What Is an MCP Connector?

MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic in late 2024, that defines a uniform way for AI assistants to connect to external tools and data sources. The “context” in the name is the live information a model needs but does not hold: your account, current market data, a company's filings, your calendar. Before MCP, every assistant-to-tool integration was bespoke. MCP makes it a shared protocol, so one assistant can plug into many tools and one tool can serve many assistants.

The protocol is open and supported across AI clients. Claude (through Claude Desktop) and ChatGPT both support MCP, as do developer tools like Cursor. That breadth matters: because the protocol is shared, a connector you build or use is not locked to a single assistant. MCP defines the conversation between an AI client and an MCP server, and that is the foundation everything below rests on.

An MCP connector is a program that implements the Model Context Protocol against one specific data source or tool. “Connector” and “server” describe the same thing from two angles: it is a server because it answers requests from an AI client, and a connector because it bridges the assistant to an outside service. A connector advertises a set of tools (functions the assistant can call, like “get positions” or “get quotes”) and resources (data the assistant can read). When you ask a question, the assistant decides which tool to call, the connector runs it against the underlying service, and the result flows back into the conversation.

A useful way to picture it: the AI client is the brain, the connector is a sense organ wired to one part of the outside world. A market-data connector lets the assistant see prices. A brokerage connector lets it see your account. A calendar connector lets it see your schedule. Each connector exposes only the tools its author chose to expose, which is also how access is scoped and kept safe.

In investing, an MCP connector matters because it turns a general AI assistant into a portfolio-aware one. Without a connector, Claude or ChatGPT can explain what a diversified portfolio is in the abstract, but it has no idea what you actually own. A brokerage MCP connector closes that gap: it gives the assistant secure access to your real account, so you can ask “how are my holdings doing this year?”, “what am I most concentrated in?”, or “which of my positions lagged the S&P 500?” and get answers grounded in your live data rather than a generic example.

Connectors differ in how far that access goes. A read-only brokerage connector lets the assistant report your positions and balances but cannot move any money. A read-and-trade connector can also place orders, usually with a safeguard that requires your explicit approval before anything reaches the broker. The connector supplies the data and, where enabled, the ability to act; the decisions stay with you. Walnut is built as a read-only connector by default for exactly this reason.

The flow is the same regardless of what the connector connects to. Concretely, for a brokerage connector:

• The AI client connects to the MCP server. Claude Desktop, ChatGPT, or another client opens a connection to the connector and reads the list of tools it advertises.
• You ask a question. For example, “how is my portfolio doing?” The assistant decides which tool answers it, such as “get positions”.
• The connector forwards the request. It calls the underlying service: your broker, directly or through an aggregator like SnapTrade, or a market-data API.
• Data flows back. The service returns your positions or the quotes, the connector hands them to the assistant, and the assistant reasons over the real numbers and answers in plain language.

Nothing in that loop requires the assistant to know your broker's API: the connector handles the translation. That separation is the whole point of the protocol. The assistant speaks MCP; the connector speaks both MCP and the service's native interface.

The most important property of an MCP connector is what its tools are allowed to do. Read-only connectors expose tools that only fetch information: positions, balances, prices. They can describe your portfolio but cannot place an order, so even if the assistant suggested a trade, the connector has no way to execute it. Read-and-trade connectors additionally expose an order-placement tool, which is more powerful and demands more care.

You are the one who authorizes the connection, and a well-built connector keeps that authorization tightly scoped. The safer designs share three traits. First, read-only is the default and any trade is gated behind your explicit approval. Second, the connector never stores your broker password: it uses scoped API tokens or a regulated aggregator, so your login stays at the broker. Third, you can keep access read-only, which means the assistant can look without being able to touch. Before connecting anything to a live account, confirm whether it can place trades and what approval controls exist.

Concrete examples make the category clearer. Brokerage MCP connectors connect a real account: Alpaca publishes an official open-source MCP server for Alpaca accounts (with paper trading), and the community has built connectors for Interactive Brokers and Tradier. Market-data MCP servers like Alpha Vantage, Polygon, and Yahoo Finance feed an assistant prices and fundamentals but never touch an account. Crypto-exchange MCP servers (often built on the CCXT library) connect exchanges like Coinbase and Kraken.

The Walnut connector is a portfolio-aware brokerage MCP connector. You connect a real broker once through the regulated aggregator SnapTrade, then read your live portfolio from Claude Desktop, ChatGPT, or Cursor, with no server code to write. It covers most major US brokers (for example Alpaca, Public, Schwab, Tradier, and Webull), with some, like Robinhood, available read-only. Access is read-only by default; your broker login stays at the broker; and where the broker supports execution, trades happen only with your approval. From a connected account you can ask an assistant about a specific stock, theme, or your whole portfolio against what you really hold. For a deeper comparison of the field, see the best MCP connectors for brokerages roundup.

Using a brokerage MCP connector follows a short, consistent path, whichever connector you choose.

• Pick a connector that fits your broker. Single-broker servers (Alpaca, Tradier) bind to one account; an aggregator-based connector like Walnut reaches many US brokers through SnapTrade.
• Authorize access. Connect your broker and choose the access level. Read-only is the safer default and lets the assistant describe your portfolio without being able to trade.
• Add the connector to your AI client. Self-hosted servers need configuration and a host you secure; a hosted connector like Walnut needs no server to run.
• Ask in plain language. Once connected, ask Claude or ChatGPT about your real holdings: what you own, how it is doing, where you are concentrated.
• Keep trade access deliberate. If a connector can place orders, keep it read-only until you intend otherwise, and make sure every trade requires your approval.

To see how to connect a broker to an assistant end to end, read how to connect your brokerage to an AI assistant, and to check which brokers already work this way, see which brokers have an AI assistant. For the bigger picture of what these tools are, see what an AI investing app is.

An MCP connector is the bridge that exposes a specific data source or tool to an AI assistant through the Model Context Protocol, Anthropic's open standard for connecting assistants to the outside world. The connector advertises tools the assistant calls, and it translates between the assistant and the underlying service, so the assistant reasons over real, live data. In investing, that is what turns a general AI into a portfolio-aware one: a brokerage MCP connector gives Claude or ChatGPT secure access to your real account, so you can ask about your actual holdings. Connectors range from read-only to read-and-trade, and the safer ones default to read-only, keep your broker login at the broker, and require your approval before any order. Walnut is one such connector, linking most US brokers through SnapTrade, read-only by default. Walnut is not an investment adviser.