Vol. 1 · No. 1.0.3 New Delhi · Friday 13 November 2026 Aug 2026 – present

Stratify

An MCP server for honest options backtesting · Personal product · Aeon Labs

  • NIFTY 1-min option bars since 2019 ▲
  • SENSEX since 2023 ▲
  • CLICKHOUSE 372M bars
  • OAUTH 2.1 PKCE · dynamic client registration
  • TIERS enforced twice
  • HELD OUT 30%
  • WALK-FORWARD folds
  • BOOTSTRAP intervals
  • SHARPE deflated ▼
  • MIN. TRADES 30 per ratio
  • TESTS 49 files
  • SERVER 1.0.3

Backtesting · Front page

Describe a strategy in plain words. It is tested on real one-minute prices.

Inside ChatGPT, Claude or Gemini. After real Indian charges and stated slippage. With an honesty panel that refuses to flatter you.

Red and green candlesticks with moving averages on a black chart.
A strategy is only as honest as the prices it meets. Photograph: a Stratify session

The problem with backtests

Most backtests flatter you. They skip the charges, test on the same data they were tuned on, and print a Sharpe ratio off a dozen trades.

I wanted the opposite: a tool that is hard to fool, even by me. You describe a strategy the way you would to a friend. It runs on 372 million one-minute option bars, NIFTY since 2019 and SENSEX since 2023, stored in ClickHouse.

The P&L comes after brokerage, STT, exchange and SEBI fees, GST and stamp duty, plus the slippage you state up front.

Then the honesty panel goes to work. It holds out 30% of the data, runs walk-forward folds, puts bootstrap intervals around the result and deflates the Sharpe. Under 30 trades, it prints no ratio at all.

“The service never calls a strategy good.”

A2 · Technology

How a request travels

From a sentence in a chat window to a P&L you can trust. Tiers are checked twice, once in the spec and once in the database.

  1. 1

    Your chat

    ChatGPT, Claude or Gemini. You describe the strategy in plain words.

  2. 2

    MCP

    Streamable HTTP or stdio. The model turns your words into a strategy spec.

  3. 3

    OAuth 2.1

    My own server: PKCE and dynamic client registration.

  4. 4

    Spec validator

    Tier check 1. The spec is checked against your plan before any query runs.

  5. 5

    ClickHouse

    Tier check 2. Row policies decide what data your plan can read. 372M bars.

  6. 6

    Backtest

    Every entry and exit on real one-minute prices.

  7. 7

    Charges

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

  8. 8

    Honesty panel

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

  9. 9

    Answer

    Back in your chat, with the caveats attached.

A3 · Markets

Gross is not net.

The gap between the two lines is the part most backtests leave out. Every Stratify P&L is reported after charges and the slippage you state.

The last 30% of the data is held back. The strategy never sees it until the end, and the result comes with a bootstrap interval, not a single number.

  • Gross P&L
  • After charges & slippage
  • Bootstrap interval

Illustrative data · not a backtest result

Illustrative chart. A gross equity curve rises above a net curve after charges; the gap widens over time. The last 30 percent is marked as held-out data, with a shaded bootstrap interval around the net line. HELD OUT · 30% IN-SAMPLE OUT-OF-SAMPLE

Walk-forward folds

traintestheld out

A4Honesty panel

  • 30%of the data held out, unseen until the end
  • Walk-forwardfolds, so tuning never peeks at the future
  • Bootstrapintervals around every headline number
  • DeflatedSharpe, corrected for how many ideas you tried
  • < 30 tradesno ratio is printed at all

The service never calls a strategy good. It tells you how much you can trust the number, and stops there.

A5 · Your money

What a trade really costs

Charges included in every Stratify P&L
ChargeWhat it isIn every P&L
BrokerageWhat the broker takes per order✓
STTSecurities transaction tax✓
Exchange feesTransaction charges from the exchange✓
SEBI feesThe regulator's turnover fee✓
GSTGoods and services tax on the charges✓
Stamp dutyLevied on the buy side✓
SlippageStated by you, up front, and shown✓

Published from CI

Every release goes out from GitHub Actions with OIDC trusted publishing: to the official MCP Registry, npm and PyPI. No long-lived tokens on my laptop.

49 test files guard the engine, the charges and the tier rules.

A6By the numbers

One-minute option bars
372M
NIFTY data since
2019
SENSEX data since
2023
Held out
30%
Trades before any ratio
30+
Test files
49
Registries
3
Tier checks
2

A7 · Pictures

The week in screens

Video. Stratify in ten seconds.
An analytics dashboard open on a laptop.
Results. A backtest summary, after charges.
A dashboard of histograms on a dark screen.
Intervals. Bootstrap distribution of returns.
A dark laptop and a phone, both showing charts.
Clients. The same server, any MCP client.
A pile of banknotes from around the world.
Costs. Every charge, counted.

A8 · Listings

Stack

Technology stack
ComponentRole
PythonThe server and the backtest engine
MCPStreamable HTTP + stdio
OAuth 2.1Own server · PKCE · dynamic client registration
ClickHouse372M one-minute bars · row policies
DockerDeployment
GitHub ActionsCI · OIDC trusted publishing
TypeScriptnpm distribution

Next slice in time

BlueRickshaw

May 2026 – presentThe operating system for importing food & drink into IndiaOpen

All work