What This Lets You Do
On May 27, 2026 Robinhood opened its brokerage to AI agents over the Model Context Protocol. Once you connect Claude, it can read your portfolio, check buying power, pull live quotes, and place real equity orders from a sandboxed sub-account you fund and control. No API code, no Python. Just a connector and your approvals.
This is the first time a mainstream US retail broker has let an outside AI agent touch a real account. Robinhood put it in beta across all 27 million customers. The setup below works from Claude Desktop and from Claude Code, and the whole point is that you stay in control: the agent can only spend what is inside the wallet you give it, and you decide whether it asks before every trade.
July 17, 2026 update: Robinhood extended Agentic Trading to crypto in early July, so eligible US traders can now let a connected agent buy and sell digital assets 24/7 — not just equities and options. The account model is unchanged: a sandboxed sub-account, a spend limit you set, and a push notification to your phone on every fill. Robinhood says more than 70,000 agentic accounts have been opened since the May launch, and CEO Vlad Tenev told CNBC on July 2 he expects agents to reach “the capability of humans” at trading. Regulators are watching: on June 23, House Financial Services Democrats Bill Foster and Brad Sherman sent the SEC 13 questions on agentic trading with a July 31 deadline, asking whether AI agents should register as brokers and whether calling them a “third-party tool” strips retail investors of protections. Nothing in that changes the setup below, but it is a reason to fund the account small and keep approvals on “ask every time” until you have watched it trade a full weekend. For how this compares to other platforms racing in, see the Coinbase for Agents setup and OKX OnchainOS agent trading guides.
How Robinhood Agentic Trading Works
Robinhood built MCP servers that expose its trading platform as a set of tools any compatible AI client can call. If you have read our MCP servers for AI trading guide, this is the same pattern as Alpaca or Gemini, with one difference that matters: it runs against a real, regulated retail brokerage that 27 million people already use.
The account model is what makes it safe to try:
| Piece | How it works |
|---|---|
| Account | A separate Agentic investing account, distinct from your main Robinhood brokerage |
| Funding | A sandboxed wallet — the agent can only spend what is inside it |
| Scope | US equities, options, and crypto (crypto added early July 2026); futures and prediction markets on the roadmap |
| Permissions | Allow all, ask every time, or deny specific actions |
| Visibility | Every agent trade streams into your regular Robinhood app as a notification |
Because the Agentic account is walled off from your main holdings, a misbehaving prompt cannot drain your long-term portfolio. The worst case is the balance you chose to fund.
Prerequisites
- A Robinhood account with agentic trading enabled (rolling out to all 27M customers through the beta)
- Claude Desktop or Claude Code installed
- A small amount you are willing to let an agent trade — start with money you would not miss
- A few minutes to read the permission prompts instead of clicking through them
Step 1: Open an Agentic Account in the Robinhood App
Inside the Robinhood app, create a dedicated Agentic investing account. It is separate from your regular brokerage by design. Fund it with only what you want the agent to work with. If you are testing, $100–$500 is plenty to confirm the wiring without putting real money at risk.
Do not skip the “fund it small” step. The sandboxed wallet is the main safety boundary here, and it only protects you to the extent that you keep the balance modest while you learn how the agent behaves.
Step 2: Add the Robinhood Connector to Claude
In Claude Desktop:
- Open Settings → Connectors.
- Click Add Custom Connector.
- Paste the Robinhood MCP link from the agentic trading section of the Robinhood app.
- Claude redirects you to your browser and the Robinhood app to authenticate.
- Approve the connection. Once that is done, Claude has access to your Agentic account.
In Claude Code, you add the same remote MCP server through the connector flow instead of editing ~/.claude/.mcp.json by hand. Robinhood’s server uses OAuth, so the browser handshake replaces a static API key. The June 2026 Claude Code release added the remote MCP connector specifically for servers like this one.
After authenticating, ask Claude something read-only first to confirm the link is live:
What is the current value and buying power of my Robinhood agentic account?If it returns your account details, the connection works.
Step 3: Set Your Permission Level
When Claude tries to take an action through the Robinhood connection, it asks for permission first. You get three options:
- Allow all — the agent acts without asking each time. Convenient, riskiest. Only use this once you trust a specific workflow.
- Ask every time — you approve each individual action. This is where you should start.
- Deny — block specific permissions entirely.
Start with ask every time. Reading every order before it fires is the cheapest insurance you have, and it teaches you how the agent reasons about sizing and order types before you hand it more rope.
