One question, one answer

Where can I test a trading bot for free?

Answered with our own numbers: what is running on MarginPad’s paper-trading API right now, and what it costs to put a bot on it.

The full Bot API referenceThe public bot arenaPaper trade by handBacktest an idea first

Why an exchange testnet is not the same test

The obvious answer to this question is "the exchange testnet", and it is the wrong one often enough to be worth saying plainly. Testnet order books are thin, so a fill you get there is a fill you would not get with real liquidity. Testnet prices drift from the real market, sometimes by percent, because nothing arbitrages them back. And testnets are reset on the operator’s schedule, which ends a forward test that was two weeks into proving something.

The alternative is to keep the prices real and simulate only the money. Everything on this page is priced from the same live multi-exchange feed our charts and our own paper terminal run on. Your bot places real orders against real prices; what is simulated is the dollar that settles.

What has to be simulated before a test means anything

A simulator that fills every order at the mid price and charges nothing will tell you that almost any strategy works. Five things decide whether a paper result survives contact with an exchange:

Fees on both legs. A taker fee is charged when you open and again when you close. On a $10,000 crypto position that is about $11 a round trip. Any strategy whose average winner is smaller than its round trip is a losing strategy no matter how often it is right, and a simulator that hides the fee hides exactly that. You can also charge the paper account at a named venue’s published schedule, so the test settles the way the live account will.

Funding. Perpetuals pay funding periodically. A position held across marks accrues it, and a carry strategy that ignores it is measuring the wrong thing.

Liquidation, on the wick. Isolated margin with a maintenance rate, checked against one-minute candle extremes rather than closes, so a spike that retraced still takes the position. A simulator that only checks closes will under-report exactly the events that kill real accounts.

The fill price, and where the liquidation sits. A market order does not land on the screen price - it walks the book, and it costs more the thinner that book is. Maintenance margin is not one number either: an exchange raises it with position size, so a large position is closed sooner than its leverage alone suggests. Both are switches here rather than assumptions. Off by default a fill lands on the live price at a flat 0.5%, which is right for learning and optimistic for proving out a strategy; turn them on and the fill moves against you the way a real book does while the liquidation moves with size, per account or per call. Every trade records which way it ran and the public bot board prints it, so no result here can be quoted without saying what it was measured under.

Orders that outlive your process. Limit entries, stop entries, stop-losses, targets and trailing stops have to be evaluated on the server. If they only exist while your script is running, you are testing your uptime, not your strategy.

Backtest, paper, live - and why the middle one gets skipped

A backtest replays historical candles. It is fast, it filters obvious losers, and it cannot catch a race condition in your order logic, a position-sizing bug that only appears after a losing streak, or leverage that quietly grows past what the margin survives. Those are the failures that empty accounts, and they only show up when the same code runs forward against prices nobody has seen yet.

Forward testing is the step most people skip, because it costs time rather than money. It is also the only stage that runs the exact code path that will run live. Do all three in order: backtest the idea, forward-test the bot for days or weeks, and fund an exchange account only once the paper result holds.

Start in three calls

The first one needs no key at all. GET /api/bot/v1/price?symbol=BTC is keyless and CORS-enabled, so it works from a terminal or a browser tab right now. Sign in with an email to mint a key, then POST /api/bot/v1/open places a leveraged position at the live price and returns its liquidation level, and GET /api/bot/v1/positions reports what happened while you were away. Stops, targets and liquidations settle on our side whether or not your bot is connected.

Official clients with no dependencies install with pip install marginpad and npm install marginpad, an MCP server at https://marginpad.io/mcp exposes the same account to Claude, ChatGPT or Cursor, and the full reference with a quickstart lives on the Bot API page. The free plan is the whole engine: 120 requests a minute, three keys, fifty open positions, the WebSocket stream, replay of a past day and every keyless market-data endpoint.

Questions

Related: the full Bot API reference · the public bot arena · paper trade by hand · backtest an idea first