Claude trading bot: can Claude trade crypto for you, safely?

The short answer

Claude, Anthropic's AI assistant, is very good at understanding a trading idea in plain words and turning it into precise rules. Since 2026, it can also place orders on some exchanges through tools built for AI agents. That makes "Claude as a trading bot" possible.

Possible is not the same as safe. On its own, Claude has no loss limits it cannot override, does not watch the market between your messages, and keeps no audited record of why it traded. Use it to draft and explain rules, and let something with hard limits place the orders.

How Claude can connect to an exchange

Claude can use outside tools through the Model Context Protocol (MCP), a standard way for assistants to call software. In March 2026 Kraken released a command-line tool with MCP support, so Claude can read balances and place orders on a Kraken account when set up by a technical user. Other exchanges have released similar agent tools.

The setup typically needs a terminal, an exchange API key and some configuration. Once done, you can ask Claude to buy or sell, and it does.

What Claude does well

• Turning a vague idea into clear rules. "Buy ETH on big dips" becomes "buy $40 of ETH when it falls 6% within 24 hours, at most 2 buys open."
• Asking what you forgot. A good prompt gets Claude to ask about the stop-loss, the budget and the exit.
• Explaining concepts. Drawdown, slippage, stop-limit orders and fee tiers, in plain words.
• Reviewing a strategy. Pointing out that a rule never sells, or that it averages down without a cap.

What Claude does not do on its own

• Enforce limits. If you write "never lose more than $25 a day" in a prompt, that is an instruction, not a limit. Nothing outside the model stops an order that breaks it.
• Watch the market. Claude acts when you message it. A rule like "buy when BTC falls 5% in a day" needs something checking every minute, all day and night.
• Give the same answer every time. Language models can respond differently to the same question. That is fine for writing rules, risky for deciding orders.
• Keep an audit trail. The exchange sees orders, not reasons, unless something records them.
• Test with fees. Claude can describe a backtest but does not, by itself, run your rules over years of exchange history with real fees.

The safer pattern: AI drafts, rules execute

The pattern that keeps the useful part of Claude and removes most of the risk:

1. Use the AI to draft the rules. Plain words in, precise rules out.
2. Confirm the rules yourself. Read them; change what is wrong.
3. Backtest the confirmed rules with real fees and compare with holding the coin.
4. Run them on paper with live prices.
5. Let a rules engine execute them, with a budget, stop-loss, daily loss limit and open-position cap checked before every order.
6. Use the AI again to explain each trade and each skipped trade.

This is how TradeCorgi works. The language model drafts your strategy card and explains what happened. It never places orders.

Claude on its own versus a dedicated agent

• Writing rules from plain words: Claude does this well. A dedicated agent does it too, and asks for any missing limits.
• Watching the market around the clock: Claude does not; it acts when you message it. A dedicated agent checks at most once a minute, all day.
• Enforcing loss limits: with Claude alone, only as instructions in the conversation. A dedicated agent enforces them in code, before every order.
• Testing with real fees: Claude can describe a backtest. A dedicated agent runs one on exchange history, with fees and slippage.
• Explaining each trade: Claude can, when asked. A dedicated agent writes a reason for every trade and every skipped trade, automatically.
• Stopping instantly: with Claude, you revoke the key or close the session. A dedicated agent has a kill switch on every screen.

If you still want to connect Claude directly

• Use a separate, trade-only key with withdrawals off and IP whitelisting on.
• Fund the account with only what you can afford to lose, because the account balance is the only real limit.
• Never leave it unattended expecting it to watch the market.
• Keep your own log of what you asked and what it did.
• Do not use leverage.

A worked example

You ask Claude: "Write rules for a BTC dip-buying strategy with $500." It suggests buying $50 when BTC falls 5% within 24 hours, selling at 4% above the buy, a 6% stop-loss and at most 2 open buys, and asks about a daily loss limit. That is a good draft.

Pasted into TradeCorgi as "Buy $50 of BTC when it drops 5% in a day. Sell at +4%, stop at -6%", your Corgi asks for the daily loss limit, you choose $25, and you confirm the card. The backtest runs with fees, a paper agent starts, and when you go live, the rules engine, not the model, places each order. This is an illustrative example, not a real result.

A checklist before letting any AI assistant trade

• The rules are written down and you confirmed them.
• Something outside the model enforces your loss limits.
• Something watches the market continuously.
• The rules were tested with fees and on paper.
• The key cannot withdraw.
• You can see why every order was placed.