Your AI is Guessing About Indian Stocks. Here's How to Make It Actually Useful

The frustration most readers have already felt

Last week I asked Claude to design a defensive Indian equity strategy. It produced a confident answer — quality stocks, low debt, momentum overlay — and quoted CAGRs and Sharpe ratios that sounded plausible.

I checked the numbers. They were either out of date by 18 months or made up entirely. The strategy itself was fine. The data was fiction.

This isn't a Claude problem. It's a context problem. And there's now a clean way to fix it.

Why your AI is guessing

AI training cutoffs are months-to-years old. Indian markets move fast. XBRL filings land 30–60 days after the period-end. Most public data sources don't preserve point-in-time. Survivorship bias is silent and brutal.

An AI without grounded data is just doing creative writing dressed as research.

And the worst part: the AI's tone never changes. It sounds equally confident whether it's quoting real CAGR or hallucinating one.

The open standard that fixes it

Model Context Protocol (MCP) is Anthropic's open spec, now supported by Claude, Cursor, Claude Code, Windsurf, VS Code, and ChatGPT. Instead of an AI guessing, it calls remote tools and gets real data back.

ftInvstr ships an MCP server for Indian equities — 11 tools covering 60+ backtested strategies, 276 data fields, 36 operators, and a full bias-safe backtest engine.

Five-minute setup (browser Claude)

No API key. No JSON file. No Node.js. Browser Claude handles the whole auth dance with one click.

1. Go to claude.ai → Profile → Connectors → Add custom connector
2. Fill in the form:
   • Name: ftInvstr
   • Remote MCP server URL: https://ftinvstr.in/mcp/
   • Leave OAuth Client ID + Secret empty under Advanced settings
3. Click Add, then Connect. You'll be redirected to ftInvstr.
4. Sign in (if you aren't already) and click Allow access on the consent screen.
5. claude.ai redirects you back — the connector is live. Open a new chat and click the 🔧 tools icon to confirm ftinvstr · 11 tools appears.

Same flow for ChatGPT (Settings → Connectors). For Claude Desktop / Cursor / Windsurf / VS Code / Python — the setup uses an API key instead of OAuth; full per-client steps at ftinvstr.in/mcp-setup/.

First conversation

Open a new chat and try this:

Claude pulls up a tool-call card, calls search_strategies against ftInvstr, and comes back with a ranked list — real CAGRs, real Sharpes, no hallucinated numbers. Five seconds. No copy-paste from a dashboard. The AI's output is now anchored to data you can verify.

The interesting part — design a strategy from plain English

Watch what unfolds in the tool-call cards:

1. get_expression_help — Claude reads the DSL reference
2. Composes an expression: order(normalize(add(add(multiply(rank('ROCE %'), 0.5), multiply(rank('OPM %'), 0.3)), multiply(neg(rank('Borrowings')), 0.2))), 0.7, 0.3, 1)
3. validate_expression(...) → {ok: true}
4. submit_backtest(...) → returns a job_id
5. Polls get_backtest_status until done
6. get_backtest_result(job_id) → real CAGR, Sharpe, equity curve, year-by-year

The AI is no longer making up numbers. It's writing a hypothesis and then running it against bias-safe Indian market data. The numbers it cites back are numbers that just got computed — against real data, with real costs modelled in.

Why this actually changes how you research

The shift is from "AI tells me a strategy" → "AI runs an experiment for me and shows me the result." The difference between asking a clever friend and asking a clever friend with a Bloomberg terminal.

Concrete examples of new workflows:

• "Now run it on NIFTY500 with weekly rebalance" — and the AI does, in real time.
• "Show me how this compares to the existing DEF_LowVol60_Quality_ROCE strategy."
• "What happens if I require ROCE > 20 AND positive momentum AND low debt?"

Every number the AI cites is reproducible. You can ask it to show you the equity curve. You can re-run the same backtest yourself on the web dashboard. There's no black box.

Where this falls short

• Backtests take 15–25 minutes when the queue is loaded. The AI polls; you wait or come back later.
• Free tier: 10 submits/day per user. Read tools (browse strategies, fetch stats) are unlimited.
• This is research, not advice. ftInvstr is not a SEBI-registered adviser. The AI cannot tell you what to buy. It can help you understand what would have worked.
• The AI can still write a bad strategy — the engine just exposes that quickly, instead of letting you find out with real money.

Set it up once. Let your AI actually work

The 5-minute investment is worth it. Once it's connected, you stop asking your AI to remember things it never knew, and you start asking it to actually do research.

Ask Claude to find you a strategy. Then ask it to test that strategy. Then ask why it works. That's the loop MCP unlocks.