How to Connect AI to Robinhood with MCP: A Safer Setup Guide

Bottom line: Robinhood now offers an official Model Context Protocol connection that lets compatible AI clients read account data and use trading tools.[2] The setup can take only a few minutes, as trader Miles Deutscher demonstrated in a widely shared X thread.[1] The connection is easy. Deciding what the agent is allowed to do is the serious part.

Robinhood’s own documentation says an attached agent can read data across all of a user’s Robinhood accounts, including account numbers, balances, positions, transactions, watchlists and scans.[2] It can place trades only in a separate Robinhood Agentic account.[2] That separation limits where the agent can trade, but it does not make automated orders safe or profitable.

This guide shows how to connect Claude, ChatGPT, Codex or another MCP-compatible client to Robinhood, run a read-only test and add a human approval gate before any live order.

Important: Agentic Trading can place real orders and lose real money.[3] Robinhood says users are responsible for every trade their AI agent places, including trades executed without a confirmation when the user has authorized that behavior.[2][3] Treat the connection as brokerage access, not as a chatbot experiment.

What does connecting AI to Robinhood actually mean?

MCP is an open standard for connecting an AI application to external tools and data.[4] In this case, Robinhood operates a remote MCP server at:

https://agent.robinhood.com/mcp/trading

The AI client connects to that server, and Robinhood handles authentication.[2] You should not paste your Robinhood password, one-time code or access token into a chat prompt. Anthropic’s connector guidance says the normal OAuth flow lets a user sign in to the external service and grant permissions without showing the password to Claude.[5]

Once connected, the model gains tools rather than magical market insight. Robinhood currently documents tools for viewing accounts, portfolios, realized profit and loss, watchlists, market data, equity positions, options, crypto and scanners.[3] The tool list also includes actions that can place or cancel real equity, options and crypto orders.[3]

That distinction matters:

• Data access lets an AI inspect information and produce analysis.
• Action access lets it change account state, including submitting an order.
• A trading strategy is the set of rules that decides whether an action should happen.
• Risk controls determine how much damage a bad instruction, stale input or model error can cause.

MCP supplies the first two capabilities. Robinhood does not claim that the connection supplies a profitable strategy or reliable risk management.[2][3]

What you need before starting

You need a Robinhood primary individual investing account in good standing before opening the additional Agentic account.[2] Robinhood says the Agentic account is a self-directed individual investing account, and its onboarding opens during the MCP authentication process.[2]

You will also need:

1. A desktop device for initial Agentic account setup and authentication. Robinhood says the account-opening flow must be completed on desktop.[2]
2. An AI platform that supports remote MCP connections. Robinhood lists Claude Code, Claude Desktop, ChatGPT, Codex, Codex CLI, Cursor and Grok, while also allowing other MCP-compatible platforms.[2]
3. The exact official endpoint: https://agent.robinhood.com/mcp/trading.[2]
4. A written policy for what the agent may read, propose and execute.
5. A small amount of capital you can afford to lose if you choose to test live trading.

For crypto trading, the Agentic account needs a corresponding Robinhood Crypto account and acceptance of the updated customer agreement and agentic disclosures.[2] Robinhood says agentic crypto trading is unavailable in some states, including New York, and the agent cannot transfer, stake or lend crypto.[2]

Step 1: Write the safety rules before connecting

Do not begin with a prompt such as “buy good stocks” or “trade whatever looks promising.” Deutscher likewise warns that a vague request is a recipe for poor results.[1]

Start with a local Markdown file called ROBINHOOD_AGENT_RULES.md. A conservative first version could look like this:

# Robinhood agent rules

## Operating mode
- Analysis and order preview only.
- Never place, modify, or cancel an order without my approval in the same chat.
- Treat approval as valid for one fully specified order only.
- If any order detail changes, ask again.

## Allowed actions
- Read portfolio, positions, balances, watchlists, quotes, and order history.
- Run an order preview when I explicitly request it.
- Do not trade options.
- Do not trade on margin.
- Do not create recurring or autonomous trading loops.

