TradingAgents: The Multi-Agent AI Trading Research Framework Everyone Is Starring

Quick answer

TradingAgents is a free, open-source framework that stages a full trading firm — analysts, debaters, a trader and a risk team — on top of large language models. It has passed 117,000 GitHub stars, but it is a research tool: approved orders go to a simulated exchange, and it places no real trades. The software costs nothing under its Apache 2.0 licence; the real bill is the language-model and data usage behind each run.

What TradingAgents actually is

TradingAgents started as an academic paper. Four researchers — Yijia Xiao, Edward Sun, Di Luo and Wei Wang — published it on arXiv on December 28, 2024, and open-sourced the code the same day. The idea was simple enough to explain and hard to build: instead of asking one language model to pick stocks, build the org chart of a trading firm out of language models and let the roles argue.

That framing is the whole product. The framework, built on LangGraph, splits a trading decision into specialised jobs, runs them in parallel, and forces the results through debate before anyone is allowed to decide. It is now one of the most-starred AI projects on GitHub: 117,328 stars and 22,472 forks at the time of writing, up from about 109,700 stars on October 4, according to a roundup that logged the counts that day. GitTrend's October 2026 trending list puts it in the top 100 repositories for the month, with roughly 6,800 new stars in October alone.

The repository is actively maintained. Version 0.7.1 landed in October 2026, the fifteenth release, and the project logs 515 commits. An independent review updated October 25 counted 43 open issues against 486 closed, and 80 open pull requests against 908 closed — a project with a real maintenance pulse, not a paper that was dumped on GitHub and abandoned.

How the firm works

A run starts with a ticker and a date. You might ask for NVDA on September 1, 2026, and the framework does the rest. Four analyst agents work at the same time, each with its own tools:

• Fundamentals Analyst — reads company financials and performance metrics, looking for intrinsic value and red flags. For US companies the statements come from SEC EDGAR, served as they were filed on the run date, so a backtest cannot peek at a restatement that came later.
• Sentiment Analyst — aggregates news headlines, StockTwits and Reddit chatter into a single read on short-term mood. With a TypeSafe key set, posts are screened by Jev first so off-topic posts are dropped before they are read.
• News Analyst — follows global news and macroeconomic indicators, including FRED data, and interprets what events mean for market conditions.
• Technical Analyst — works with indicators such as MACD and RSI to detect patterns and forecast price moves.

Once the analysts report, two researcher agents — one bullish, one bearish — debate the findings in structured rounds. The point of the debate is to make sure the risks get argued as hard as the upside. A trader agent then synthesises the debate into a proposal, a risk-management team reviews the exposure, and a portfolio manager makes the final call. The output is a rating on a five-tier scale — Buy, Overweight, Hold, Underweight or Sell — plus a written report in Markdown and HTML.

Two details matter more than the org chart. First, the framework keeps a memory log: every completed decision is written down, and later runs settle those decisions against realised returns and generate a one-paragraph reflection, so the system carries forward what worked and what did not. Second, runs are not guaranteed to agree with each other. The README says so plainly: two runs of the same ticker and date can differ, because language-model sampling is non-deterministic and live data moves. Lowering the temperature helps, but the current default models are reasoning-first and largely ignore it.

The one thing it does not do

This is the part that gets lost in the hype, so it deserves its own section. TradingAgents does not trade. When the portfolio manager approves a proposal, the README says the order "will be sent to the simulated exchange and executed." Simulated is the operative word. An independent source audit of the codebase found no broker, exchange or order code anywhere in the project; the execution the README describes is the paper's simulation, not a live order.

The README carries its own warning, and it is unusually blunt for an open-source project: "TradingAgents framework is designed for research purposes. Trading performance may vary based on many factors, including the chosen backbone language models, model temperature, trading periods, the quality of data, and other non-deterministic factors. It is not intended as financial, investment, or trading advice."

That means the correct way to think about it is as a research scaffold. It produces a structured, argued investment memo and a simulated decision you can score over time. It does not connect to your brokerage, it cannot move your money, and nobody should treat its output as a signal to buy or sell. Anyone telling you otherwise — including the YouTube videos calling it the trading bot that "trades for you" — is describing a different product from the one in the repository.

