I backtest what an agent proposes?
Turn the proposal into a fixed rule first, since an agent's reasoning changes from run to run. Then connect an assistant to Keel and ask it to build and backtest the rule on Hyperliquid history with fees, slippage and funding charged. The Claude Code guide shows the full loop from a terminal.
Sources
Facts about each project come from its own repository or documentation, read on the date shown. Open-source projects change quickly, so check the source before you rely on a detail. We left out anything we could not confirm there.
- AI Hedge Fund (virattt): https://github.com/virattt/ai-hedge-fund (read 18 October 2026)
- TradingAgents (TauricResearch): https://github.com/TauricResearch/TradingAgents (read 18 October 2026)
- FinRobot (AI4Finance): https://github.com/AI4Finance-Foundation/FinRobot (read 18 October 2026)
- FinGPT (AI4Finance): https://github.com/AI4Finance-Foundation/FinGPT (read 18 October 2026)
- Vibe-Trading (HKUDS): https://github.com/HKUDS/Vibe-Trading (read 18 October 2026)
- TradingAgents paper: https://arxiv.org/abs/2412.20138 (read 18 October 2026)
- LangGraph and LangChain: https://docs.langchain.com/oss/python/langgraph/overview (read 18 October 2026)
- CrewAI: https://docs.crewai.com/en/introduction (read 18 October 2026)
- AutoGen: https://github.com/microsoft/autogen (read 18 October 2026)
Test the idea, not just the reasoning
Connect your assistant to Keel, turn an agent's idea into a rule and backtest it on Hyperliquid history.