## Order limits
- Use only the Robinhood Agentic account.
- Maximum order value: [enter your limit].
- Maximum total open exposure: [enter your limit].
- Allowed symbols: [enter an explicit list].
- Allowed order types: limit orders only.
- Never average down automatically.

## Required pre-trade response
Before requesting approval, show:
1. account name;
2. symbol and asset type;
3. buy or sell;
4. quantity and estimated notional value;
5. order type and limit price;
6. current bid, ask, and quote timestamp;
7. estimated remaining buying power;
8. why the order satisfies every rule;
9. any warning or uncertainty returned by the preview tool.

## Approval
Ask: "Approve this exact order?"
Accept only an explicit yes after the full order summary.
Silence, ambiguity, a prior approval, or approval for another order means no.

The numbers and asset list must come from the user, not from the model. A percentage that is sensible for one portfolio can be reckless for another. If you do not know what limit to use, remain in analysis-only mode.

Step 2: Connect Robinhood to Claude

Robinhood publishes slightly different paths for Claude Desktop and Claude Code.[2]

Claude Desktop or Claude web

For an individual Claude plan:

1. Open Customize and select Connectors.
2. Select +, then Add custom connector.
3. Paste https://agent.robinhood.com/mcp/trading.
4. Add the connector.
5. Select Connect and complete Robinhood’s authentication flow.

Anthropic says custom remote connectors are available across its individual plans, although Free users are limited to one custom connector.[5] Team and Enterprise workspaces require an owner to add the connector before members can authenticate it.[5]

Robinhood’s help page describes the current Claude Desktop route as Settings → Connectors → Add custom connector.[2] Menu labels can differ between the desktop app and the web interface, so use the official connector page if the labels have moved.

Claude Code

Run Robinhood’s documented command:

claude mcp add robinhood-trading --transport http https://agent.robinhood.com/mcp/trading

Then enter /mcp, select robinhood-trading and authenticate.[2]

After connecting, attach or paste your rules file into the project or conversation. Tell Claude that the rules are mandatory and that no write tool may be used without a fresh confirmation.

Step 3: Connect Robinhood to ChatGPT

Robinhood’s current instructions for ChatGPT are:[2]

1. Open Settings.
2. Select Security & login.
3. Turn on Developer Mode.
4. Open Plugins and select +.
5. Add https://agent.robinhood.com/mcp/trading.
6. Authenticate with Robinhood and complete Agentic account onboarding if prompted.

OpenAI has renamed and reorganized connectors, apps and plugins several times. Its current help material says availability and controls can vary by plan, region and workspace configuration.[6] If the Robinhood path is not visible, check that the ChatGPT client is current and that a workspace administrator has not disabled custom MCP connections.

Before enabling the plugin in a trading conversation, review its action permissions. OpenAI’s current permission model can be configured to ask before meaningful external effects, including financial transactions.[6] For a brokerage connection, use the strictest available approval setting rather than “allow all actions.”

Step 4: Connect with Codex, Cursor or Grok

Robinhood also publishes these setup paths:[2]

Codex desktop

1. Open Settings → MCP servers.
2. Choose Streamable HTTP.
3. Add the official endpoint.

Codex CLI

codex mcp add robinhood-trading --url https://agent.robinhood.com/mcp/trading

Enter /mcp, then select robinhood-trading.[2]

Cursor

1. Give the agent the official MCP endpoint.
2. Open Settings → Cursor Settings.
3. Select Tools & MCPs → Connect.[2]

Grok

1. Start a chat and select +.
2. Choose Add connector → Custom.
3. Add the official endpoint.[2]

Whichever client you use, confirm that the hostname is exactly agent.robinhood.com.[2] Do not use a copied endpoint on a lookalike domain, an unsolicited server sent through direct message or a community proxy that asks for Robinhood credentials.[5]

Step 5: Authenticate and open the Agentic account

When you connect, Robinhood should open its own authentication and onboarding flow. Its help center says users are prompted to open an Agentic account during AI-agent authentication.[2]

Check the browser address before signing in. The login and consent flow should be on an authentic Robinhood domain.[2] Review the account and permissions presented on screen, then complete the Agentic account setup.

