Install TradingAgents and Start Researching Stocks

TradingAgents: one-prompt setup + the best prompts

How to use this file: drag it into Claude Code and type:

Set this up for me.

Claude installs TradingAgents, connects it to Claude or your ChatGPT plan, runs a quick test, and shows you the prompts below. Afterwards, paste any prompt from the second half of this file to put it to work.

TradingAgents is a free, open-source AI trading firm: analysts, a bull and a bear, a trader, a risk team, and a portfolio manager who makes the final call. It's a research tool, not financial advice. Nothing here places a real trade.

PART 1: SETUP (instructions for Claude)

You are setting up TradingAgents for someone who may not code. Do every step yourself. Before each step, say in one short line what you're doing. Never ask the user to paste an API key or password into the chat: keys go into the .env file, which you open for them. Stop and explain plainly if a step fails.

Shell: every command below is for a Bash shell (Mac, Linux, or Git Bash on Windows, which is Claude Code's default shell there). On Windows, replace `.venv/bin/` with `.venv/Scripts/` everywhere.

Step 1: Prerequisites

• Check that git works (git --version). If it's missing: on Mac, run xcode-select --install and ask the user to click Install; on Windows, have them install Git for Windows from git-scm.com.
• Check that uv works (uv --version). If it's missing, install it:
• Mac/Linux: curl -LsSf https://astral.sh/uv/install.sh | sh, then source $HOME/.local/bin/env
• Windows (PowerShell): powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex", then open a new shell
• uv downloads the right Python version itself. TradingAgents needs Python 3.11+, and many computers ship with an older one, so don't use the system Python.

Step 2: Download TradingAgents

Install into a folder called TradingAgents in the user's home folder. If it already exists, ask before touching it.

cd ~ && git clone --branch v0.7.0 --depth 1 https://github.com/TauricResearch/TradingAgents.git

This pins version 0.7.0, the version these prompts were tested on.

Step 3: Install it

cd ~/TradingAgents && uv venv --python 3.13 && uv pip install .

Then check it: cd ~/TradingAgents && .venv/bin/tradingagents --help should print a list of options, including the backtest and login commands.

Step 4: Connect an AI

Ask the user: "Do you want to run this on your ChatGPT Plus or Pro subscription, or on a Claude API key?"

Option A: ChatGPT subscription (no extra bill; counts toward the plan's usage limits)

1. Run cd ~/TradingAgents && .venv/bin/tradingagents login chatgpt in the background. It prints a link starting with https://auth.openai.com/ and then waits.
2. Give the user that link. They open it, sign in, and approve TradingAgents. The browser returns to their computer on its own.
3. Wait until the command prints Signed in to ChatGPT as .... If it hasn't after 10 minutes, stop it and run it again.
4. List the models their plan offers. The first word on each line is the model id:

   cd ~/TradingAgents && .venv/bin/python -c "from tradingagents.llm_clients import chatgpt_auth as a; [print(i, '-', n) for n, i in a.list_models()]"

Use the first model id listed for both model settings in Step 5.

Option B: Claude API key (pay per run; check the cost in the Anthropic console after the first one)

1. Tell the user to create a key at console.anthropic.com → API Keys. Create it inside a workspace. Keys with "Organization" scope get rejected.
2. Create the .env file from Step 5, open it for them (open ~/TradingAgents/.env on Mac, notepad ~/TradingAgents/.env on Windows), and wait while they paste the key after ANTHROPIC_API_KEY= and save.
3. Check the key works without printing it:

   cd ~/TradingAgents && .venv/bin/python -c "from dotenv import load_dotenv; load_dotenv('.env'); import anthropic; print(anthropic.Anthropic().messages.create(model='claude-sonnet-5-5', max_tokens=5, messages=[{'role':'user','content':'Say OK'}]).content[0].text)"

If it fails with "not scoped to a workspace", they need a new key created inside a workspace. "invalid x-api-key" means the key was mistyped or revoked.

Step 5: Settings

Create `~/TradingAgents/.env`. TradingAgents reads it automatically when it runs from that folder. These settings let it run start to finish without stopping to ask questions. Write values plainly: no quotes, no trailing spaces.