The paper behind it, and the 30.5% number

The arXiv paper reports that the multi-agent framework beat its baselines on cumulative returns, Sharpe ratio and maximum drawdown across the tested stocks. Tauric Research, the team behind the project, puts a sharper figure on its homepage: "up to 30.5% annualized returns with robust risk management." That is Tauric's own claim about its own paper's backtest, not an independent measurement, and it should be read that way.

The README is honest about the gap between a published backtest and your machine. "Backtest results are not guaranteed to match any published figure," it says. "Returns depend on the model, the temperature, the date range, data quality, and the sampling above. Treat the framework as a research scaffold for studying multi-agent analysis, not as a strategy with a fixed, replicable return."

The framework does include a backtesting tool of its own — run_backtest runs the same pipeline over a grid of tickers and dates and scores each decision on realised alpha against the instrument's regional benchmark, entered at the first close after the decision date. That is the right way to use it: score your own decisions on your own dates, rather than quoting the paper's numbers as if they were yours.

What it costs

The framework itself is free. It is licensed under Apache 2.0, there is no paid tier, no hosted service and no subscription in the repository. The money goes to the services behind each run, and there are three of them.

Language models. The defaults in the current README are OpenAI's GPT-6 family: Sol for deep thinking, Luna for quick thinking. On OpenAI's published API rate card, Luna costs $0.10 per million input tokens and $0.50 per million output tokens, while Sol costs $2 and $10. You can point the framework at almost any provider instead — Google, Anthropic, xAI, DeepSeek, Qwen, GLM, MiniMax, OpenRouter, Mistral, Moonshot, Groq, NVIDIA, AWS Bedrock, Azure OpenAI, or a local model through Ollama. Each tier can run on its own provider, so the managers can run on Claude while the rest run on OpenAI. No per-run cost is published, because it depends on the model, the debate rounds and the depth you choose; the cheap default pairing is the point.

A ChatGPT plan instead of an API key. Since version 0.7.0, you can run TradingAgents on your ChatGPT Plus or Pro plan rather than paying per token. You sign in once with tradingagents login chatgpt, and runs count toward your plan's usage limits. Plus costs $20 a month and Pro starts at $100 a month on OpenAI's pricing page. The trade-off is control: the plan route does not accept temperature or max-tokens settings.

Data. Yahoo Finance covers prices and most markets at no cost, and SEC EDGAR needs no account or key. The optional extras are cheap or free: Alpha Vantage offers a free key covering 25 requests a day, and the FRED macro feed is free with a key. The TypeSafe Jev screening for social sentiment needs a key, and is optional — without it, posts pass through unscreened.

Put together, the honest price of TradingAgents is zero for the software plus whatever your model provider charges. A casual researcher running a handful of tickers a week on Luna will spend very little; someone running deep-thinking Sol models across a backtest grid will spend real money. There is no free lunch here, only a free framework.

How to run it

Installation is a clone and a pip install. You need Python 3.11 or later, then git clone https://github.com/TauricResearch/TradingAgents.git and pip install .. Docker is supported with a compose file, and each GitHub release since 0.7.0 carries a wheel and source archive.

The interactive CLI is the main way in: run tradingagents, pick your tickers, analysis date, provider and research depth, and watch the agents work. For scheduled or scripted runs, flags and environment variables answer every prompt — for example, tradingagents --ticker NVDA --date 2026-09-23 --analysts market,news,fundamentals --save --no-show. Python users can import TradingAgentsGraph and call propagate() directly.

It works with any market Yahoo Finance covers, using exchange-suffixed tickers: US tickers plain, Hong Kong with .HK, Tokyo with .T, London with .L, India with .NS or .BO, and crypto as BTC-USD or ETH-USD. You can also hand it your actual portfolio as a JSON file, so the trader, risk analysts and portfolio manager work against your real book instead of a hypothetical one. Checkpoints are opt-in with --checkpoint, so a crashed run resumes from the last successful step instead of starting over.