Robinhood makes an important permission distinction: the agent receives read access to all Robinhood accounts but can place trades only in the dedicated Agentic account.[2] If you are uncomfortable exposing account numbers, positions, balances and transaction history to the chosen AI provider, stop here. A separate trading account does not separate the read access.

MCP security guidance recommends least-privilege scopes because broad tokens increase the damage from compromise and make audit trails harder to interpret.[7] If the consent screen allows narrower permissions, begin with the minimum needed for account discovery and analysis.

Step 6: Run a read-only smoke test

Do not make the first tool call a trade. Start a new conversation with the connector enabled and ask:

Use Robinhood read-only tools. List the available accounts by type and show
buying power for the Agentic account. Do not reveal full account numbers.
Do not preview, place, modify, or cancel any order.

Next, ask for a portfolio summary:

Read my current Agentic account positions and buying power. Explain which
Robinhood tools you used and the timestamp of the data. Take no action.

Compare the result with the Robinhood app. Check the account, positions, buying power and timestamp. A correct response proves that the connector can read data; it does not prove that the model will always select the correct account or interpret instructions consistently, a risk Robinhood includes in its disclosures.[2]

Step 7: Test an order preview without placing it

Robinhood documents separate review or preview tools for equities, options and crypto, alongside the tools that place real orders.[3] Use the preview path first.

Example:

Review, but do not place, a limit order to buy [quantity] of [symbol] at
[limit price] in my Agentic account. Show every pre-trade warning, estimated
notional value, buying power after the order, and the exact tool you used.
Do not call any place-order tool.

Verify that the agent used review_equity_order for an equity or preview_crypto_order for crypto rather than a placement tool.[3] If the client cannot show tool-call details clearly, do not grant it live trading authority.

Step 8: Keep human approval on for live orders

A safe first live test has five parts:

1. Use the dedicated Agentic account.
2. Fund it with a deliberately small amount.
3. Choose a liquid asset and a small, fully specified limit order.
4. Require a preview and inspect every field.
5. Approve that one order only, then verify it in Robinhood’s Activity section.

Robinhood says all agent orders appear in the Agentic account’s Activity section and in Robinhood history.[3] Check the broker record rather than trusting the chat’s success message.

Do not authorize a perpetual instruction such as “run this strategy every hour” during initial testing. Robinhood warns that an agent can trade without confirmation when the user asks it to operate that way, and AI-driven strategies may be difficult to monitor or stop in real time.[2]

The biggest risks are not setup errors

The model can misunderstand a valid instruction

Robinhood explicitly warns that AI agents can make errors, misread instructions, rely on incomplete or outdated information and behave unexpectedly.[2][3] A syntactically correct order can still be the wrong trade.

The connector exposes sensitive financial data

The MCP connection can read account numbers, balances, positions, transactions, watchlists and scans across all Robinhood accounts.[2] Review the AI provider’s data controls and avoid placing account details into shared projects or logs.

Prompt injection can turn data into instructions

A malicious page, tool response or connected data source can contain hidden instructions intended to redirect an agent.[5] Anthropic tells users to connect only trusted servers, inspect tool inputs and outputs and avoid permanent approval for tools they do not trust.[5] Keep unrelated web-browsing tools out of a trading session when possible.

Automation can repeat mistakes quickly

An autonomous loop can repeat an unintended action or keep trading after market conditions change, which is why Robinhood warns that agentic strategies can move quickly and be hard to stop in real time.[2] Add order-count, notional and time limits outside the prose strategy wherever the client supports them.

