Not on their own. ChatGPT and Claude are chat models: by themselves they have no exchange account, no order tools and no live price feed, so they cannot place a trade for you. Connected to a trading platform through an MCP connector or an API key, they can write a strategy, backtest it and run it on paper, and the step to real money stays your decision.
The rest of this guide explains where that line sits, what a connector changes, what can go wrong, and how the whole loop works on OptAlgo, with one real backtest and one real walk-forward as the example.
What a chat model can and cannot do on its own
Without any tools, a chat model works with text. That is more useful for trading than it sounds, and less than many videos claim.
- It can explain a strategy idea, such as trend following or mean reversion, and list the rules it needs.
- It can write strategy code in Python or Pine Script from a one-sentence description.
- It can read code you paste and point out bugs, including some forms of lookahead.
- It can explain a result you show it: what a Sharpe ratio or a drawdown means.
What it cannot do without a connection to a real system:
- Place, change or cancel an order. It has no exchange account and no way to reach one.
- See the current price or your open positions with the accuracy trading needs.
- Run a backtest. It has no market history and no engine to run code on it.
- Watch the market while you are away. A chat answers when you write, and stops when you stop.
The third point matters most. Ask a chat model with no tools to “backtest this strategy” and it may still answer with a table of returns. Those numbers were not computed. They are text that looks like a backtest.
What a connector adds
The Model Context Protocol (MCP) is an open standard that lets an AI assistant call tools from another service. A trading platform that offers an MCP server gives the model real tools: instead of guessing a number, it asks the platform and reports what came back. An API key does the same job for terminal agents and your own programs.
With the OptAlgo MCP connector, Claude or ChatGPT can list markets and history, check your plan limits, read the strategy SDK, validate a strategy for free, run backtests, read the honesty grade, the overfit verdict and the equity chart, create paper bots, open and close paper positions, check live prices and list trades.
What the connector can never do is just as specific. It is paper-only by design: it cannot place a live order, switch a bot to live or touch a live bot. So the honest version of “the AI trades for me” is this: the model writes and tests the rules, the platform runs those rules at every bar close, and on paper no real money is involved.
Who decides about real money
You do. On OptAlgo a strategy goes live only when you confirm it in the app, on a Plus or Pro plan. Your assistant can suggest it, and only you can approve it.
API keys for terminal agents such as Claude Code or Codex are a separate path. A key carries scopes you choose (read, backtest, paper bots, live trading), and a key that may trade live carries limits the server enforces: maximum allocation, maximum leverage and a daily loss limit. Every call is logged. The security page lists what OptAlgo can and cannot do with your account.
The risks, plainly
Trading is risky and you can lose money, with or without an AI. Three risks are specific to building strategies with a chat model.
- Invented backtests. A model with no engine behind it can produce results it never ran. Only trust numbers that come from a tool call to a real backtester, and ask the model to say which run they came from.
- Overfitting. A model will happily try “a few more variations” fifty times on the same data. The best of fifty looks good partly by luck. See backtest overfitting for why that fails live.
- Key safety. Never paste an exchange API key or secret into a chat. Use a connector that signs in with OAuth, and give any exchange key trade permission without withdrawal permission.
There is a fourth, quieter risk: fluent text reads as confidence. A model can describe a losing strategy in the same calm tone as a sound one. That is why the grade and the verdict should come before any return figure.
How it works on OptAlgo, step by step
- Create a free account: sign up here. No card and no exchange account are needed to start.
- Connect your assistant. In Claude.ai or ChatGPT, add https://mcp.optalgo.com/mcp as a custom connector, sign in to OptAlgo and approve. It uses OAuth, so no key is pasted into the chat. The connector docs show the setup for each app.
- Describe the idea in one sentence. The assistant writes a strategy.py with the OptAlgo SDK: plain Python you can read and download.
- Validate for free. Validation runs the code once on synthetic bars and lists every problem with its fix, including lookahead and non-deterministic code. It costs no units.
- Run one backtest. It is charged by run time: one unit is up to 10 seconds, and the Free plan has 50 units a month. The result leads with an honesty grade from A to D and a search verdict, then the numbers.
