Run Claude on Robinhood agentic trading: the prompt, the /loop, the record

You want your AI agent to place trades on Robinhood for you. The short, honest answer: an AI agent can run a rules-based trading program on a schedule — reading the market, deciding within fixed rules, and routing orders through your broker — but it should never be a black box you hand your money to. The good versions are software you operate, on your machine, with your broker and your capital. Coil is one of those. Here's how it works and what's realistic.

If your agent will read Coil alongside Robinhood, the agent quickstart has working Python, TypeScript and curl for it: the MCP connector, a free key and pay-per-read.

The short version

“Agentic” means an AI model can act through tools, not only chat: a scheduled Claude session pulls fresh data, checks a fixed rule set and, only if the rules say so, places an order at your broker. The sound version runs a deterministic, pre-written strategy inside hard rules the model cannot override; the dangerous one improvises trades from a vague prompt. Coil is built the first way (basics: what agentic trading is, can Claude trade stocks?).

Two pieces sit on your side: scheduled Claude sessions that read Coil’s board and rulebook over MCP, and a broker connector, built for Robinhood, which officially supports agentic trading through dedicated accounts. Robinhood “does not control, supervise, monitor, recommend, or audit these AI agents,” so the discipline comes from Coil’s served, versioned rulebook (exits first, sizing by rule, a fail-closed feed) while the account limits stay yours. It is long-only, buys only leaders at real entries, and uses leveraged ETFs only at reduced size (how they decay); with nothing qualifying, cash is the position (how it works).

You keep control: your keys never leave your machine, the capital is yours alone, and you set the schedule and can stop it. A human watching is still the strongest safety layer. Budget about an hour from a connected broker to a scheduled dry-run: connect Robinhood’s MCP (the setup guide), add Coil with claude mcp add --transport http coil https://mcp.coil.trade/mcp (the free tier is a delayed board with no key; Coil Live serves the live board), paste the standing prompt below, watch the dry-run, and go live only as a separate, deliberate step. The pairing recipe is the short form.

Keeping it running: the prompt and the /loop schedule

Connected is a one-shot. A Claude session with two MCP servers loaded — Robinhood's for the account, Coil's for the board — does one honest pass when you ask it to: read the regime, read the ranked list, read the names it is serious about, size from your rules, show you the plan. The prompt below encodes that read order. Paste it as the session's standing instruction (a CLAUDE.md in the folder you launch from works well) and keep position size and exit rules in your half of it. Coil publishes scores and states, never sizes, stops or targets.

Before considering any trade:
1. Call coil get_market_regime. If the ladder does not permit names, stop
   and report that the board says stand down. Do not place orders.
2. Call coil get_buy_list and consider ONLY names that appear on it.
3. Call coil get_stock_read on each candidate you are serious about.
4. Size positions from MY rules below, never from Coil - it publishes
   scores and states, not position sizes, stops or targets.
5. Place orders through Robinhood only within the budget I set, and
   show me the plan before you act.

My rules: [your risk limits, max position size, max daily orders]

To turn that one pass into a standing one, use Claude Code's /loop. It is a bundled skill that re-runs a prompt for as long as the session stays open. Give it an interval and the prompt runs on a fixed schedule; leave the interval off and Claude picks the next wait itself, somewhere between a minute and an hour, and prints its reason at the end of each pass. Against the free board there is nothing to gain by looping. It is one market day delayed, so one pass before the open is the whole job. The intraday board, re-scored about every five minutes through the US market day, comes with a Scanner or Coil Live license key, and the same MCP tools upgrade in place. That is the version worth a loop:

/loop 15m re-run the standing instructions: regime first, then the list, then the reads. Compare against the plan you last showed me and act only on a change, inside my rules.

Fifteen minutes is deliberate. Every pass re-sends the whole conversation, so what a loop costs is passes times context, and the context only grows. A five-minute loop across a full session is close to eighty passes. Our own scheduled sessions taught us where the money actually goes, and it was never the tick: it was the session that failed to end. One review session ran for four days and, by itself, cost as much as every other scheduled session put together. So give every run a hard end — a time budget, after which it reports and stops — and make the regime read the first thing each pass does, so a stand-down day produces one line saying so and nothing else. Most passes should end there. The board's first job is to tell the agent when to do nothing.

The mechanics, from the Claude Code documentation: a loop fires only while the session is open and idle, so closing the terminal stops it, and a fire that comes due while Claude is busy runs once when it is free rather than once per missed slot. Loops that run more often than hourly are jittered by up to half their interval, so 15m means roughly every fifteen minutes, not a metronome. A recurring task expires after seven days, firing one last time and deleting itself, which is a feature: a forgotten loop cannot run all quarter. Press Esc to stop a self-paced loop while it waits; a fixed-interval loop is cancelled by asking, in plain words, to cancel it. For a pass that has to run with no terminal open, schedule it outside the session instead — a Claude Desktop scheduled task, or a cron entry that launches claude -p with the same standing instructions — and allow the exact tools it needs ahead of time. A headless run cannot answer a permission prompt; it just sits there until something kills it, and it looks from the outside like an agent that decided to do nothing. The full version of this — the Desktop task, the cron line with the flags a headless run needs, and what a pass should cost — is on its own page: running an always-on trading agent with Claude Code’s /loop.

