Letting AI Day-Trade for You: Setting Up GitHub's Trending TradingAgents Monorepo Safely

What “赛博炒股” Actually Means

“Letting AI trade for you” sounds like free money. In practice, TradingAgents is a multi-agent trading simulation: large language models (LLMs) role-play fund employees and produce a research narrative + mock order. You are running a LangGraph workflow, not depositing cash into a bot.

Why beginners still love it:

• Drama: Bull researcher vs bear researcher vs risk officer — readable logs
• Education: See how pros separate analysis, risk, and execution
• Low stakes: Default path uses historical/market data APIs and simulated execution — treat outputs as fan fiction unless you built an audited production stack

Red line: If a fork asks you to download a .exe or .dmg installer from a non-official repo, stop. Use only github.com/TauricResearch/TradingAgents.

The Virtual Trading Desk (Simplified)

Market data APIs ──► Analyst agents (fundamental, sentiment, technical, …) │ ▼ Bull vs Bear debate (N rounds) │ ▼ Trader agent (investment plan) │ ▼ Risk committee debate │ ▼ Portfolio manager ──► Simulated BUY/HOLD/SELL

Typical run: ~12 LLM calls (more if you crank debate depth). Wall-clock 5–25 minutes depending on model and API latency. RAM on your machine: ~2–5 GB for Python + LangGraph — GPU optional if you use cloud APIs only.

External deep dive: TradingAgents paper (arXiv).

Safe Mode Rules Before You Click Run

Eight-Step Beginner Runbook

Step 1 — Clone the official monorepo

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

Verify remote:

git remote -v # origin https://github.com/TauricResearch/TradingAgents.git

Step 2 — Python environment (3.11–3.13)

python3 -m venv .venv source .venv/bin/activate # Windows: .venv\Scripts\activate pip install --upgrade pip pip install .

Conda alternative (from upstream README):

conda create -n tradingagents python=3.13 -y conda activate tradingagents pip install .

Step 3 — API keys in .env (never in chat)

cp .env.example .env

Minimum for default configs (check your chosen data vendor in CLI):

# LLM — pick ONE provider you already use OPENAI_API_KEY=sk-... # ANTHROPIC_API_KEY=... # DEEPSEEK_API_KEY=... # Market data (often Alpha Vantage in examples) ALPHA_VANTAGE_API_KEY=your_key

Load env:

set -a && source .env && set +a

Step 4 — Launch the interactive CLI

tradingagents # or: python -m cli.main

You should see prompts for:

• Ticker (start with AAPL)
• Date / horizon
• LLM provider (OpenAI, Anthropic, DeepSeek, Ollama, …)
• Research depth (fewer debate rounds = cheaper)

Step 5 — Run your first “挂机” simulation

Select:

• Depth: shallow (1–2 debate rounds) for the first run
• Model: mid-tier (e.g. GPT-4o-mini class or Claude Sonnet) — not the most expensive flagship until you like the logs

Wait for completion. Output artifacts usually include debate transcripts and a final recommendation string such as FINAL TRANSACTION PROPOSAL: HOLD.

Step 6 — Read the log like a story, not an order

Open the run directory / console export. Check:

1. Did analysts cite real ticker data or hallucinate?
2. Did risk agents actually object, or rubber-stamp?
3. Does the final action match your intuition?

If the bull case cites nonsense metrics, fix data API keys before blaming the model.

Step 7 — Optional: Ollama for $0 inference

For API-free experiments (lower quality, slower):

ollama pull qwen2.5:14b export OLLAMA_BASE_URL=http://127.0.0.1:11434/v1

In CLI, pick provider Ollama. Budget 16 GB RAM on Apple Silicon for 14B-class models; see M4 memory matrix.

Step 8 — Schedule runs on an always-on machine (optional)

Laptops sleep; long debates die mid-graph. A small always-on Mac (local or remote) keeps LangGraph checkpoints intact. SSH in, run inside tmux, and pull logs next morning — same pattern as OpenClaw on remote M4.

tmux new -s tradingagents tradingagents # Detach: Ctrl-b d

What You Get: Sample Output Shape

After a successful run you typically have:

This is not a bank statement. Drop the JSON into your own backtest notebook if you want statistics — TradingAgents does not replace compliance review.

Cost and “Will AI Lose Money for Me?”

For comparing model platforms (not install), read TradingAgents vs FinGPT 2026. For social “what if everyone did X?” swarms, see MiroFish setup — complementary, not interchangeable.

Troubleshooting

ModuleNotFoundError: tradingagents

Fix: Activate venv and reinstall from repo root:

source .venv/bin/activate && pip install -e .

CLI exits immediately / empty tool list

Fix: Confirm .env loaded; print keys without revealing them:

python -c "import os; print('OPENAI' in os.environ, 'ALPHA' in os.environ)"

Run hangs after “Risk committee”

Fix: Provider rate limit. Switch model, reduce debate rounds, or wait 60 s and retry. Check provider status page.

Suspicious GitHub fork with Windows .exe

Fix: Delete it. Official project is source-only Python from TauricResearch/TradingAgents.

Recommended Path

If X → Y: If you have not read the bull/bear log once, do not connect a live broker.

FAQ

Is TradingAgents legal to use?

The framework is open source (Apache-2.0). Your compliance depends on jurisdiction, data licenses, and whether you trade real money. This guide covers research simulation only.

How is this different from Robinhood or a trading bot?

Broker apps execute regulated orders. TradingAgents is an LLM role-play orchestrator for research narratives and simulated decisions — closer to a desk meeting simulator than a certified auto-trader.

Do I need a Mac?

No. Linux or Windows work. Mac is convenient for Python data stacks and matches other KuzCloud guides if you already rent a remote M4.

Can I use Claude instead of GPT?

Yes. Upstream supports multiple providers via .env and CLI selection. Use the key for the provider you pick.

What tickers work out of the box?

Liquid US equities (e.g. AAPL, MSFT, SPY) are the least painful starters. Crypto variants exist in community configs — read README before enabling.