Bring your own AI: using Claude, GPT or Gemini to trade perps

You already pay for a frontier model. You use Claude, GPT, or Gemini every day. So here is a fair question: why can that same intelligence answer your emails but not read a market and propose a trade? The answer, until recently, was that there was nowhere safe to let it. This post is about the "bring your own AI" approach, and why it is the right shape for automated trading.

The model is not the moat

It is tempting to think the smartest model wins. In trading, that is mostly wrong. Every capable model can look at price action, funding, and order flow and produce a reasonable trade idea. Proposing a trade is close to a commodity.

What separates outcomes is everything after the proposal: how the position is sized, whether a bad idea gets vetoed, how cleanly the order executes, and how honestly the result is reported. That is why forcing you onto one proprietary model is a weakness dressed up as a feature. The better design lets you bring the model you already trust.

What "bring your own AI" means in practice

Concretely, you connect your own API key for the model of your choice. The platform feeds that model live market context, the model returns a proposed trade with a thesis, and then the platform's own risk engine takes over: sizing, vetoing, and executing under rules you set.

The benefits are direct:

• You control the brain. Prefer Claude for its reasoning, or GPT, or Gemini? Your call. Swap any time.
• You control the cost. With your own key, you pay the model provider directly. No opaque markup.
• You are not locked in. If a better model ships next month, you point at it. The desk, the risk engine, and your track record stay the same.

The guardrails that make it safe

Handing any model the ability to trade is only sane with hard constraints around it. A serious setup enforces, at minimum:

• A mandatory stop-loss on every position.
• A maximum leverage cap the model cannot exceed.
• A per-trade risk limit and a daily loss limit, both hard.
• One position per symbol, so the model can never stack or average into a loser.
• A separate trade-only permission that can manage orders but cannot authorize an external withdrawal.

With these in place, the question stops being "do I trust this model completely" and becomes "am I comfortable with the worst case, which is bounded by rules I set". That is a much better question.

A note on prompt injection

There is a subtle risk unique to AI systems: a malicious market data source or instruction could try to manipulate the model. A well-built platform treats model output as untrusted, validates every field, and never lets a model's text talk its way past the risk engine. The engine is deterministic code, not a persuadable assistant. That separation is what makes bring-your-own-AI safe rather than reckless.

Try it

The honest way to start is with fake money. Connect a model, set your limits, and watch it run on testnet before a single real dollar is involved.

• New to perps? Read the perpetuals explainer first.
• See why the risk engine matters more than the model.
• Start free on testnet and point your own AI at live markets.

Not financial advice, just tools.