For ChatGPT (use the model id from Step 4):

TRADINGAGENTS_LLM_PROVIDER=chatgpt
TRADINGAGENTS_DEEP_THINK_LLM=<model id>
TRADINGAGENTS_QUICK_THINK_LLM=<model id>
TRADINGAGENTS_MAX_DEBATE_ROUNDS=1
TRADINGAGENTS_MAX_RISK_ROUNDS=1
TRADINGAGENTS_OUTPUT_LANGUAGE=English
TRADINGAGENTS_NO_UPDATE_CHECK=1

For Claude:

ANTHROPIC_API_KEY=
TRADINGAGENTS_LLM_PROVIDER=anthropic
TRADINGAGENTS_DEEP_THINK_LLM=claude-opus-5-5
TRADINGAGENTS_QUICK_THINK_LLM=claude-sonnet-5-5
TRADINGAGENTS_ANTHROPIC_EFFORT=medium
TRADINGAGENTS_MAX_DEBATE_ROUNDS=1
TRADINGAGENTS_MAX_RISK_ROUNDS=1
TRADINGAGENTS_OUTPUT_LANGUAGE=English
TRADINGAGENTS_NO_UPDATE_CHECK=1

Optional, free: a FRED key (fred.stlouisfed.org → My Account → API Keys) adds interest rates, inflation and the VIX to the analysis. If the user wants it, add FRED_API_KEY= and let them paste it in.

Step 6: Test run (cheap, about 2-7 minutes)

Run one small analysis to prove everything works. Use the most recent weekday as the date (YYYY-MM-DD). The test writes to a throwaway memory file so it doesn't count as a real call.

Claude (the deep model is swapped for Sonnet to keep the test cheap):

cd ~/TradingAgents && TRADINGAGENTS_DEEP_THINK_LLM=claude-sonnet-5-5 TRADINGAGENTS_MEMORY_LOG_PATH=.setup_test_memory.md \
  .venv/bin/tradingagents --ticker NVDA --date YYYY-MM-DD --analysts market --save --no-show < /dev/null

ChatGPT:

cd ~/TradingAgents && TRADINGAGENTS_MEMORY_LOG_PATH=.setup_test_memory.md \
  .venv/bin/tradingagents --ticker NVDA --date YYYY-MM-DD --analysts market --save --no-show < /dev/null

Run it in the background and wait for it to finish (it prints "Analysis Complete"). Then find the report folder and read the rating:

REPORT=$(ls -td ~/.tradingagents/logs/reports/NVDA_* | head -1) && head -5 "$REPORT/5_portfolio/decision.md"

Tell the user the rating in one line, then open $REPORT/complete_report.html in their browser (open on Mac, start on Windows).

Step 7: Save the instructions for next time

Copy Part 2 and Part 3 of this file, word for word, into ~/TradingAgents/CLAUDE.md. Claude Code reads that file automatically whenever it is opened in the ~/TradingAgents folder, so the prompts keep working in future sessions without this file being dragged in again. If a CLAUDE.md already exists there, add these sections to the end of it instead of replacing it.

Step 8: Done

Tell the user: "TradingAgents is set up. Here are the prompts you can use:" Then show them the prompt titles from Part 3 as a short numbered menu, one line each. Finish with one line: "Next time, open Claude Code in your TradingAgents folder and paste any prompt. The setup is remembered there."

PART 2: HOW TO RUN AN ANALYSIS (reference for Claude)

Whenever a prompt asks you to run TradingAgents, follow these rules.

The command. Always start with cd ~/TradingAgents && so the .env settings load. Without them the run stops with "No terminal to answer the setup questions".

cd ~/TradingAgents && .venv/bin/tradingagents --ticker TICKER --date YYYY-MM-DD --analysts LIST --save --no-show < /dev/null

