Agentic Trading on Robinhood: Connect Claude via MCP, Safely

Steps

1. Understand what you're enabling. The MCP server exposes your agentic account to whatever agent you connect: market data, positions, and order placement for equities and single-leg options (crypto began rolling out in July 2026). The account is deliberately walled off — the agent can only touch funds you move into it. That isolation is the single most important safety property of the whole design, so never wish it away.
2. Check eligibility and open the agentic account. You need a US individual Robinhood account in good standing; the feature is free. In the Robinhood app, find the Agentic tab, accept the (real, worth reading) terms, and open the dedicated agentic account. Fund it with a small amount you can genuinely afford to lose — this is also your test budget, because there is no paper-trading mode.
3. Connect your agent over MCP. The Agentic tab shows the MCP connection URL. For Claude Code it's one command (below), then an OAuth flow you must complete on desktop — approve in the browser, verify in the Robinhood mobile app, and paste the redirect URL back to the agent. Gotcha: the final localhost page may show a 'can't connect' error; that's expected — copy the full URL from the address bar anyway.
4. Switch every guardrail on before the first prompt. In the agentic account settings: turn ON manual approval for each trade, keep per-trade push notifications on, and note where the instant-disconnect button lives. With manual approval on, the agent proposes and you dispose — which turns 'an LLM trades my money' into 'an LLM drafts orders I confirm'. Leave it that way at least until the novelty wears off, and preferably forever.
5. Start read-only: research before orders. Spend the first sessions on prompts that can't spend money: summarize the account, quote tickers, analyze a position, propose (not place) a trade with reasoning. You're learning the agent's failure modes — stale data, misread instructions, overconfident reasoning — while they're free. Only after the agent has been boringly reliable for a while should you let it place a first, small, manually-approved order.
6. Supervise like it's an intern with your card. Watch the live activity feed and P&L after every session. The agent only works while it's running (Claude Code open on your machine) — it is not a 24/7 bot, and that's a feature: no session, no trades. If anything looks off, disconnect first and investigate second. And keep the funding small: the strategy here is a prompt, and prompts don't have backtests.

The code

Copy these verbatim and swap in your own values — licence ID, symbol, risk and stop settings. Double-check the symbol name matches your broker exactly, or the order will be rejected.

# One-time setup — requires a desktop browser for the OAuth flow.
claude mcp add robinhood-trading --transport http https://agent.robinhood.com/mcp/trading

# Then restart Claude Code and follow the OAuth prompts:
#   1. Approve access in the browser (desktop only)
#   2. Verify in the Robinhood mobile app
#   3. If the final localhost page errors, copy the FULL URL
#      from the address bar and paste it back to Claude.
# Verify the current endpoint in Robinhood's Agentic tab before connecting.

# Prompts that cannot spend money — use these for the first sessions:

"Summarize my agentic account: balance, positions, open orders."

"Quote NVDA and AAPL and compare their last month's range."

"PROPOSE (do not place) a trade idea for this account with your
 reasoning, position size under $100, and where you'd put a stop."

# Only after days of boring reliability, with manual approval ON:
"Place a limit buy for 1 share of VTI at $X. Wait for my approval."