AI Trading Agent: How to Trade Crypto With Claude and MCP
AI assistants such as Claude and ChatGPT can now connect to trading accounts, read them and place orders for you. Before you hand one your own exchange account or a prop firm challenge (a paid trading test run by a firm that funds the traders who pass), it helps to know how that connection works and what the assistant adds.
Whichever model you pick, the result rests on the access you give it and the strategy you write down for it.
Key takeaways
- An AI trading agent is an AI assistant, such as Claude or ChatGPT, that places trades through a connection you allow and follows a strategy you give it.
- The connection often runs over MCP (Model Context Protocol), and a service that signs in by token works in Claude but cannot be added through ChatGPT's custom-server form.
- The model brings no strategy of its own and cannot see where price is going, so it trades only as well as the written rules you give it.
- Every order an agent places counts as yours, so the account's loss limits and its rules on automation apply to each one.
What is an AI trading agent?
An AI trading agent is an AI assistant that can act on a trading account, reading it and placing orders, instead of only answering questions in a chat. The assistant runs on a large language model (the software behind Claude and ChatGPT, which reads instructions and writes replies in plain language).
Once the assistant is connected to an account, it works much like a clerk who has been handed a key to that account and a page of written instructions. It can act within what the key allows, and it knows your plan only as far as the page spells it out.
An agent and a trading bot
A trading bot and an agent can both place orders for you, but they decide what to do in different ways. A bot runs fixed rules written in code, such as "buy when the price closes above this level". It sends each order through the exchange's API (the connection an exchange offers to software), and the guide to crypto API trading walks through that setup step by step.
An agent reads its instructions in plain language and chooses an action each time it runs. That makes it easier to set up, because you describe the plan in words rather than code. It also makes it less predictable, because loose wording leaves room for an action you did not intend.
What an agent does not bring
An agent brings no strategy of its own, so it can follow only the page you give it. A language model cannot see where price is going, and even with clear rules it can misread one or pick the wrong action.
It also does not watch the market between runs, because an assistant in a chat window waits for your next message unless you set it up to run on a schedule. That is why an order that protects a position, such as a stop loss, belongs on the account itself.
How an AI agent connects to an account through MCP
Whether an assistant can trade an account depends on the connection offered by the exchange, broker or prop firm that holds it. That connection is often MCP (Model Context Protocol, an open standard for linking AI assistants to outside services). Anthropic introduced it in November 2024 so that a service could build one connection for every assistant instead of a custom one for each.
What the server and the token decide
The company that holds the account runs an MCP server, which lists the actions an assistant may take there, such as reading the balance or placing an order. The assistant can call nothing outside that list, so no agent can do more than the list allows.
Within that list, a token or a sign-in decides how far one assistant may go. A token is a long code tied to your account, and its permissions say whether the assistant may only read the account or may also trade. Anyone who holds it gets the same access, which is why you keep it private and why a well-built service lets you cancel it at any time.
How Claude and ChatGPT sign in
Claude and ChatGPT can both add an MCP server, but they sign in to it differently, and that decides which services each one can use. In Claude, you add the server as a custom connector with its address. If the service signs in with a token, you paste the token under request headers, and Claude sends it with every request.
In ChatGPT, you add a custom MCP server on the web, if your account's settings allow it. ChatGPT signs in through OAuth (you log in on the service's own page and approve access) or with no sign-in at all. Its form has no field for a token, so a service that connects only by token cannot be added there. In ChatGPT, an agent can therefore trade only an account whose company offers an MCP server with OAuth sign-in or none.
How to set up an AI trading agent, step by step
The setup follows the same order whichever assistant you use, once it can sign in to the account's server. It moves from reading the account to trading it only once the agent has shown that it reads the account correctly.
1. Write the strategy down as rules
An agent is worth running only on a strategy you already trade by hand. The page you give it should say, in plain sentences:
- when you enter a trade, and in which markets
- how much you risk on each trade
- where each stop loss goes (set on the account together with the entry)
- how much it may lose in a day before it stops trading
- what it reports to you after each order
If you cannot fill in those lines yet, the agent has nothing clear to follow.
2. Connect it with read access
Start with a token or a sign-in that lets the assistant read the account but not trade, so you can see how it behaves before it can change anything. Connect only to a server run by the company that holds your account, because a server can carry hidden instructions that push the assistant into actions you did not ask for. Both Anthropic and OpenAI tell users to connect only to servers they trust.
