Open-source MCP server brings live NSE/BSE data to AI assistants

Govind Sisara built Tapetide, an open-source MCP server that gives AI assistants real-time access to Indian stock market data from NSE and BSE. The motivation was straightforward: asking models like Claude or ChatGPT live financial questions either produces a refusal or a confidently wrong answer, because they have no native access to market data past their training cut-off. By implementing the Model Context Protocol — described in the post as "USB-C for AI tools" — the server lets Claude, ChatGPT, Cursor, Gemini, and Grok all call the same data source through a consistent interface. The tool catalog includes `search_stocks`, `screen_stocks` (with 326 fundamental filters), `screen_stocks_technical`, `get_company_profile`, `get_financials`, `get_stock_quote`, `get_price_history`, `get_shareholding`, `get_fii_dii_detail`, `market_ipo`, `market_fno_ban`, `add_to_watchlist`, and more than 21 tools in total, covering fundamentals, technicals, institutional flows, and portfolio tracking across all ~8,200 listed stocks.

For stdio-only clients, a ~300-line npm package with zero runtime dependencies proxies requests to the Worker, auto-detecting message framing differences across clients and refreshing tokens before expiry.

The architecture centers on a stateless Cloudflare Worker that accepts JSON-RPC over HTTPS, using the MCP SDK's `WebStandardStreamableHTTPServerTransport` in stateless mode so any edge location can serve any request identically. Authentication handles two client archetypes: chat apps like Claude.ai and ChatGPT expect a full OAuth flow with Dynamic Client Registration (RFC 7591) and discovery endpoints (RFC 8414, RFC 9728), while code editors typically accept a bearer token. The server triggers the OAuth handshake by returning a `401` with a `WWW-Authenticate` header on the first `initialize` call, and supports both short-lived HMAC access tokens and long-lived personal tokens distinguished by prefix (`tpt_rt_…`). Rate limiting uses two independent layers — a per-minute burst smoother backed by KV (fails open) and a daily/monthly quota enforced via an atomic D1 UPSERT (fails closed) — and notably refunds quota when a tool call fails due to invalid arguments, since LLMs frequently hallucinate bad parameters. For stdio-only clients, a ~300-line npm package with zero runtime dependencies proxies requests to the Worker, auto-detecting message framing differences across clients and refreshing tokens before expiry.

Key facts

1. 01Govind Sisara built and open-sourced an MCP server providing live NSE and BSE stock data to AI assistants
2. 02Covers all ~8,200 NSE + BSE stocks across fundamentals, technicals, institutional flows, and portfolio tracking
3. 03Exposes 21+ tools including `screen_stocks` with 326 fundamental filters, `get_fii_dii_detail`, and `get_stock_quote`
4. 04Runs as a stateless Cloudflare Worker; chat-app users connect by pasting one URL with no installation required
5. 05Supports both full OAuth (RFC 7591/8414/9728) for chat clients and bearer tokens for code editors
6. 06Two-layer rate limiting: per-minute KV burst smoother (fails open) and daily/monthly D1 quota (fails closed)
7. 07Failed tool calls caused by invalid arguments are automatically refunded against the user's quota

Topics