Can you let Claude trade your brokerage account — and should you?

Yes, you can: several large retail brokers now let an AI agent such as Claude connect to a live account, and at some of them it can place real orders. Whether you should depends on one thing more than the model you pick: how much the agent is allowed to do without you. This guide covers what is possible today, what goes wrong, and a staged way to try it. It contains no promise of returns, because there is none to make. We build business automations for a living and we do not sell trading bots or signals; the part we know well is how to let software act on something valuable without getting burned.

What changed in 2026?

Brokers opened the door themselves. By Finance Magnates' count, at least ten retail brokers and platform vendors connected AI agents to live client accounts between January and June 2026, including Interactive Brokers, Robinhood, eToro, Public, moomoo and TradeStation. Claude was named in nine of the ten launches, ChatGPT in five, Grok in three, Gemini in two.

Most of these connections use the Model Context Protocol (MCP), an open standard that lets a model call external tools. The broker publishes a server exposing actions such as "get positions", "get quote" or "draft order", and any compatible AI client can call them. Before this, a "Claude trading bot" meant writing your own code against a broker API such as Alpaca or Interactive Brokers, and hosting it yourself. That route still exists and still gives the most control. The new one takes minutes instead of weekends, which is exactly why the limits deserve attention.

What can the agent actually do in your account?

It depends on the broker, and the differences are large. Finance Magnates groups the launches in three tiers:

• Read-only. The agent sees balances, positions and history. It can analyse and explain, and it cannot trade.
• Human-approved. The agent drafts an order and you sign it off. At Interactive Brokers, according to StockBrokers.com's hands-on review, the order is submitted by you from the broker's own platform, outside the AI session.
• Autonomous inside a fence. The agent trades alone, but only in a dedicated sub-account. Robinhood confines agent trading to a separate agent account; eToro uses funded sub-accounts.

Two details from those reviews are worth knowing before you connect anything. First, in every launch Finance Magnates looked at, the agent cannot deposit, withdraw or move client money. Second, the defaults are not uniform. StockBrokers.com found one broker's server defaulting to paper trading and another's defaulting to live production trading, and it notes that a confirmation step at one broker proves the order is accurate, not that a human approved it. Read the defaults of the specific server you install. Do not assume.

Will it make money?

No honest source can tell you that, and we won't either. What the evidence so far shows is narrower and more useful.

In June 2026, researchers at Elm Wealth ran a "Crystal Ball Challenge", reported by Fortune: across roughly 200 sessions, models were given historical newspaper front pages containing market-moving news and asked to trade on them. Claude beat human players in 76% of sessions, ChatGPT in 63%. The same research found the models generally took too much risk for the context, with leveraged positions of 7x to 12x in stocks. A strategy that is right most of the time and sized like that can still end at zero.

The second weakness is tone. In StockBrokers.com's testing, Claude described an extra week on an options spread as free when the short leg still carried value, which made the later expiry more expensive. The reviewer's point applies to every model: flawed reasoning arrives in the same confident voice as sound reasoning. You will not hear the difference. You have to check.

So the realistic value of an agent today is the unglamorous work: pulling your positions into one view, checking an order against rules you wrote, journaling every trade with the reasoning behind it, flagging when your exposure drifts. It is a fast analyst that needs supervision, and it has no edge to sell you.

How do you try it without getting hurt?

The same way we put any automation in front of something that matters: in stages, with evidence at each step.

1. Read-only first. Connect the agent with no trading permission. Ask it to describe your portfolio and compare its answers with the broker's own screens. If it misreads data here, stop.
2. Paper trade for weeks. Use the broker's paper or sandbox mode. Keep a log of every order the agent wanted to place, the reason it gave, and what happened next. This is shadow mode: real inputs, no real consequences.
3. Real orders, each one approved by you. Keep the approval outside the chat where the broker allows it. Read the order ticket itself; the agent's summary of the ticket is not the ticket.
4. Hard limits at the broker. Maximum order size, a symbol whitelist, a capped sub-account. Set them where the agent cannot change them. A rule written in a prompt is a request, and a rule enforced by the broker is a limit.
5. Autonomy last, small, and revocable. If you ever let it act alone, do it in a fenced sub-account holding only money you can afford to lose, and know how to revoke its access in one step.

If you write your own bot against a broker API, the same order applies, plus two engineering habits: log every decision with its inputs, and build a kill switch before you build a strategy.

What about the rules?

According to Finance Magnates, no regulator has yet written rules specifically for AI agents trading retail accounts; existing frameworks from the SEC, FINRA, ESMA and others still apply, and the UK's FCA has a review in progress. In practice that means you remain responsible for every order placed in your account, whoever or whatever typed it. Public's own documentation, as quoted by StockBrokers.com, puts it plainly: executed transactions cannot be reversed.

One boundary is easy to cross without noticing. Automating your own account is one thing. Running money for other people, or selling personalised buy and sell recommendations, is a regulated activity in most jurisdictions and needs a licence. A bot does not change that. If that is your plan, talk to a lawyer before you talk to a developer.

Where do we fit in?

Not in your personal trading account. We build automations for companies, on the client's own cloud, and the discipline above is the one we use when software touches invoices, payments or customer data: shadow mode, then human approval, then autonomy only where the evidence supports it. If you run a business in finance and want to know which of your processes could safely be handed to an agent, the free AI Audit maps that in about five minutes.

This article is general information, not investment, legal or tax advice. Running Dolphins is not a broker or an investment adviser, does not recommend any broker, security or strategy, and has no commercial relationship with the companies mentioned. Broker features and defaults change: check the current documentation before connecting anything to an account that holds real money. Trading involves the risk of losing money.