Options and crypto add another layer of risk

Robinhood currently allows agentic long equity, options and crypto orders.[3] Options can expire worthless and some strategies can lose more than the original investment, while crypto held through Robinhood Crypto is not FDIC insured or SIPC protected.[2][3] A first connectivity test does not need either product.

A practical operating model

The most useful starting point is not a fully autonomous trading bot. It is a three-stage assistant:

1. Research: gather account and market data, then state what is known and uncertain.
2. Proposal: construct one order preview that satisfies the written rules.
3. Approval: wait for the user to approve the exact order before calling a placement tool.

This design keeps the speed advantage of an AI interface while leaving the final decision with a person. It also creates a readable trail: data, reasoning, preview, approval and broker confirmation.

After each session, review the Robinhood Activity page, open orders and current positions. Disconnect the MCP connector if the client behaves unexpectedly. Anthropic says connector permissions can be revoked from Claude’s connector settings or the third-party service’s security settings.[5] Robinhood’s troubleshooting guide also recommends disconnecting and reconnecting the Trading MCP when connection problems occur.[2]

What Miles Deutscher’s post gets right

Deutscher’s thread correctly identifies the low-friction setup: choose an MCP-capable AI client, add Robinhood’s official endpoint, authenticate, open the Agentic account and define a ruleset.[1] It also warns against vague prompts, recommends supervised use and suggests requiring confirmation before trades.[1]

The part worth emphasizing is that a quick technical setup does not justify quick autonomy. The value of this connection is controlled execution and easier access to portfolio data.[2][3]

“A 24/7 trading desk” is Deutscher’s useful metaphor for availability, but Robinhood does not present the agent as proof of a profitable edge.[1][2]

Robinhood AI connection FAQ

What is the official Robinhood Trading MCP URL?

The official endpoint published by Robinhood is https://agent.robinhood.com/mcp/trading.[2] Check the hostname before adding it to any AI client.

Can the AI see my other Robinhood accounts?

Yes. Robinhood says the connected agent has read access to all Robinhood accounts, including account numbers, balances, positions and transaction history. It can place trades only in the Agentic account.[2]

Can an AI trade without asking me?

Yes, if you authorize it to act without confirmation. Robinhood warns that an agent can place trades without direct approval when the user has instructed it to do so.[2][3] Keep per-order confirmation enabled while testing.

Does Robinhood Agentic Trading guarantee better returns?

No. MCP provides access to data and order tools, not a profitable strategy. Robinhood does not guarantee the accuracy or suitability of agent output and says the user assumes the trading risk.[2][3]

Can an agent trade stocks, options and crypto?

Robinhood currently documents long equity, options and crypto orders for Agentic accounts.[3] Eligibility and product restrictions still apply, and agentic crypto trading is unavailable in some states.[2]

How do I stop the connection?

Disable or remove the Robinhood connector in the AI client's settings, then review open orders and positions directly in Robinhood.[5] If you suspect unauthorized access, use Robinhood's security and support channels rather than relying on the chat session.

Disclosure: This article is educational and does not constitute financial, investment, legal or tax advice. Agentic Trading can execute real orders and may result in the loss of your entire investment.[2][3] Product availability, menus and permissions can change. Verify current instructions and disclosures on Robinhood's official website before connecting an AI client.

Sources

[1] https://x.com/milesdeutscher/status/2094794129498833248 — Miles Deutscher: Robinhood AI agent setup thread [2] https://robinhood.com/us/en/support/articles/agentic-trading-overview — Robinhood: Agentic Trading overview [3] https://robinhood.com/us/en/support/articles/trading-with-your-agent — Robinhood: Trading with your agent [4] https://www.anthropic.com/news/model-context-protocol — Anthropic: Model Context Protocol [5] https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp — Anthropic: Custom connectors using remote MCP [6] https://help.openai.com/en/articles/11487775-connectors-in-chatgpt — OpenAI: Apps and plugins in ChatGPT [7] https://modelcontextprotocol.io/specification/latest/basic/security_best_practices — Model Context Protocol: Security best practices