honest-backtest-mcp

MCP server for the Honest Backtest API — backtest trading strategies against real market data, and get told in the same response what the numbers do not support.

Most backtesters will report a Sharpe ratio computed over eight trades as though it meant something, and will happily let you zero out fees and show a 400% return. The numbers are real; the impression they create is not. This server returns the usual metrics plus a warnings list naming the specific reasons a result is weak — and it puts the warnings first, because an agent that skims a tool result and stops early should hit the caveats before the returns.

Install

{
  "mcpServers": {
    "honest-backtest": {
      "command": "npx",
      "args": ["-y", "honest-backtest-mcp"],
      "env": {
        "RAPIDAPI_KEY": "your-key"
      }
    }
  }
}

Works with any MCP client — Claude Desktop, Claude Code, Cursor, Cline, Windsurf.

Getting a key

Every endpoint except /health is served through an API marketplace, so a key is required. Subscribe on RapidAPI, then set RAPIDAPI_KEY. Without it, calls return forbidden — the server starts and lists its tools normally, so you can inspect the interface before subscribing.

Configuration

| Variable | Default | Purpose | |---|---|---| | RAPIDAPI_KEY | — | Your marketplace subscription key | | RAPIDAPI_HOST | API host | Override when calling through a different gateway | | HONEST_BACKTEST_API_URL | production Worker | Point at a different deployment | | HONEST_BACKTEST_TIMEOUT_MS | 30000 | Request timeout |

Tools

list_strategies

The catalogue with every tunable parameter and its default. Call this before run_backtest rather than guessing parameter names.

run_backtest

Simulate a strategy against real historical data.

| Parameter | Notes | |---|---| | strategy | dca_trend, grid, rotation, buy_hold | | symbol | e.g. BTCUSDT. Required for every strategy except rotation | | symbols | 2–10 symbols. Required for rotation | | interval | 1h, 4h, 1d, 1w. Default 1d | | capital | Default 10000 | | candles | 30–720. Default 500. Under 200 triggers SHORT_HISTORY | | feeRate | Default 0.001 (0.10%). Setting 0 triggers NO_COSTS | | slippageRate | Default 0.0005. Setting 0 triggers NO_COSTS | | params | Strategy-specific. Unknown keys are rejected, not ignored |

Bounds are enforced in the tool schema, so an out-of-range value fails before a network call is spent.

How the engine avoids flattering itself

• Every fill pays commission and slippage, in the direction that hurts
• Orders execute at the next candle's open — never the close that triggered the signal
• Long-only, no leverage, cash can never go negative

Warning codes

| Code | Fires when | |---|---| | SMALL_SAMPLE | Fewer than 30 closed trades — risk metrics are noise | | DEEP_DRAWDOWN | Max drawdown over 30%, stated in account terms | | IMPLAUSIBLE_SHARPE | Sharpe over 3 — usually overfitting, not edge | | NO_COSTS | Fees and slippage both zeroed; unachievable live | | SHORT_HISTORY | Fewer than 200 candles — likely a single market regime | | PAST_PERFORMANCE | Always |

A note on grid

The grid centres on the first candle and never re-centres, so its lifetime result is largely decided by the price on its start date. In testing over 2021–2026, a 5-level 3% grid on ETHUSDT executed zero trades across 5.6 years while ETH rose 563%. On SOLUSDT the same configuration returned +1759%, but that was an accidental buy-and-hold with a 95% drawdown, not grid behaviour. Documented because it is true, not because it helps sell.

Not investment advice

A backtest describes the past under assumptions. It is not a prediction.

Development

npm test              # 31 unit tests, no network
npm run test:contract # verifies tool schemas against the live API; needs RAPIDAPI_KEY

License

MIT