MCP for Trading: How Indian Brokers Are Opening Up to AI
Model Context Protocol (MCP) is steadily becoming the standard way software hands live data to AI assistants. Indian broking is now its most visible consumer use case. Zerodha's Kite MCP was the first broker-backed MCP server in India. Here is what the protocol actually is, what your broker exposes through it, and where it goes next.
MCP in one paragraph
Before MCP, an AI assistant knew two things: its training data and whatever you pasted into the chat. MCP (Model Context Protocol, open-sourced by Anthropic in late 2024) adds a third channel. The assistant calls tools on a server — get_holdings, get_quotes — and answers from live results. For trading, this is the difference between "generally, diversification matters" and "your top holding is 31% of your portfolio and you have no stop-loss on it."
Why Zerodha going first matters
Kite MCP (May 2025) made India's largest retail broker the first to offer a broker-run MCP server. It is free, at mcp.kite.trade. Its design choices are a model for how broker-AI integration should work:
- Credentials stay home. Login and 2FA happen on Zerodha's own pages. The AI tool receives only a revocable session — never your password.
- Read-mostly by default. ~22 tools covering profile, margins, holdings, positions, MF holdings, quotes, OHLC, historical candles, instrument search and order history.
- GTT-only writes. Order tools are limited to conditional GTT orders. Market-order automation is left out on purpose — that is the regulatory and liability answer to "can AI trade for me?"
- Consumer-grade access. No developer app, no API key, no ₹-per-month — unlike the classic Kite Connect path.
What AI assistants can and can't do with it
Can: answer account-specific questions from live data, produce portfolio analysis and risk reviews, maintain a trade journal, monitor GTT coverage, and draft protective GTT orders for approval.
Can't: place market orders, run unattended strategies, or act without your approval on anything involving orders. That is not an MCP limit. It is Zerodha's (and SEBI's) rule, built into the toolset itself.
Where this goes next
- More brokers. The pattern (broker-run MCP, read-only scope, GTT-style writes) is repeatable. Expect other Indian brokers to follow with their own servers.
- Richer surfaces. MCP is gaining UI parts — tools that render as widgets rather than raw JSON — which suits trading dashboards well.
- Agent standards for money. The hard problem is not technical. It is the approval and recovery model for financial actions. GTT with approval is the first serious answer, and it generalises.
- Regulatory framing. SEBI's algo rules will keep automatic execution out of consumer AI. The winning products will be the ones that make the human-in-the-loop effortless.
Quick answers
Do I need to know what MCP is to use it? No — with KitePilot you just log in and chat.
Is MCP the same as an API? It is a protocol on top of APIs: same data, standardised tool interface for AI assistants.
Is my money at risk if I connect? The connection is read-scoped and revocable. GTT is the only write path, and each order needs your approval. The main risk to manage is which MCP client you trust. Prefer the official server or a UI that is transparent about it.