Pros and cons

What it does well:

• A genuinely interesting architecture — four parallel analysts, a bull-bear debate and a risk review produce a more argued memo than any single-model prompt.
• Free and permissively licensed under Apache 2.0, with no hosted upsell anywhere in the project.
• Point-in-time discipline is taken seriously: EDGAR fundamentals are served as filed, and backtests are scored from the first close after each decision, which is how you avoid look-ahead bias.
• Broad provider support, including local models through Ollama and a ChatGPT-plan route for people who would rather not manage API keys.
• Active maintenance — fifteen releases, a changelog that follows Keep a Changelog, and a healthy closed-to-open issue ratio.

Where it falls short:

• It places no real trades, so anyone looking for an execution engine has the wrong tool — and the hype around it often suggests otherwise.
• Results are not reproducible run to run. The same ticker and date can produce different decisions, and the README admits it.
• The paper's backtest numbers are the authors' own, on a small set of stocks, and the README warns your results will not match them.
• Running it well is not free: the model bill is yours, and deep-thinking models across a backtest grid add up fast.
• Setup is a developer's job. Python, API keys and a terminal are required; there is no hosted version to click through.

TradingAgents vs the other starred trading repos

TradingAgents is the most-starred of the AI trading repositories, but stardom is not a ranking of usefulness — each of these tools does a different job. A roundup that opened all five put it plainly: match the job to the tool first.

*Star counts as logged on October 4, 2026, in the roundup linked above. TradingAgents has since passed 117,000. The sensible chain the roundup suggests is OpenBB for data, TradingAgents or FinRobot for the research memo, and Nautilus Trader only after a strategy has survived a real backtest — with a human owning the risk at every step.

FAQ

Does TradingAgents place real trades?

No. Approved orders go to a simulated exchange inside the framework. A source audit found no broker, exchange or order code in the repository. If you want live execution, Nautilus Trader is the tool in the comparison table that actually connects to brokers.

How much does TradingAgents cost?

The framework is free under Apache 2.0. You pay for the language models and any optional data keys: API usage at your provider's rates, or a ChatGPT Plus ($20 a month) or Pro (from $100 a month) plan if you use the subscription route. No per-run price is published, because it depends on the model and depth you choose.

Can I run it with local models?

Yes. Set the provider to Ollama and point it at your local server, or use any OpenAI-compatible endpoint such as vLLM, LM Studio or llama.cpp. Local models remove the API bill, but the README notes that reasoning-first models vary the most run to run, so a non-reasoning local model can actually be the more repeatable choice.

What models does it use by default?

The current defaults are OpenAI's GPT-6 Sol for deep thinking and GPT-6 Luna for quick thinking. Both can be swapped for any model your provider serves, and each tier can run on a different provider.

Is TradingAgents financial advice?

No, and its own README says so: it is designed for research purposes and "is not intended as financial, investment, or trading advice." Its output is a structured research memo and a simulated decision, not a recommendation to buy or sell anything.

Key takeaways

• TradingAgents stages a full trading firm — four analysts, bull and bear debaters, a trader, a risk team and a portfolio manager — on top of large language models, and it is free under Apache 2.0.
• At 117,328 stars it is the most-starred AI trading repository on GitHub, with roughly 6,800 of those stars arriving in October 2026 alone.
• It does not place real trades: approved orders go to a simulated exchange, and an independent source audit found no broker or order code in the project.
• Tauric's headline claim of up to 30.5% annualised returns is the paper's own backtest, and the README warns your results will not match the published figures.
• The real cost is model usage, not software: Luna at $0.10/$0.50 per million tokens, Sol at $2/$10, or a ChatGPT Plus or Pro plan instead of an API key.

TradingAgents earned its stars by turning a research paper into something you can actually run, and the October 2026 releases — ChatGPT-plan support, point-in-time fixes, backtesting that scores decisions honestly — show a project still being built rather than merely starred. Use it the way its authors ask you to: as a research scaffold that argues with itself, scores its own decisions, and never touches your money. As of October 28, 2026, that is exactly what it is — and exactly what it is not.

Sources: TradingAgents on GitHub; the arXiv paper (2412.20138); Tauric Research; OpenAI ChatGPT pricing; OpenAI API pricing; Alpha Vantage support; Anchor Terminal's TradingAgents review; GitTrend October 2026. See also our 7 best AI coding agents in 2026 and how OpenAI and Anthropic cut AI prices up to 80% in four months.