Short answer
Give the agent a simulated account with virtual money, let it trade real market prices with a real broker's fees for a few weeks (or replay past days to go faster), and judge it on results after fees: net profit, profit factor, win rate and how many days it ended green. Papertrade does this for 45 stock markets, and your agent connects over MCP or a REST API.
Why a convincing AI can still trade badly
Large language models are good at explaining a trade. That's not the same as making money from it. An agent can give a confident, well-reasoned case for every entry and still lose over a month, because markets reward being right more often than wrong, after costs, again and again. The only way to know is to watch it make real decisions and count the results.
Costs matter more than most people expect. A day-trading agent that makes many small trades can be profitable before fees and lose money after them. Any test that ignores commission, exchange charges and taxes will flatter the agent.
What a fair test needs
- Real prices. The agent should trade the actual market, live or replayed, not made-up data.
- Real costs. The commission of the broker you'd actually use, plus the market's charges and taxes (for example STT and stamp duty in India, the SEC fee in the US, stamp duty in the UK and Hong Kong).
- The broker's rules. Market hours, holidays, intraday leverage limits and auto square-off times, so the agent can't do things a real account wouldn't allow.
- Enough sessions. One good day proves nothing. Judge an agent over weeks of trading days or dozens of replayed days.
- A record of every decision. Each order, with the time, price and which app placed it, so you can see why results came out the way they did.
Step by step with Papertrade
- Create a free Papertrade account. There's no card and no real money involved.
- Connect your AI. Add Papertrade to ChatGPT, Claude, Gemini, Cursor or another MCP app using the server URL
https://papertrade.world/api/mcp(setup guides). Building your own bot instead? Use the REST API with an API key. - Create a simulator. Pick a market (45 to choose from), the broker whose fees you want (172 brokers and apps) and the starting capital. Choose live to trade today's prices or replay to trade a past day.
- Give the agent clear rules. Tell it what it may trade, how much to risk, where to put stops and when to stop for the day. Keep the same rules for the whole test.
- Let it trade. Every tool call shows up on the Connect AI page with the app that made it, and every order appears in the simulator in real time.
- Read the report. Net P&L after fees, charges by type, win rate, profit factor, average win and loss, a calendar heat map of green and red days, streaks, per-stock results and plain-language insights.
Using Papertrade, create a live US simulator called Agent Test with $25,000 on Robinhood. Trade only large-cap stocks, hold at most 3 positions, put a stop loss 2% below every entry and close everything by 3:45 pm ET.
Live or replay?
Live simulators trade real-time prices, so they're the most honest test, but they only move during market hours and give you one trading day per day. Replays play back any past trading day minute by minute at up to 120x speed, so you can put an agent through many different days in an afternoon.
One catch with replays: a language model may remember what happened on well-known past days. Use replays to iterate quickly, then confirm the agent on live sessions it can't have seen.
Mistakes that make a test meaningless
- Judging the agent on a single session, or on its best day.
- Leaving fees out, or using a cheaper broker than the one you'd really use.
- Changing the prompt or rules halfway through and keeping the same results. Start a new simulator for each version.
- Comparing two agents on different markets, brokers or dates. Give them the same conditions.
- Looking only at win rate. See how to judge an AI trader.
Frequently asked questions
- Does testing an AI trading agent on Papertrade use real money?
- No. Every simulator uses virtual money. Papertrade never connects to a real broker or exchange, never places real orders and never moves money.
- Which AI apps can I test?
- Any app that supports MCP, including ChatGPT, Claude, Gemini, Grok, Perplexity, Le Chat, Claude Code, Codex, Cursor, VS Code and Windsurf. Your own bot or script can use the REST API instead.
- How long should I test an AI trading agent?
- Long enough to see it on good and bad days: a few weeks of live trading days, or dozens of replayed days, and enough round trips that one lucky trade doesn't decide the result.
- Is Papertrade free?
- Yes. Papertrade is free to use and doesn't ask for a credit card.
- Can I compare two AI models?
- Yes. Create one simulator per model with the same market, broker, capital and rules, and compare their reports side by side.