• Date: today, unless the user names one. Data is used as of that day's close.
• Analysts:
• Stocks: market,sentiment,news,fundamentals.
• Crypto (BTC-USD, ETH-USD, SOL-USD) and funds/ETFs (SPY, QQQ, GLD, TLT, sector ETFs): market,sentiment,news. They have no company fundamentals.
• "Quick mode" means --analysts market only. Cheaper, good for screening a list.
• The user's holdings: add --portfolio my_portfolio.json. The trader, risk team and portfolio manager then judge the position the user already holds. See the file format below.
• Timing: a run takes about 2-8 minutes depending on the model and how many analysts run (ChatGPT plans tend to be slower). Always run it in the background and wait for "Analysis Complete". Run at most 2 tickers at the same time; more triggers rate limits on Reddit data.
• Finding the report: the printed path wraps across lines, so don't parse it. Take the newest folder for that ticker instead (folders are named by when the run happened, not the analysis date):

  REPORT=$(ls -td ~/.tradingagents/logs/reports/TICKER_* | head -1)

Inside it:

• 1_analysts/ (market, sentiment, news, fundamentals reports)
• 2_research/ (bull.md, bear.md, manager.md)
• 3_trading/trader.md (entry, stop-loss, sizing)
• 4_risk/ (aggressive, conservative, neutral)
• 5_portfolio/decision.md (final rating, summary, price target, time horizon)
• complete_report.html: the whole report as one page. Open it for the user when a prompt asks for a report.
• Ratings, from most to least bullish: Buy, Overweight, Hold, Underweight, Sell. Overweight means "own more than a normal-sized position", a milder Buy. Underweight means "own less", a milder Sell. Explain this the first time a rating comes up.
• Backtests (scoring past calls):

  cd ~/TradingAgents && .venv/bin/tradingagents backtest TICKER --start YYYY-MM-DD --end YYYY-MM-DD --every 7 --analysts market < /dev/null

• It runs one analysis per date, so each date costs about as much as one quick run.
• For crypto tickers (BTC-USD, ETH-USD and so on) add --asset-type crypto. A normal analysis detects crypto by itself; a backtest assumes stocks unless told.
• Each call is scored over the 5 trading days after the date, so the end date must be at least 8 trading days ago.
• It prints a summary per rating (number of calls and average "alpha", the return against the market) and the path of a log with each date's call and outcome. Read that log to build a date-by-date table.
• Memory: the first run for each ticker and date is saved in ~/.tradingagents/memory/trading_memory.md. Later runs score those calls once their holding period has passed and learn from them. A repeat run of the same ticker and date isn't saved again.
• When you summarise a run: plain English, short. Lead with the rating. Say when an analyst reported missing data (for example "news unavailable"), so the user knows what the call is based on.
• Never place real trades. This is research, not financial advice.

Portfolio file (~/TradingAgents/my_portfolio.json). Help the user fill it in when a prompt needs it:

{
  "cash": 5000,
  "currency": "USD",
  "positions": [
    {"ticker": "NVDA", "quantity": 40, "average_price": 197.58},
    {"ticker": "BTC-USD", "quantity": 0.15, "average_price": 60000}
  ]
}

Investor profile (~/TradingAgents/investor_profile.md). A short plain-English document about the user: goals, age, time horizon, how much risk they're comfortable with, their investing style, and any fixed rules (for example "keep 10% cash at all times", "hold 20% gold", "at least 10% Bitcoin"). TradingAgents doesn't read it; you do. When a prompt mentions the investor profile, help the user write it if it doesn't exist yet, then weigh the desk's ratings against it and point out any rule a suggested change would break.

Troubleshooting

PART 3: THE PROMPTS

Paste any of these into Claude Code. Replace anything in [BRACKETS].

1. Second opinion: before you buy anything

Run TradingAgents on [TICKER] for today with the full team. I'm thinking about buying $[AMOUNT] and I currently hold [NOTHING / X shares at $Y]. Tell me in plain English: the final rating, the bull's best argument, the bear's best argument, the price that would prove the trade wrong, and whether the desk would make my buy smaller. Then open the full report.

2. The alpha hunt: rank my watchlist

Run TradingAgents in quick mode on each of these: [NVDA, AMD, AVGO, TSM, PLTR, META]. Rank them from most to least bullish in a table with one line on why. Then run the full team on the top 2 and tell me which one has the stronger case and what would make it wrong.

3. Full portfolio review

Help me fill in my_portfolio.json with my holdings and write my investor_profile.md if I don't have one. Then run TradingAgents on every holding, with the portfolio file, so each call takes my whole portfolio into account. Give me a table with each holding's rating and a one-line reason, then rank the 3 biggest changes the desk suggests in line with my investor profile, starting with the most urgent. Flag any change that would break one of my rules.

