Screen NSE stocks with Claude over MCP

"Find NSE stocks above their 200-day average where delivery has stayed over 55% all week" is a prompt, and an agent connected to PatternsRadar will answer it against that evening's data. The screener exposes a Model Context Protocol server, so the whole NSE cash market arrives in the conversation as a set of tools.

This works on every plan, including Free. One API key authenticates both the REST API and MCP, and creating one takes a minute in your account settings.

The one-line setup

For Claude Code:

claude mcp add --transport http patternsradar https://api.patternsradar.com/mcp --header "Authorization: Bearer prdr_YOUR_KEY"

Any other MCP client that speaks streamable HTTP takes the same URL and header. The MCP docs page has the JSON config and a TypeScript SDK example.

Why the agent doesn't need a manual

The usual failure mode of "LLM plus market data API" is the model guessing at a query syntax it has never seen. I watched one invent three different spellings of a moving average inside a single session before I built the fix for it.

The fix is a tool called sift_reference. It returns the complete, machine-readable reference for the Sift query language, every field and indicator and operator and pattern, generated from the same catalog the compiler itself reads. An agent calls it once and then writes valid scans for the rest of the session.

That matters because Sift is the interesting part. The language has event operators most screeners cannot express:

delivery_pct has been above 55 for 5 bars

That one line is the whole of the sustained delivery scan. An agent that can write these will compose setups you would never bother clicking together by hand.

A real session

About two minutes of conversation:

1. Ask for candidates. "Run a scan for NSE stocks above the 200-day average with delivery over 1.5× their 20-day norm." The agent calls run_scan. What comes back is deliberately narrow, symbol, close, change, volume, RSI, delivery, so it does not drown its own context.
2. Make it prove the setup. "Before I look at these, has that scan actually worked?" The agent calls hitrate, which replays the query across the last 250 sessions and reports what matches did over the next 1, 5 and 20 days.
3. Keep what survives. "Save it as 'quiet strength' and I'll watch it." The save_scan tool writes it to your account, where end-of-day alerts can pick it up.

The server's own tool descriptions carry the caveat and so does this page. The hit rate is a sketch rather than a backtest: close-to-close returns, no costs, no slippage, the universe as it stands today. It answers whether a setup has been worth a look. It cannot tell you what you would have earned.

What this is actually for

Screens like delivery surge and quiet accumulation exist as pages because they are common questions. The MCP server is for the questions that are yours alone: the setup you half-believe in and want tested, or the nightly research note you want assembled by something that reads like a colleague. The data is end-of-day and NSE only. Inside that boundary, an agent holding these ten tools is a competent research assistant with thirty years of memory behind it.

The full tool list, parameter shapes and limits are on the MCP page. The same capabilities over plain HTTP live in the API reference.