Until recently an AI agent for trading meant something you assembled yourself. That changed when brokers began shipping agents inside their own apps. On 29 September 2026 Robinhood launched Agents in its app, with a marketplace of paid data apps. A trader now has two routes, and they differ in ways worth understanding before you pick.
The agent inside the broker’s app
This agent lives where your account lives. Setup is short because the broker has done it. The data it works with comes from the broker and from the apps in the broker’s marketplace. Billing for those apps runs through the broker. The launch is covered in Robinhood Agents, explained.
The appeal is convenience. Everything is in one place under one login. The agent is close to the account it may act on, and the broker’s own approvals sit around it.
The agent you run yourself
The other route is an AI app you already use, with connectors you choose. An MCP client is given a server’s URL in its connector settings, and it can hold more than one server. You pick which market data it reads. You pick whether it has an action side at all.
Robinhood supports this route too. It runs its own trading MCP server for agents built outside its app. So an outside agent can read from one server and act through the broker’s, which is the setup in pairing a market-data connector with your broker.
Who picks the data
This is the main difference. Inside a broker’s app, you choose from what the marketplace offers. In your own AI app, you choose from any server that speaks the protocol. Neither is better in the abstract. One trades choice for ease. The other asks you to do the choosing, and to judge each server on its merits.
When you run the agent yourself, the standing instructions are yours to write. You can tell it to quote the age of every number, to name its tools and to stop short of advice. You can read every tool call it makes. That control is the reason some traders take the longer route. What to do with it is in what to log to check an agent’s claims.
Read and act, together or apart
An agent you run can be read-only if you want. Connect a data server and nothing else, and it cannot place a trade. Add the broker’s server later if you decide you want that. Each step is a separate decision, and you can stop at any of them.
The same questions apply to both
Wherever the agent lives, ask it the same things. Where did that number come from. How old is it. What could you not read. An agent that answers those plainly is worth using. One that cannot is a risk in either home.
Where NoVo sits
NoVo is on the second route. The MCP & API is a server any MCP client can connect to, including Claude, ChatGPT and Cursor. It only reads. It has no broker link and places no orders, so it is the read side in either arrangement.