Can Claude trade stocks?

The short answer

By itself, no. Claude is a language model, not a broker. Out of the box it has no account, no market data, no order pipe, and no way to move a single share. It can reason about markets and explain a strategy, but it cannot touch your money. Claude can only place a trade when you deliberately wire it to a brokerage through a connector — a tool interface such as an MCP server — and grant it permission to act. Even then, what it does next depends entirely on the instructions and guardrails you give it. An agent with a broker connection but no disciplined rules is far more likely to lose money than make it.

How an AI agent actually places a trade

Modern models can do more than chat — they can call tools. A tool is a function the model is allowed to invoke: "get a quote for NVDA," "list my positions," "submit a buy order." The connector standard that exposes these tools to the model is commonly called MCP (Model Context Protocol). When a brokerage publishes an MCP connector — Robinhood's Agentic Trading is one example — an agent can be authorized to call those broker tools on your behalf.

In practice the loop looks like this:

• Read. The agent calls data tools to fetch quotes, recent price history, and your current positions and buying power.
• Decide. It applies some logic — either improvised reasoning, or a fixed set of rules — to determine whether to act.
• Submit. If the logic says trade, it calls the broker's order tool with a symbol, side, quantity, and order type. The broker, not the model, executes and clears the order.
• Confirm. It reads back the fill, updates its view of the account, and waits for the next scheduled run.

The model never holds your cash or shares; the brokerage does. The connector simply lets the model ask the broker to do things, the same way a mobile app does — except the "finger on the button" is software you scheduled. For a concrete walkthrough of this Claude-plus-broker setup, see our guide on Claude and Robinhood agentic trading.

Improvising agent vs. fixed rules engine

This is the distinction that matters most, and it is where most of the risk lives. There are two very different ways to let an agent trade, and they are not equally safe.

An improvising agent treats every run as a new judgment call. That sounds powerful, but trading rewards consistency and punishes drift — and a model that decides differently depending on how a prompt is worded is the opposite of consistent. A rules engine flips the arrangement: the strategy is fixed, testable code, and the agent's only job is to run it on schedule and call the broker. The model becomes the hands and the timer, not the trader. We go deeper on this trade-off in self-tuning vs. static trading bots.

What has to be true for it to be safe

Whichever approach you choose, a few things should be non-negotiable before an agent is allowed near a live account:

• Hard spend and size limits. A cap on how much capital a single trade — and the account overall — can put at risk, enforced in code, not just hoped for in a prompt.
• Approvals or a kill switch. A way to require confirmation before orders, and a single switch that halts all activity immediately.
• Credentials that stay local. Your broker keys and capital live on your own machine and in your own account — never handed to a third party or a hosted "trade for you" service.
• Risk circuit-breakers. Automatic rules that reduce size or stand down to cash after a drawdown, so one bad stretch can't compound unchecked.
• A tested edge — not optimism. Some evidence the strategy works before you risk money, ideally a cold backtest the agent can't peek at while trading.

None of these make trading safe in absolute terms. They make it bounded. The market can still go against you, stops can still gap through in a fast move, and leveraged products can lose value with brutal speed — see how leveraged ETFs decay for why.

Where Coil fits

Coil is built for exactly the second column of that table. It is not an AI that improvises trades, and it is not a managed account or a signal service. It is a rules-based, long-only engine that scans the whole market — the S&P 500, the Nasdaq-100, and a Macro book (bonds, income, gold and metals, commodities) — scores every name, and buys leaders at real entries. You install and run it yourself, with your own broker credentials and your own capital. Claude (running on a schedule, via Claude Code) and a broker connector like Robinhood do the executing; the strategy itself is fixed, tested code, not a prompt.

That engine ships with a research-backtest harness — a fresh process per market window, so the strategy can't accidentally learn from data it shouldn't see — plus a safety self-test suite (account whitelist, position caps, drawdown ladder, kill switch) you can run yourself. Both exist because they matter: we publish the honest rider right next to the headline, and we would rather disclose a lean stretch than quote inflated results.

On entries and exits, Coil is long-only. It buys leaders scored READY or SETUP — at support, at a real entry-window — and refuses the two ways people lose: it never buys a FALLING name (a falling knife) and never CHASEs an extended one pinned at its highs. It accelerates the strongest leaders with long leveraged vehicles (for example NVDA→NVDL, QQQ→TQQQ) at reduced notional, scales out on a ladder, and trails a stop. When the pool is thin it raises cash and rotates defensively into the Macro book rather than shorting. Sizing is conviction-scaled and clamped by an immutable safety core: account whitelist, position caps, a drawdown ladder, kill switches — and it ships disarmed.

In that research backtest, the leadership-rotation backbone Coil's scoring is built on compounded +638% versus SPY's +282%, with a shallower worst drawdown (−23% vs −32%) and a positive result in 9 of 10 years (worst −1%, in 2018). Read the honest rider with it: through the end of 2025 it ran roughly even with SPY at about one-third less drawdown — the outperformance concentrates in leadership regimes, where the strongest years land (2025 +51%, 2026 H1 +86%). On no-setup days the account sits in cash, where the broker's own variable sweep may pay yield (for example, Robinhood Gold advertised ~3.35% APY in early 2026 — the broker's rate, variable, not paid by Coil, and not risk-free).

Want to compare letting an agent improvise against running a tested engine, or weigh Coil against other tooling? Start with Coil vs. trading bots and signal services, or read the full mechanics on how Coil works.