TradeAgent

English | 简体中文

An AI analyst picks the trade. Plain, tested Python places it.

TradeAgent scans US stocks and ETFs, asks a coding agent (Codex) to review the most unusual moves, and trades the survivors on Robinhood through the official agentic trading API, with sizing limits and an automatic exit. Every forecast is written down before the market answers and scored against a random pick and SPY, so you find out whether the AI is actually adding anything.

Try it · Why not something else? · Connect Robinhood · Everyday use · What protects your money · Results so far

Read this first. TradeAgent is a beta tool for learning whether an AI can pick short-term trades. It has not shown that it can make money: no strategy has a proven edge yet. Use amounts you would be happy to lose. Not financial advice.

How it works

flowchart LR
    A[Market data<br/>31 stocks & ETFs] --> B[Scanner]
    B -->|top 3 movers| C[Codex review]
    B --> D[Quant signals]
    C --> E{TradePlan}
    D --> E
    E -->|NO_TRADE| F[Recorded & scored later]
    E -->|BUY| G[Risk checks]
    G --> H[Robinhood order]
    H --> I[Timed exit & P&L]

The split is the whole point: the model decides what is worth a look; it never touches the broker. Sizing, limits, order placement, exits and reconciliation are ordinary Python that behaves the same way every time and is covered by 1,000+ offline tests.

Why not use something that already exists?

TradeAgent reuses what already works: Robinhood's official agentic trading MCP for data and orders, Codex for the analysis (on your ChatGPT plan, no per-token API bill), and standard libraries for market calendars. What it adds is the piece none of the alternatives below provide: a guarded path from an AI suggestion to a real order and back out, plus an honest record of whether the suggestions were any good.

When TradeAgent is the wrong tool: you want years of historical backtests, crypto or another broker, long-term investing, or a strategy that is already known to be profitable.

Try it in two minutes

Needs Linux or WSL2 on Windows, and Python 3.12+. No accounts.

pipx install tradeagent         # on Ubuntu/WSL: sudo apt install pipx first

tradeagent                      # list of commands, grouped by what you want to do
tradeagent demo --demo-dir ~/tradeagent-demo
tradeagent run-once --paper --state-dir ~/tradeagent-paper

That covers the demo, Robinhood sign-in and single live trades. The Codex analyst and the daily autopilot use the skill and installer in this repository, so for those clone it:

git clone https://github.com/danielye0010/TradeAgent.git
cd TradeAgent
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Each command prints a full JSON record, and a short summary when you run it in a terminal:

Demo passed: synthetic data only, 0 broker calls.
  150 forecasts recorded and scored, 15 learning runs, 5 challenger strategies created.
  These numbers exercise the machinery; they are not evidence of profit.

Paper round trip: COMPLETED (simulated broker, no real orders)
  buy  2 IWM @ 10.01 (filled)
  sell 2 IWM @ 9.99 (filled)
  Realized P&L: -$0.04

tradeagent inspect --state-dir ~/tradeagent-demo shows what the demo lab learned. Keep state directories on a local Linux disk (ext4, btrfs or xfs); TradeAgent refuses network shares and Windows mounts such as /mnt/c because SQLite locking is unreliable there.

Connect Robinhood

You need a Robinhood individual investing account in good standing. When you approve TradeAgent, Robinhood walks you through opening a separate agentic account; that is the only account TradeAgent can trade in. Put in it only what you are willing to risk.

tradeagent login        # opens Robinhood in your browser; approve TradeAgent
tradeagent doctor       # offline checklist: what is set up, what to fix next
tradeagent live-check   # read-only: account, permissions, quotes. Never places orders.

login uses the standard OAuth sign-in Robinhood offers to AI apps. Your password stays with Robinhood; TradeAgent keeps a short-lived access token and a refresh token in ~/.config/tradeagent/robinhood-oauth.json, readable by you alone (mode 600). It also creates ~/.config/tradeagent/tradeagent.toml, the one file that controls real trading:

[live]
symbols = ["QQQ", "IWM"]   # the only symbols TradeAgent is allowed to buy
max_notional = "25"        # hard dollar ceiling for any single entry

[entry]
dollar_amount = "5"        # each live entry spends exactly this much

[exit]
hold_seconds = 3600        # sell automatically after one hour

On a machine without a browser, use tradeagent login --paste and paste back the address Robinhood redirects you to. tradeagent logout forgets the session.

