Personal product · Aug 2026 – now

Stratify

An MCP server for honest options backtesting. Describe a strategy in plain words inside ChatGPT, Claude or Gemini. It's tested on real 1-minute prices.

  • ChatGPT
  • Claude
  • Gemini
chat · stratify connected
You

Sell a NIFTY straddle at 9:20, stop at 30%, exit by 3:15. Test it since 2019.

stratify · backtest_spec
underlying  NIFTY
structure   short_straddle · ATM
entry       09:20
stop_loss   30% of premium
exit        15:15
data        1-min option bars, 2019 →
costs       india_fno + slippage
holdout     30%
spec validated · tier checked twice
stratify · run_backtest

Report ready: net of every charge, 30% held out, folds and intervals attached. No verdict.

on theMCPRegistry
372Mone-minute option bars
2019NIFTY since · SENSEX since 2023
30%of data held out, always
49test files

The problem

Most backtests flatter you.

They skip the charges. They fit the whole history, then grade on the same history. They print a Sharpe ratio from eleven trades.

I wanted the opposite: a backtester that shows you every rupee of cost, keeps data you never fit on, and refuses to give a ratio it can't defend.

The service never calls a strategy good.

How it works

From a sentence to a ledger.

  1. 01

    Ask in plain words

    Inside ChatGPT, Claude or Gemini. The MCP server speaks streamable HTTP and stdio.

  2. 02

    Sign in properly

    My own OAuth 2.1 server: PKCE and dynamic client registration.

  3. 03

    Check the tier twice

    Once in the spec validator. Again in ClickHouse row policies.

  4. 04

    Replay real minutes

    372 million one-minute option bars. NIFTY since 2019, SENSEX since 2023.

  5. 05

    Charge every rupee

    Brokerage, STT, exchange & SEBI fees, GST, stamp duty, plus stated slippage.

  6. 06

    Show the doubt

    Held-out data, walk-forward folds, bootstrap intervals, deflated Sharpe.

Playground

Build a strategy. Watch it get audited.

Simulation

A synthetic, in-browser toy. Prices, premiums and charges here are made up from a seeded random walk — not Stratify's data and not advice.

Underlying

Variants tried: 1 — every tweak counts against the deflated Sharpe.

Where the money went illustrative ₹

    Honesty panel

    Trades
    —
    Net, in-sample
    —
    Net, held-out
    —
    Folds in profit
    —
    Sharpe (per trade)
    —
    Deflated Sharpe
    —

    No ratio under 30 trades.

    Verdict: none. The service never calls a strategy good.

    Numbers

    One-minute option bars in ClickHouse372 million
    NIFTY historysince 2019
    SENSEX historysince 2023
    Data held out from fitting30%
    Minimum trades before any ratio30
    Test files49
    Where it's publishedMCP Registry · npm · PyPI
    How it's publishedCI, OIDC trusted publishing
    Strategies called good0

    Stack

    Boring tools, strict rules.

    • Python
    • MCP · streamable HTTP + stdio
    • OAuth 2.1
    • ClickHouse
    • Docker
    • GitHub Actions
    • TypeScript

    Next project

    BlueRickshaw

    The operating system for importing food & drink into India — from a supplier abroad to a kirana shelf.