What running by rule looks like over months is not a claim on this page; it is a page of its own. The live engine record is published every session next to SPY and QQQ, win or lose, rendered from the public API with an append-only commitment log so a bad stretch cannot be quietly edited out later. Read it before you budget a dollar. It is the honest answer to “does this work”, and the honest answer moves.

The honesty layer: the research backbone

The research below comes from a research-backtest harness, and the rules behind it are written down and served as a versioned rulebook. Coil is not an open-ended AI inventing new trades: new tested structure arrives only as a deliberate new version, with a changelog. Account limits (which account, caps, any loss halt) are yours to set in your agent and at your broker. That's the brand — the honest one.

The research backbone (2017–2026 H1)

This is a research backtest — not live results, not client returns, and not a prediction. The leadership-rotation backbone Coil's scoring is built on was replayed point-in-time and survivorship-free (delisted names included), with next-open fills and costs modeled. It was benchmarked against SPY over the same window.

The same study, with its assumptions, is on the honest review of AI trading bots.

We print no backtest performance figures for Coil's strategy; the live record (/results) is the only performance this site shows. Our published studies report their own test results, failures included. Read the honest rider with it: the engine itself is newly live; this is research, not a live track record, and past performance does not predict future results.

The risks, stated plainly

• Leverage cuts both ways. The leveraged ETFs Coil uses to accelerate a leader move roughly 2–3x the underlying before gaps and slippage; they decay on multi-day holds and can lose rapidly, up to total loss. Coil uses them at reduced notional for that reason — but reduced size is not no risk.
• Circuit-breakers reduce damage; they don't remove risk. Coil's breakers tie to your account's high-water mark to cap single-day pain — but stops can gap through on a hard open.
• This is not passive income. No guaranteed return, no autopilot wealth. You're operating risk software, and you can lose money.

If you're weighing this against off-the-shelf bots and paid alerts, the Coil vs trading bots and signal services comparison lays out the differences honestly.

Common questions from operators

Why does my AI agent refuse to place a trade?

Robinhood agentic trading "not working" with Claude is, in almost every case, correct behaviour, not a bug. Work down this list in order — each cause looks identical from the chat window ("it just won't trade"), and each has a different fix:

1. No authorised path to your brokerage. A general-purpose assistant cannot place an order because nothing has connected it to one, and inventing a way to move your money would be exactly the wrong instinct. If you have not added the broker's MCP server to your agent, this is your case — see the setup guide, including the --transport http flag most people miss.
2. The MCP is connected, but the account isn't agentic. Robinhood's place_equity_order requires an account with agentic trading enabled and rejects orders from any other account — so the tools all appear, every read works, and every order is refused. Enable agentic trading on the specific account you named, in the Robinhood app. (Verified against Robinhood's Trading MCP as of July 2026; their docs are authoritative.)
3. Nobody was there to confirm. The documented order flow is review_equity_order → present the estimate and alerts → explicit user confirmation → place_equity_order. A scheduled, unattended run with nobody present to confirm is supposed to place nothing on that path — that is the broker's design working. An unattended routine needs its own deliberate, pre-authorised rules for what it may do (which is exactly what an engine like Coil's arming flow is for), not a bypass of the confirmation.
4. If you are running Coil: it ships deliberately disarmed. Live trading is off by default. The engine scans, scores every name, shows you precisely the plan it would execute — and places nothing — until you complete the arming flow yourself.
5. You asked for a decision, not an execution. "Should I buy NVDA?" asks the model to originate an investment recommendation, and it will decline — that refusal is correct and is not a configuration problem. Rules you wrote, executed by rule, are a different request from advice.

The fix is not prompt-wrangling. An agent that can be talked into placing orders it was never configured to place is a liability, not a feature: the same flexibility that lets you argue it into a trade lets a bad instruction do the same. What you want is a real, authorised broker connection plus a deliberate arming step that you perform knowingly.

For the six symptoms and their fixes, see the Robinhood MCP troubleshooting guide.

Can Claude actually execute trades on Robinhood?

Yes — through a broker connector, not by asking nicely: Coil is built for Claude Code and places orders through Robinhood’s MCP connector (any broker with an equivalent MCP works). Your credentials stay on your own machine and Coil never collects them or sees your balances or positions; the full answer is in can Claude trade stocks?

How do I let the agent place real orders?

Arming is deliberately a separate, several-step act rather than a toggle: you allowlist the single account it is permitted to touch, run an integrity re-pin, pass a green safety self-test, and type an explicit acknowledgment that you can lose money. Only then does it trade — and a single command stops it again. The friction is the point. Anything easier would be easier to do by accident.

Is it safe to let an AI agent trade my account?

Safety here is bounded, not guaranteed, and we would rather say so plainly. Coil restricts itself to one allowlisted account, enforces hard position and exposure caps, runs a drawdown breaker tied to your equity high-water mark, and fails closed — any ambiguity and it does nothing rather than guess. Those bounds reduce and cap damage. They do not remove market risk, and you can still lose money.

