Go back to Learn

Tutorial · Agentic AI · 25 min

TradingAgents: an AI trading desk on your computer

Install TradingAgents, pick a model and a ticker, and read how its analyst, researcher, trader and risk agents reach a decision. Open source and free to run.

Start tutorial
Six ivory seats around one vermilion oval table seen from above, a small report card at each seat, on navy.
Paulo Cauhi ·

TradingAgents is an open-source framework that runs a whole trading desk on your computer, with an AI agent in every chair. Analysts read the fundamentals, the news, the social chatter and the technicals. A bull and a bear researcher argue over their reports. A trader turns the debate into a proposal, a risk team stress-tests it, and a portfolio manager gives the final call. You choose the ticker, the date and the model. It was built by researchers at UCLA and MIT and is free to run.

01Install Clone the repository and install it in a Python environment.02Configure Add one API key, or point it at a free local model.03Run Answer six questions in the CLI and read the decision.

Before you start

  • A computer with macOS, Windows or Linux, plus git.
  • Python 3.10 or newer (the project's own example uses 3.12). Check with python3 --version.
  • One LLM provider: an API key from OpenAI, Anthropic, Google, xAI, DeepSeek, OpenRouter, Groq or another supported provider, or Ollama for a free local model.
  • No brokerage account. TradingAgents only analyses and recommends; it does not place orders.

Every run costs tokens. A single analysis makes dozens of model calls (the screenshots below show 45 for one full run). Start with the cheaper models and the shallow research depth, then go deeper once you know what a run costs you.

This is a research tool, not investment advice. The project says so itself. Treat the output as one more opinion to check, never as an order to execute.

Install TradingAgents

Clone the repository and enter it:

Terminal
git clone https://github.com/TauricResearch/TradingAgents.git && cd TradingAgents

Create a virtual environment and activate it (on Windows, run .venv\Scripts\activate instead of the source line):

Terminal
python3 -m venv .venv && source .venv/bin/activate

Install the package and its dependencies:

Terminal
pip install .

That gives you the tradingagents command inside this environment. Prefer Docker? Copy .env.example to .env, add your keys, and run docker compose run --rm tradingagents instead.

Add a model provider

TradingAgents reads API keys from a .env file in the project folder. Copy the template:

Terminal
cp .env.example .env

Open .env and fill in the key for the provider you will use, for example:

.env
OPENAI_API_KEY=sk-...

One key is enough. The file also has slots for ANTHROPIC_API_KEY, GOOGLE_API_KEY, DEEPSEEK_API_KEY, OPENROUTER_API_KEY, GROQ_API_KEY and others, and an optional FRED_API_KEY for macro data from the Federal Reserve.

Free local models. Install Ollama, pull a model with ollama pull <name>, and pick Ollama as the provider in the CLI. It talks to http://localhost:11434/v1 by default. Any other OpenAI-compatible server (vLLM, LM Studio, llama.cpp) works through the OpenAI-compatible provider, where you type the server's URL.

The same file can preset choices so the CLI skips a question: TRADINGAGENTS_LLM_PROVIDER, TRADINGAGENTS_DEEP_THINK_LLM, TRADINGAGENTS_QUICK_THINK_LLM and TRADINGAGENTS_OUTPUT_LANGUAGE are the useful ones.

Run an analysis

With the environment active, start the CLI:

Terminal
tradingagents

If the command is not found, python -m cli.main does the same from the project folder.

TradingAgents CLI welcome box listing the workflow: analyst team, research team, trader, risk management, portfolio management.
Screenshot: TradingAgents project.

The CLI walks you through six questions:

  1. Ticker. SPY is the default. Use Yahoo Finance symbols: AAPL, NVDA, BTC-USD, or exchange-suffixed tickers such as AZN.L or 7203.T.
  2. Analysis date. YYYY-MM-DD. Pick a trading day; weekends and holidays have no data.
  3. Output language. The reports and the final decision come out in the language you choose; English, Spanish and Portuguese are on the list.
  4. Analysts team. Space toggles each of Market, Social, News and Fundamentals; a selects all.
  5. Research depth. Shallow (1 debate round), Medium (3) or Deep (5). Deeper means more calls and more tokens.
  6. LLM provider and models. Pick the provider, then a quick-thinking model for the analysts and a deep-thinking model for the debate and the decision.