Everyday use

There are three ways to use TradeAgent. Only the last two can spend real money, and only when you start them yourself.

1. Ask the analyst (Codex)

Install the Codex CLI, open it in the TradeAgent folder and talk to the trade-opportunity-analyst skill:

The same steps also run by hand: tradeagent opportunity scan, decide, show and compare. See the opportunity workflow.

2. Let the daily autopilot paper-trade

A systemd timer runs the research loop every trading day with no chat window and no real orders. It needs tradeagent login, the Codex CLI signed in with your ChatGPT account, and on WSL2 systemd enabled:

python scripts/install_shadow.py --daily \
  --python "$PWD/.venv/bin/python" \
  --state-dir "$HOME/tradeagent-research/daily"

At 10:00 ET it scans the market and asks Codex to rank the top three candidates, then records four competing picks with identical entry, exit and costs:

After each run, daily-report.md in the state directory shows every arm's P&L after costs and its excess over SPY for the same hour. If Codex cannot beat Random, it is not helping. A second protocol, --protocol daily-news-v2, asks Codex the same question with and without live web search, which measures exactly what reading the news is worth. Details: daily autopilot.

3. Place one trade yourself

tradeagent run-once --live   # one entry from your config, then the automatic exit
tradeagent status            # read-only view of the current run
tradeagent recover           # finish an interrupted run; only ever sells

This is the execution path without any AI involved, useful for checking that orders, fills and exits work on your account before trusting the analyst. Full reference: live execution.

What protects your money

AI output is unpredictable; the code between it and your account is not. Before every order the engine checks, and refuses to trade if anything is off:

• Hard caps: your dollar amount, a per-entry ceiling, position and exposure limits, a cash reserve and a daily loss limit. Nothing is resized upward; there is no leverage, no short selling and no margin.
• Allowlist: only the symbols in your config, only in regular market hours.
• Fresh data: account, quotes, spread and depth must be seconds old and agree with each other, or the run halts.
• No double orders: every order is journaled with a unique ID before it is sent and is never blindly retried, so a timeout cannot turn one buy into two.
• A built-in way out: the exit is part of the same run, due after your hold time or shortly before the close. If your computer sleeps or the network drops, tradeagent recover resumes the run and can only sell. Create a file named KILL in the state directory (~/.local/state/tradeagent/live) to stop buying and exit now. No software can promise a fill in a halted or illiquid market, which is one more reason to keep amounts small.

Robinhood's own approval settings still apply on top of this. When Robinhood updates its API, TradeAgent checks every tool it uses: additive changes keep working and are noted in live-check; a change that could alter what an order means stops trading with a clear message, while recover can still exit.

Results so far

No strategy has shown a repeatable edge after costs. The best historical rule made four profitable trades in its holdout period, too few to mean anything, and its gains look like ordinary market exposure rather than skill (alpha findings). The daily autopilot started collecting forward evidence in October 2026. Telling a real edge of a few hundredths of a percent per trade from luck takes on the order of a year of daily observations, which is why every pick is recorded and compared with Random and SPY rather than judged by a good week.

Experimental research

These run alongside the main loop in their own state directories, place no orders and make no profit claims:

Project layout

.agents/skills/trade-opportunity-analyst/   the Codex skill
src/tradeagent/
  research/          scanner, signals, TradePlans, outcome scoring
  prospective/       daily autopilot and market-data providers
  oneshot.py         live entry, exit and recovery engine
  risk.py            position, exposure and cash limits
  robinhood_login.py browser sign-in and token helper
scripts/             systemd installers and release checks
docs/                design notes and operating guides

More reading: architecture, research protocol, changelog, release workflow.

Development

pip install -e ".[dev]"
pytest -q
ruff check src tests scripts && ruff format --check src tests scripts

The tests use synthetic markets, a mock broker and a fake OAuth server, so they run offline in about two minutes. CI covers Python 3.12 through 3.14 and smoke-tests the built wheel. See CONTRIBUTING.md.

License

Apache 2.0, with some vendored MIT components (details). TradeAgent is an independent project, not affiliated with Robinhood Markets, Inc. or OpenAI. Trading involves risk of loss, including loss of your entire investment.