Does Robinhood agentic trading work with Claude Code?

Yes. Robinhood’s Trading MCP server is a standard remote MCP endpoint, so any MCP-capable client can use it — Claude Code, Claude Desktop, Cursor, or another host. Claude Code is the better fit for an unattended routine, because it can run on a schedule and execute code between tool calls; Claude Desktop suits a human-in-the-loop session where you approve each step. The exact command and the --transport http flag that makes it work are in the claude mcp add guide; confirm the address on Robinhood’s own agentic-trading page before you paste it.

Can I connect Robinhood agentic trading to Gemini CLI, Codex, or Antigravity?

Yes, in the sense that matters: Robinhood's endpoint is a standard remote MCP server, so this is a config-format question, not a compatibility question. Any client that speaks remote (streamable-HTTP) MCP can connect. Where each one puts it, as of July 2026 — the vendor's own MCP docs are authoritative and these are theirs to change:

The URL is the one in the claude mcp add guide — take it from Robinhood's agentic-trading page and confirm it there before pasting it into any of these clients, because it is theirs to change. One honesty note: Coil is built and tested for Claude Code; other agents can read the same board and the same broker MCP, but we make no promises about how well they run an unattended routine.

What prompt or parameters should I give Claude for agentic trading?

The prompt that matters is not a magic phrase — it is a read order plus your own risk limits, stated explicitly. An agent told only "trade my account profitably" will screen whatever it happened to see last; an agent told to check the market regime first, take candidates only from a ranked list, and size from rules you wrote will behave consistently.

A copy-paste prompt that encodes exactly that discipline is above, with the /loop schedule that keeps it running, and on the Coil × Robinhood pairing page along with the read order it depends on. Two things belong in your half of it and never the publisher's: position size and exit rules. Coil publishes scores and states — never position sizes, stop prices or target prices.

Why can't the agent place a trailing stop on Robinhood?

Because the order type isn't there — and the workaround is better than the order type. As of July 2026, Robinhood's Trading MCP place_equity_order accepts market, limit, stop_market and stop_limit, with a stop_price and good-til-cancelled support. What it does not accept is the trailing variant. Two more constraints worth knowing before you design around it: market and stop orders are regular-hours only (they're rejected if tagged to an extended session), and the overnight session executes limit orders only — so a resting stop cannot even be parked overnight. Robinhood's documentation is authoritative on all of this and it is theirs to change.

The honest reframe: a trailing stop is arithmetic, not an order type. A scheduled agent recomputes its exit level from current structure each session and sends a plain order when the condition trips — which is exactly how Coil is built: deterministic per-session exits, no resting stop at the broker (see trailing stop for the mechanic). One caveat no order type escapes: nothing defends a price the market gapped over, whether the instruction lives at the broker or in your agent.

Can I have it day-trade for me?

By rule, in the tape that calls for it — not as a scalping alert stream. Coil runs two horizons and lets the market pick: above the long-term trend line it swings the leaders, holding for days; below it, it shortens its horizon and trades session leaders intraday by rule, because in a falling tape overnight exposure is the risk. Either way a name has to prove it's ready — score a real entry — before anything is bought. What Coil is not: a scalper, an open-fader, or an intraday alert stream you react to (the Trade Ideas comparison makes that case in full). The free board is one market day delayed — do not wire an unattended intraday loop against delayed data; the live intraday board is the $19/mo Scanner. Your broker's own day-trade limits apply to your account regardless of what your prompt says.

Can Claude Code keep the agent running on a schedule?

Yes. Inside an open session, /loop 15m followed by your prompt re-runs it on a fixed schedule, and it only fires while the session is open and idle, so an unattended run wants a Claude Desktop scheduled task or a cron job instead (the prompt is above). The flags, the limits and what a pass costs are in the /loop trading-agent guide.

Which Robinhood MCP tools does an agent actually call, and can it run on a free account?

Robinhood’s Trading MCP exposes read tools alongside the order tools, so a well-built routine spends most of its calls reading, sizes from get_portfolio, and places orders rarely, through review_equity_order and then place_equity_order. Two names people search for do not exist: get_account (it is get_accounts) and get_orders (it is get_equity_orders); every tool is listed in Robinhood MCP tools: the complete reference.

On our side, Coil’s MCP board has a free tier with no key, wallet or account: a sample of the board (top 3 per book, headline fields), one market day delayed, enough to see how it reads before a swing-timeframe agent pays for the whole thing. A swing routine is a handful of tool calls a few times a day (read the regime, read the ranked list, check positions, maybe place an order); an always-on intraday watcher polling every few minutes is a different and costlier shape, as the /loop section above explains. Whether a given account is eligible is Robinhood’s call, not Coil’s, and the setup guide’s cost section breaks it down.

Pricing and next step

Coil sells two subscriptions. The Scanner is $19/mo or $159/yr: the scored board of 650+ US names, read in your browser. Coil Live is $49/mo or $419/yr: the engine-grade board served over MCP to your own agent, with the served discipline. Both can be cancelled at any time, and checkout runs through Gumroad as merchant of record. Read how it works first if you'd rather see the method before you decide.