Then it runs. The left panel tracks each agent, the right panel streams every tool call and reasoning step, and the report grows at the bottom as each analyst finishes.

TradingAgents run in progress: progress table on the left, tool calls on the right, news analysis report below.
Screenshot: TradingAgents project. The news analyst's report, with the other analysts still pending.
TradingAgents market analyst report with moving averages, MACD, RSI, Bollinger Bands and ATR.
Screenshot: TradingAgents project. The market analyst pulls indicators (SMA, MACD, RSI, Bollinger, ATR) before writing.

Read the decision, then the memory

When every team has finished, the CLI asks whether to save the report to disk and whether to show it in full. The last section is the Portfolio Management Decision: the recommendation, the key arguments from the risky, safe and neutral analysts, and the plan.

TradingAgents final portfolio management decision: recommendation, summary of key arguments and refined investment plan.
Screenshot: TradingAgents project. A completed run ending in a Sell recommendation with a staged plan.

It remembers. Every completed run is appended to ~/.tradingagents/memory/trading_memory.md. The next time you analyse the same ticker, TradingAgents fetches the realised return since that decision, writes a one-paragraph reflection, and feeds the recent decisions and lessons into the portfolio manager. That is where you see why it was right and why it was wrong.

Long runs can resume. Add --checkpoint and the state is saved after each step, so a crashed run continues from where it stopped:

Terminal
tradingagents analyze --checkpoint

Use --clear-checkpoints to start fresh.

TradingAgents architecture diagram: data sources feed bullish and bearish researchers, then the trader, the risk management team, the manager and execution.
Diagram: TradingAgents project.

Use it from Python

The CLI is a front end for a LangGraph graph you can call from your own code:

from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG

config = DEFAULT_CONFIG.copy()
config["llm_provider"] = "openai"
config["deep_think_llm"] = "gpt-5.6"
config["quick_think_llm"] = "gpt-5.6-luna"
config["max_debate_rounds"] = 1

ta = TradingAgentsGraph(debug=True, config=config)
_, decision = ta.propagate("NVDA", "2026-01-15")
print(decision)

tradingagents/default_config.py lists every option: providers, models, debate rounds, output language, temperature and the paths for the cache and the memory log.

Common problems

  • The CLI asks for an API key. The key for the provider you picked is missing from .env (or the environment). Add it and run again.
  • "Not a trading day (weekend or holiday)". The date has no market data. Pick a weekday when the market was open.
  • Two runs give different answers. Expected: the models are sampled, and news and social data change over time. Lower TRADINGAGENTS_TEMPERATURE in .env for less variation; reasoning models mostly ignore it.
  • Runs are slow or expensive. Choose Shallow depth, fewer analysts and a cheaper quick-thinking model.
  • Ollama is not answering. Make sure the Ollama app or ollama serve is running and the model was pulled.
  • Windows: "no Windows console available". Run the CLI from Windows Terminal, PowerShell or cmd.exe, not from a piped or embedded terminal.

TradingAgents is open source under the Apache-2.0 licence, built by Tauric Research and described in the paper TradingAgents: Multi-Agents LLM Financial Trading Framework by Yijia Xiao, Edward Sun, Di Luo and Wei Wang. Code and documentation: github.com/TauricResearch/TradingAgents. This guide was written for version 0.4.0.

One email per update

When I publish something new about AI and agents, you get one email. No spam, and you can unsubscribe with one click.

Your address is stored by Brevo, the email service I use, only to send you these emails.

Engagement is unavailable right now. The article remains available.

Comments

Comments

Engagement is unavailable right now. The article remains available.

Comment privacyCommenting guidelines