4. Concentration check: where am I overexposed?

Look at my_portfolio.json and my investor_profile.md. Group my holdings by sector, asset type and theme (for example AI, crypto, gold, big tech) and show me what percentage of my portfolio each one is. Then run TradingAgents in quick mode on my 3 largest exposures and tell me where I'm more concentrated than my profile says I should be, and which position the desk would trim first.

5. Devil's advocate on my biggest position

My biggest position is [TICKER], about [X]% of my portfolio. Run TradingAgents on it with my portfolio file. Give me only the bear's case and the risk team's view on my position size. Don't soften it. Finish with the one price level or event that would make the bear right.

6. Earnings prep: the night before a report

[TICKER] reports earnings [DATE]. Run TradingAgents on it for today. Tell me what the bull and bear each expect, the price levels the desk says matter after the report, and whether it would hold through earnings or trim first.

7. Sector scan: the best 3 in a sector

I'm bullish on [SEMICONDUCTOR] stocks. Make a list of the 50 largest US-listed [SEMICONDUCTOR] companies by market cap. Run TradingAgents in quick mode on each one, at most 2 at a time, and rank them from most to least bullish in a table. Then run the full team on the top 3 and tell me which one has the strongest case and what would make each one wrong. (This scans a lot of stocks, so it costs more than a single run.)

8. Sector rotation: where's the strength?

Run TradingAgents in quick mode on these sector ETFs: XLK, XLF, XLE, XLV, XLI, XLY, XLP, XLU. Rank them from most to least bullish with one line each. Then run the full team (market, sentiment, news) on the top and bottom one and explain the difference.

9. The honesty test: does it actually work?

Test whether TradingAgents' calls on [TICKER] would have made money. Use its backtest command in quick mode for every 7 days over [THE LAST 2 MONTHS], ending at least 2 weeks ago. Show me a simple table: date, rating, what the price did afterwards, and whether the call beat the market. Tell me honestly whether it looks better than just holding.

10. Claude vs ChatGPT: same stock, two brains

*(needs both set up: a Claude key in .env and a ChatGPT sign-in)*

Run TradingAgents on [TICKER] for the same date twice: once on Claude, once on my ChatGPT plan (for the ChatGPT run, put TRADINGAGENTS_LLM_PROVIDER=chatgpt and both model settings in front of the command). Compare the two final calls side by side: where they agree, where they differ, and which arguments each one weighted most.

11. Morning brief: every day, automatically

Set up a daily job on this computer that runs at [7:00am] on weekdays. It should run TradingAgents on each holding in my_portfolio.json, with the portfolio file, and save a short summary file: each holding's rating, any rating that changed since yesterday, and the top 3 suggested changes. Use the scheduler built into my operating system, show me how to turn it off, and run it once now so I can see the summary. Tell me plainly that it only runs while this computer is on and awake at that time.

12. Paper trading: let the desk place practice trades (advanced)

*(needs a free Alpaca account switched to Paper Trading, and its paper API keys)*

Connect TradingAgents to my Alpaca PAPER trading account. Install alpaca-py into the TradingAgents environment, add ALPACA_API_KEY and ALPACA_SECRET_KEY to .env, open .env so I can paste my keys, and never ask for them in chat. Refuse to run if the key doesn't start with "PK" or the account isn't a paper account. Then write a small script that reads the latest rating for each holding and shows me the paper orders it would place (Buy/Overweight adds, Underweight/Sell trims). Show me a dry run first and only place the paper orders after I type YES.

13. One paper order: act on a call

*(needs prompt 12 set up first)*

TradingAgents just rated [TICKER] a [RATING]. Place a PAPER order on my Alpaca account to buy $[1,000] of [TICKER] at the current market price. Check the account is a paper account first, show me the order before sending it, and only send it after I type YES. If the market is closed, tell me the order will fill at the next open.

*Built for the Miles Deutscher Finance video on TradingAgents (github.com/TauricResearch/TradingAgents). Research only. Not financial advice. Paper trading only.*