3. Check what it reports
Ask the agent for your balance, your open positions and the room left under the account's loss limits (the most the account may lose in a day and in total). Compare its answers with what the account itself shows. If they differ, the agent is misreading the account and is not ready to place orders on it.
4. Turn on trading with approvals on
Once its reports match, give it trading access. Both assistants ask you to approve an action such as an order before it runs, and both let you skip the approval for a tool you trust.
Anthropic advises choosing "Allow always" for a tool only when you trust it to run unsupervised. Until the agent has placed orders the way your rules describe, keep the approvals on.
5. Know whose rules apply
An order an agent places is your order, so every rule of the account applies to it. A loss limit closes an account whether you or the agent made the losing trade.
On a prop firm challenge, the firm's rules also say whether automation is allowed and whose strategy a bot or agent may run. Firms differ here, and some ban automated orders, so check which prop firms allow bots and AI agents before you build an agent for a challenge.
Running an AI trading agent at HyroTrader
We run HyroTrader, a crypto prop firm, so this is how an agent connection looks in practice at one firm.
On our challenges you trade USDT perpetual futures (futures contracts with no expiry date, settled in USDT) on a simulated (demo) account that runs on live market data, from the challenge through the funded stage. The payouts to funded traders are real money, at an 80 percent profit split.
Bots and AI agents are allowed as long as the strategy is your own, and an AI assistant connects through Hyro, our own trading platform. Because Hyro is ours, it needs no exchange account, and it fills orders against live Bybit prices with simulated fees and slippage (the gap between the price you expected and the price you got).
On Bybit or BloFin, a bot trades your own demo account at that exchange, over the exchange's own API, and our overview of the platforms a challenge can run on compares every route.
How the connection works
You turn on two-factor authentication (a second login check, usually a code from an app) and then create an agent token under Settings, then Agent tokens. The token can read the account only or trade as well, and you can cancel it at any time.
You then add our MCP server and the token to your assistant's MCP settings. From there the assistant can check your challenges, results and trade history, and it can place trades if the token allows it.
Our help pages name Claude and any other assistant that supports MCP, and the connection uses a token, so the assistant has to be able to send one. ChatGPT's form for a custom server has no place for a token, so with today's setup the route runs through Claude or another assistant that accepts one, not through ChatGPT.
Our help pages do not yet carry a full setup guide, so for the server details you can ask our support team by email, live chat or WhatsApp.
The limits on a $10,000 One-Step challenge
An agent trades inside the same limits you do, so the numbers it has to respect are the account's own. All of our loss limits are static. Each one is a fixed dollar amount set from the starting balance, so the room you start with is the room you keep.
One-Step is a challenge with a single test phase. On a $10,000 account its entry payment is a $129 Refundable Challenge Deposit, which comes back with your first payout and is not returned if you fail.
Rule | $10,000 One-Step | How it works |
|---|---|---|
Daily loss limit | 4%, or $400 | Equity (balance plus open positions at their current value) may fall at most $400 below its level at 00:00 UTC each day |
Maximum loss | 8%, or $800 | The account closes if equity falls to $9,200 |
Profit target | 10%, or $1,000 | Only closed trades count toward it |
Minimum trading days | 3 | They do not need to be consecutive |
Profit Distribution Rule | 40% of the target, or $400 | No single day counts for more than this toward the target |
The daily limit counts open positions at their current value, so an agent that tracks only closed trades can cross it without noticing. The Profit Distribution Rule never closes the account. Profit above that share stays in the account and does not count toward the target, so you keep trading to make up the difference.
Once you are funded, the margin your open positions tie up may not exceed 25 percent of the starting balance ($2,500), and their total value may not exceed twice the starting balance ($20,000).
What the agent must not do
The strategy has to be your own, whoever places the orders. A bought or third-party system that runs someone else's strategy is banned, and so are copy trading, account mirroring and signal services (services that send trade ideas for you to copy).
Exploits are banned as well, including tricks that profit from delays or price gaps in the simulation and very fast order entry aimed at it. So is the Martingale method (raising the position size after a loss to win it back). A setup that breaks these rules can lead to an account review, restrictions or disqualification.
Is an AI trading agent worth setting up?
An agent is worth setting up if you already trade a strategy by hand and want its rules applied the same way on every order. If you are still looking for that strategy, the agent cannot supply it, so that work comes first.
If you have that strategy, start a HyroTrader challenge on the Hyro platform and connect Claude, or another assistant that accepts a token, with read access first.
Note: the price and rule figures above are our published values as of November 2026. The trading rules page has the official, current version of every limit, so check it before you set any number in an agent's instructions.
