ChatGPT and Claude can describe a backtest; connected to Basis, they run one — on exchange candles, with fees and slippage, on closed bars only — and compare your rule with 33 built-in strategies on data they were not tuned on. You get trades, win rate, drawdown and the return against simply holding, not a story.
Ask a chat assistant without tools to backtest a strategy and one of two things happens. It writes code you then have to run yourself, or it produces statistics that look like a backtest and were never computed.
The third failure is quieter: a real test, tuned until it looks good on the same data it is judged on. That strategy has learned the past, not the market.
One server URL for every assistant. There is nothing to sign up for and no key to paste.
https://www.basischarts.com/api/mcpPlugins → the plus button → Add custom MCP server. Paste the URL, accept the notice and press Create as a plugin. Then switch Basis on in a new chat.
ChatGPT asks you to confirm before a tool that changes something runs, and building a chart counts as one. Available on plans that allow custom MCP servers.
Customize → Connectors → + Add → Add custom connector. Name it Basis, paste the URL, then Continue and Add.
Works on every Claude plan, including Free (one custom connector). There is nothing to authorise: the connector has no login step because Basis does not ask for one.
Run this once in any project. It is saved for every session afterwards.
claude mcp add --transport http basis https://www.basischarts.com/api/mcpAdd to ~/.cursor/mcp.json, or .cursor/mcp.json inside one project.
{
"mcpServers": {
"basis": {
"type": "http",
"url": "https://www.basischarts.com/api/mcp"
}
}
}Every client, tool and limit is on the Basis MCP page.
Any of 33 built-in strategies with your parameters, on exchange candles, with fees of 4 basis points a side and 1 of slippage by default, and stops, targets or time stops if you want them.
Signals use closed bars only. A rule that would have needed the next candle to decide never trades on it.
Your idea against 33 built-in strategies across timeframes, judged on held-out data with a correction for having tried many.
Replay a market-wide screen bar by bar and trade its matches as an equal-weight basket, with turnover costs counted.
Position size, liquidation price and risk at stop for the trade you would actually take.
Each one asks the assistant to check something specific, and says what it will look at — so you can tell a measured answer from a fluent one.
Backtest buying BTC on the 4h when RSI(14) falls to 30 and selling when it recovers to 50, long only, with fees. Show trades, win rate, drawdown and the return against holding.
Checks: The built-in RSI strategy on exchange candles, with costs.
Which built-in strategies held up on BTC and ETH on the 1h and 4h when judged on data they were not tuned on?
Checks: A strategy comparison with held-out validation and a multiple-testing correction.
If I had bought every coin among the 50 most traded on Binance whose 1h RSI(14) was below 30, re-checked every bar, how would that basket have done after costs?
Checks: A historical replay of the screen with turnover costs.
Run the same RSI rule with a 2% stop and a 4% target, then without them. Which is more robust, and on how many trades?
Checks: Two backtests with the same costs, side by side.
Count the trades first. Under thirty is an anecdote, whatever the win rate says.
Read drawdown before return. The deepest fall is what you would have had to sit through, and it is where most strategies are actually abandoned.
Compare with holding. A strategy that made 10% while the market made 30% lost money against doing nothing.
Trust results on data the strategy was not tuned on. If it only works where it was optimised, it learned the past.
Not reliably on its own: without tools it either hands you code or describes results it never computed. Connected to Basis, it runs the backtest on exchange history and reports what actually happened.
Yes. The default is 4 basis points a side in fees and 1 in slippage, and you can ask for different values.
Signals use closed bars only, so a rule never acts on a candle that had not finished.
Crypto spot and perpetuals on Binance, Bybit, OKX and Hyperliquid, on the timeframes the chart offers.
It is a reason to keep studying it. Look at the number of trades, the drawdown, the comparison with holding and how it did on data it was not tuned on — and size anything live small.
No. The backtester and strategy research run on the free Basis MCP server.
Paste the URL into your assistant, then ask it what the market is doing — and to show you what it checked.
https://www.basischarts.com/api/mcpSetup steps are as OpenAI and Anthropic document them in October 2026; menus change. Basis reports what venues publish and what the maths says about it. Nothing here is financial advice.