How to backtest a crypto strategy with an AI agent

The slow part of testing a trading idea was never the idea. It's the plumbing: write the indicator, pull historical candles, configure a backtester, run it, read the output, tweak a parameter, run it again. An hour per idea, and most ideas don't survive contact with real data. So the thing that actually limits you is how fast you can test an idea and throw it away.

RiskGate collapses that loop. You connect an AI agent — in Claude Code or Cursor — to your account, describe the idea in plain English, and the agent writes the strategy, backtests it on years of real candles, reads the trades back, and refines it. You read the results.

What you connect

The bridge is an MCP server — a small Model Context Protocol server scoped to your account. One install command registers it with Claude Code and Cursor. After that, your agent has a set of tools: read the strategy API, write and save a strategy, validate it, run a backtest, read the full trade list back, and manage price alerts.

You don't write code. You describe what you want; the agent translates it into a TypeScript strategy and runs it. And because it's real TypeScript with the indicators built in, what comes out is readable, editable code you own — not a black box.

The loop, concretely

Say you want to test: go long when the fast EMA crosses above the slow EMA on BTC 4h.

The agent reads the strategy API first, so it knows how to fetch candles, read positions, and place orders. It writes the strategy in TypeScript, validates it (the validator flags look-ahead and anything that can't run in the sandbox), saves it, and runs a backtest across years of 4-hour candles. Back come the equity curve, win rate, profit factor, max drawdown, and every trade it took. From there it can tune the EMA lengths, try a different entry — a Donchian channel breakout, say — and line the variants up side by side. All in one conversation.

Backtests you can trust

RiskGate's engine fills every order at the next bar's open and shows the strategy only the bars up to the current cursor, so the result reflects how the idea would genuinely have traded, with hindsight kept out of it. You get the whole trade list — every entry and exit — so you can audit exactly why the equity curve moved. When a strategy holds up here, you can believe it.

Where it fits

Charting is free, and registered users can backtest. The AI agent is a paid add-on — the MCP server. None of it requires exchange keys: research and backtesting are entirely separate from live trading, which is its own optional tier. So you can take an idea from a sentence to a tested strategy without ever touching a live account.