Step 4: Place Your First Order
With the connector live and permissions set, you can trade in plain language. Try a small, explicit order:
Buy 1 share of NVDA at market in my agentic account.
Show me the order details before you place it.Claude pulls a live quote, builds the order, and shows it to you. When you approve, it routes through Robinhood and the fill streams into your regular Robinhood app as a notification. The MCP tool set covers the common order types, so you can also say:
Place a limit order to buy 2 shares of AAPL at $5 below the current ask,
good for the day.Now that crypto is live, the same plain-language flow works for digital assets:
Buy $25 of BTC in my agentic account. Show me the quote and fee first.Crypto orders route the same way and fill 24/7, so an agent with standing approval can trade overnight — read the summary just as carefully here.
Read the order summary every time before you approve, especially the share count and price. The agent is good but not infallible, and a fat-fingered prompt becomes a fat-fingered order if you wave it through.
A Sensible First Workflow
The strongest early use is not “let it trade on its own.” It is using Claude as a research-plus-execution layer where you stay the decision-maker:
- Ask Claude to summarize your agentic positions and unrealized P&L.
- Have it pull quotes and recent context on two or three names you are watching.
- Ask for a sized order proposal that respects a rule you set, like “no single position above 20% of the account.”
- Review, then approve or adjust.
This keeps the human in the loop on the decision while the agent handles the tedious parts — pulling quotes, doing the math, formatting the order. If you want a fully scheduled version that runs without you clicking, our managed agents scheduled trading guide covers running a Claude agent on a cron schedule.
Limits and Risks Worth Knowing
Equities, options, and crypto — not everything yet. The beta launched with US stocks, added options, and extended to 24/7 crypto trading in early July 2026. Futures and prediction markets are still on the roadmap, not live. Crypto trades run around the clock, so an agent with standing approval can act while you sleep — a good reason to keep permissions tight on the crypto side first.
No paper mode. Unlike Alpaca or Bybit testnet, there is no sandbox simulation — the Agentic account trades real shares with real money. The sandboxed wallet limits how much, but every fill is live. This is why the “fund it small” rule matters more here than on exchanges with a testnet.
Regulators are watching this space. Both the SEC and CFTC have warned about AI trading bot fraud. Robinhood is a legitimate, regulated broker, which puts it on the right side of that line, but the warnings are aimed at the behavior of handing an agent unsupervised control of real capital. Keep approvals on until you have watched the agent work.
The agent can be wrong. An LLM placing orders is still an LLM. It can misread a prompt, misjudge size, or act on stale context. Permission prompts exist for exactly this reason. Use them.
How This Compares to Other Agentic Brokers
Robinhood is not the only broker shipping native agent access in 2026. The shape of each differs:
| Broker | Asset coverage | Sandbox | Best for |
|---|---|---|---|
| Robinhood Agentic | US equities, options, crypto | Sandboxed sub-account, no paper mode | Retail traders who already use Robinhood |
| Gemini Agentic Trading | Crypto spot, US regulated | Paper mode available | Crypto traders who want a regulated US venue |
| Alpaca MCP V2 | Stocks, ETFs, options, crypto | Full paper trading | Builders who want testable code-level control |
If you want to compare crypto-side agentic execution, see Gemini Agentic Trading with Claude Code. If you would rather build a programmable stock bot with a real paper-trading sandbox, the Alpaca MCP stock trading bot tutorial is the better starting point.
FAQ
Is my main Robinhood portfolio at risk? No. The Agentic account is separate from your regular brokerage, and the agent can only spend what is inside the sandboxed wallet you fund. Your main holdings are walled off.
Can Claude withdraw money? The connection is for trading inside the Agentic account. Keep permissions on “ask every time” so you see and approve every action the agent attempts.
Does it work with Claude Code or only Claude Desktop? Both. Claude Desktop uses the Connectors UI; Claude Code uses the remote MCP connector added in the June 2026 release. Same Robinhood account, same OAuth handshake.
What does it cost? The agentic feature is part of Robinhood’s platform and rolling out to its 27 million customers in beta. Standard Robinhood trading terms apply to the orders the agent places.
Bottom Line
Robinhood Agentic Trading is the first time you can point Claude at a mainstream retail brokerage and have it place real equity orders, and the sandboxed-sub-account design makes it reasonable to actually try. Open a small Agentic account, connect the MCP connector, keep approvals on “ask every time,” and let the agent prove itself on tiny orders before you give it more. Start there, watch how it reasons, and scale only what earns your trust.
