Zerodha Kite MCP: Zerodha's Official AI Trading Assistant Server

Most "MCP server for [brand]" posts on the internet describe unofficial, community-built wrappers around a company's public API. Zerodha's is different: Kite MCP is a first-party server built and maintained by Zerodha itself, publicly announced on Zerodha's own Z-Connect blog and open-sourced on GitHub as zerodha/kite-mcp-server. It launched on May 20, 2025, making it one of the earliest official MCP integrations shipped by a major Indian financial company.

What Kite MCP Actually Is

Kite MCP exposes your live Zerodha trading account — the same account behind the Kite web and mobile apps — to any MCP-compatible AI client. Instead of building a custom ChatGPT plugin or a bespoke API integration, Zerodha built one standards-based server that works with Claude Desktop, VS Code with GitHub Copilot, and any other MCP host at once. Zerodha's own framing is that it lets AI assistants "interact with real-world data and services" rather than reasoning from static, pre-trained knowledge about the market.

The Full Tool List — 22 Tools Across Five Groups

The zerodha/kite-mcp-server repository's own README documents exactly 22 tools, grouped as:

• Setup & authentication (1): login
• Market data (5): get_quotes, get_ltp, get_ohlc, get_historical_data, search_instruments
• Portfolio & account (5): get_profile, get_margins, get_holdings, get_positions, get_mf_holdings
• Orders & trading (7): place_order, modify_order, cancel_order, get_orders, get_trades, get_order_history, get_order_trades
• GTT orders (4): get_gtts, place_gtt_order, modify_gtt_order, delete_gtt_order

In practice this means you can ask an AI assistant something like "what's my current portfolio allocation across sectors?" or "show me my open GTT orders" and get an answer computed from your actual live account, not a hallucinated estimate.

Hosted vs. Self-Hosted: A Distinction That Actually Matters

Here's the detail that most casual coverage of Kite MCP glosses over: the codebase itself includes full order-placement tools (place_order, modify_order, cancel_order), but Zerodha's own hosted instance at mcp.kite.trade deliberately excludes potentially destructive trading operations for security, per the project's own README. That's a meaningful design choice: the public, zero-setup version most people will actually connect to is scoped down from what the open-source code is technically capable of.

If you self-host the server instead — running the open-source code yourself rather than using Zerodha's hosted endpoint — you get access to the full 22-tool set, including live order placement, but you're also the one responsible for whatever safeguards you put around an LLM that can now place real trades. This is exactly the "capability vs. exposure" tradeoff worth understanding before connecting any financial MCP server to an autonomous agent loop.

Transport Modes and Self-Hosting Requirements

The server supports four transport modes: stdio, http, sse, and a hybrid mode recommended for production. HTTP is Zerodha's own recommendation for better performance and reliability over stdio.

For the hosted version, Zerodha's README states plainly: "no installation or API keys required" — you just add the remote MCP URL to your client. For self-hosting, you'll need your own Kite Connect developer credentials, supplied as KITE_API_KEY and KITE_API_SECRET environment variables — meaning a Kite Connect developer subscription, which Zerodha prices separately from a regular trading account. The repository is MIT licensed, so self-hosting and modifying it is explicitly permitted.

Read-Only by Design (For the Hosted Version)

Zerodha built three explicit safety constraints into the hosted Kite MCP experience, and they matter more than any single feature:

• Limited access — the AI client can only reach data and actions you've explicitly authorized.
• No credential storage — your Zerodha credentials never pass through Claude, ChatGPT, or any other AI client. Authentication happens directly with Zerodha's own two-factor login flow.
• Read-only by default — on the hosted instance, everything except GTT (conditional, pre-defined) orders is read-only. Full order placement exists in the open-source code but is excluded from the public hosted endpoint specifically for security.

If a guide anywhere tells you to wire an LLM directly into unattended live order placement, understand that requires self-hosting and deliberately opting back into the destructive-operations tools Zerodha excluded by default — not something to do without your own explicit confirmation step on every trade.

Connecting Kite MCP to Claude Desktop

Zerodha's documented setup path for Claude Desktop, using the hosted instance, is:

1. Install Node.js if you don't already have it.
2. Add Zerodha's remote MCP server URL (mcp.kite.trade) to Claude Desktop's developer/MCP configuration.
3. Restart Claude Desktop so it picks up the new server entry.
4. Authorize the connection through Zerodha's normal two-factor authentication flow — this happens on Zerodha's own login page, not inside Claude.

VS Code follows the same pattern: add the server to your MCP settings in settings.json, then verify the connection from the Copilot Chat panel. If you're running the stdio transport of a self-hosted build instead, the project's own README specifically warns to use the full absolute path to your built binary in the client config, rather than a relative path that may not resolve correctly across different working directories.

Why an Official Brokerage MCP Server Matters

Zerodha is India's largest stockbroker by active client count, so an official, first-party MCP server from them is a meaningfully different signal than the many unofficial wrappers that exist for other Indian brokers and platforms — several of which (Groww, Upstox, ICICI Direct) are covered elsewhere on this site and are genuinely useful, but are community-maintained rather than shipped by the broker itself. Kite MCP is a real, working example of exactly the pattern this site's guides describe elsewhere: an existing production API (Kite Connect) exposed through a standard protocol, with authentication kept outside the AI client and the blast radius of what the model can do deliberately capped on the default, hosted path.