$ stratify --version 1.0.3

Stratify

An MCP server for honest options backtesting.

Describe a strategy in plain words inside ChatGPT, Claude or Gemini. It is tested on real 1-minute prices, after real Indian charges.

  • MCP Registry official
  • npm published
  • PyPI published
  • 13 Nov 2026 server 1.0.3
372M one-minute option barsNIFTY since 2019SENSEX since 202330% data held out49 test filesOAuth 2.1 PKCE · DCRv1.0.3 MCP Registry · npm · PyPI

// the problem

Most backtests flatter you.

  1. 01

    They ignore the costs.

    Every Indian charge, itemised: brokerage, STT, exchange & SEBI fees, GST, stamp duty — plus stated slippage.

  2. 02

    They test on the data they were tuned on.

    30% of the data is held out. Walk-forward folds. Bootstrap intervals.

  3. 03

    They quote a Sharpe from a handful of trades.

    Deflated Sharpe, and no ratio at all under 30 trades.

// how it works

One request, traced.

From a sentence in a chat window to a report. Tiers are enforced twice — once in the spec validator, once in the database.

  1. chat clientChatGPT · Claude · Gemini
  2. mcp transportstreamable HTTP + stdio
  3. oauth 2.1my own server · PKCE · dynamic client registration
  4. spec validatortier check 1 of 2
  5. clickhouse372M one-minute bars · row policies = tier check 2 of 2
  6. p&l enginereal Indian charges + stated slippage
  7. honesty panelheld-out · folds · bootstrap · deflated Sharpe
  8. reportback to the chat

Illustrative span layout. Shapes show order, not measured timings.

// the report

Printed, itemised, unimpressed.

P&L comes after every real charge an Indian options trader pays, and after slippage I state up front.

Then the honesty panel. It reports what the strategy did on data it never saw, with intervals instead of a single heroic number.

“The service never calls a strategy good.”

STRATIFY

BACKTEST REPORT · server 1.0.3

================================

Strategy
NIFTY short straddle
Data
1-min option bars
History
NIFTY since 2019

--------------------------------

P&L AFTER CHARGES

  • Gross P&Lfrom fills
  • − Brokerageitemised
  • − STTitemised
  • − Exchange feesitemised
  • − SEBI feesitemised
  • − GSTitemised
  • − Stamp dutyitemised
  • − Slippagestated
  • = Net P&Lafter all

--------------------------------

HONESTY PANEL

  • 30% of data held out
  • Walk-forward folds
  • Bootstrap intervals
  • Deflated Sharpe
  • No ratio under 30 trades

================================

The service never calls
a strategy good.

// by the numbers

Numbers

372,000,000 one-minute option bars in ClickHouse — NIFTY since 2019, SENSEX since 2023.
30%held-out data on every run
30trades minimum before any ratio is shown
2×tier enforcement: validator + row policies
49test files
3registries from CI: MCP Registry, npm, PyPI — OIDC trusted publishing
3chat clients: ChatGPT, Claude, Gemini

// stack

requirements.txt

# runtime
python
mcp            # streamable HTTP + stdio
oauth-2.1      # PKCE, dynamic client registration
clickhouse     # 372M rows, row policies
docker
# ship
github-actions # OIDC trusted publishing
typescript     # npm package

Next application · 03 / 07

BlueRickshaw
bluerickshaw.os

The import chain, as software.