- Ask for one walk-forward research run instead of tuning by hand. It optimises on a training window and judges on test windows the optimiser never saw.
- Deploy to paper. OptAlgo runs the code at every bar close in a sandbox, as a paper bot in My Bots. The deploy gate refuses a losing backtest, and refuses one judged likely overfit unless an out-of-sample research run of the same code passed.
- Decide about live yourself, in the app, once the paper record holds up.
A worked example: one good backtest, then a harder test
Here is what that loop produced on OptAlgo on 2026-11-22, on Binance data. The idea was a simple trend strategy: go long when the 20-bar EMA crosses above the 80-bar EMA, and exit on a 3 ATR trailing stop. It ran on the BTCUSDT perpetual, on 4h bars, from 2022-01-01 to 2026-01-01.
- Honesty grade: A.
- Return after fees and funding: +70.8%, a CAGR of 14.3%.
- Max drawdown: -15.0%. Sharpe: 0.85.
- 55 trades, win rate 40%, profit factor 1.79.
This is one backtest, and one backtest is in-sample. The grade A says the fills were settled on real data with fees, funding and exchange rules. It does not say the idea has an edge. The 20 and 80 were picked by a person looking at the same history the test ran on.
So the next step was a walk-forward research run of the same idea from 2021-01-01 to 2026-01-01. Each of 6 rolling folds trained on 608 days and tested on the next 180 days. In each fold, 30 configurations were tried over a fast EMA of 10 to 40, a slow EMA of 60 to 160 and a trail of 2, 3 or 4 ATR.
- Verdict: likely overfit.
- Out-of-sample return: +23.4%. Out-of-sample annualised Sharpe: 0.49.
- 4 of 6 test folds were positive, with 30 out-of-sample trades in total.
- It failed the deflated Sharpe check, with warnings for a short track record and for train-to-test degradation.
- The chosen parameters changed in every fold: for example fast 30 and slow 120, then 25 and 160, then 40 and 140.
Read those two results together. The out-of-sample return was still positive, but much weaker than the single run, the sample was only 30 trades, and no parameter set stayed best from one fold to the next. A grade A run with +70.8% is honest about fills. It is not proof of an edge.
This is also the answer to the title question in practice. The model did the work it is good at: it wrote the code, ran the tools and reported the verdict. The verdict told a person not to put money on that version of the idea.
So should you let an AI trade for you?
Let it write and test. Keep the decision. A reasonable way to use ChatGPT or Claude for crypto looks like this:
- Use the model for code, experiments and reading results.
- Trust only numbers that came from a real engine, graded and net of fees.
- Treat a single backtest as a screen, and a walk-forward as the test.
- Run on paper before anything else, and compare the paper trades with the backtest.
- Go live only by your own confirmation, with limits on allocation, leverage and daily loss.
The AI trading bot page describes the product side of this. If you work in a terminal, the Claude Code guide walks through the same loop there.
Frequently asked questions
Can ChatGPT place trades on Binance for me?
Not by itself. ChatGPT has no exchange account and no order tools. Through the OptAlgo connector it can write, backtest and paper-trade a strategy. Live trading on Binance, Bybit or OKX starts only after you confirm it in the OptAlgo app, on a Plus or Pro plan.
Can Claude see live crypto prices?
A chat model alone has no live price feed. With the OptAlgo connector added, Claude can check live prices, list markets and history and read your paper trades, because it calls OptAlgo tools instead of answering from memory.
Is it safe to give an AI my exchange API key?
Do not paste an exchange key into a chat. The OptAlgo chat connectors sign in with OAuth and are paper-only, so no key enters the conversation. Exchange keys connected to OptAlgo can trade but cannot withdraw.
Are backtest results from ChatGPT or Claude real?
Only when they come from a real backtest engine through a tool call. A model with no engine can write numbers that look like a backtest without running anything. On OptAlgo every result carries an honesty grade from A to D and an overfit verdict.
Does an AI trading strategy make money?
No strategy wins by construction, and trading is risky. A single backtest is in-sample and never proof. A walk-forward and a paper record show whether an idea held up on data it never saw.