Most AI apps that support the Model Context Protocol will hold more than one server at a time. That one fact changes what an assistant can do for a trader, because the server that reads the market and the server that holds your account do not have to be the same company.
The shape of it
One server answers questions about structure: where dealers are positioned, where the walls and the flip sit, what funding looks like across venues. The other is your broker's, and it knows your positions and can stage orders. The assistant reads from the first and acts through the second, in a single conversation.
Neither server needs to know the other exists. The assistant is the only thing that sees both.
Connecting it
Dr. NoVo is at https://novo-options.trade/mcp. Robinhood's trading server is at https://agent.robinhood.com/mcp/trading.
In Claude Desktop, add each one under Settings, then Connectors, then Add custom connector. In Claude Code it is one line per server: claude mcp add novo --transport http https://novo-options.trade/mcp. Cursor, ChatGPT and other MCP-capable apps take the same two URLs through their own connector settings.
Twenty of the thirty-three tools answer with no key at all. The rest take an API key, which can go in a header or in the URL itself, so a client that cannot send custom headers still works.
Why the split is the safe part
It is natural to worry about an assistant that can both read and trade. The pairing actually separates the two cleanly. The data server has no access to your account and cannot place anything. Execution stays behind your broker's own controls, and on Robinhood that means a dedicated Agentic account and the trade approvals you configured.
Two habits keep it honest. Ask which tool a figure came from before you act on it; a grounded answer can name it. And treat anything the assistant proposes as a proposal. The approval step exists so that a misread never becomes an order on its own.
None of this is advice, and automated trading carries the same risk of loss as trading by hand.