How to let an AI trade for you on Hyperliquid, without giving it withdrawal authority
A practical guide to connecting an AI model to Hyperliquid perpetuals with separate owner authority, trade-only permission, and a risk engine that can veto.
The idea of an AI that trades for you sounds either exciting or terrifying, and usually it should be both. The exciting part is real: modern models can read market context and reason about a setup faster than any human, and they never get tired, greedy, or scared. The terrifying part is also real: many "AI trading" products ask you to deposit money with an operator and trust a black box. This guide walks through how to let an AI trade Hyperliquid perpetuals while keeping owner withdrawal authority separate from automation.
The two things that go wrong
Almost every bad outcome in automated trading comes from one of two failures.
First, custody. If you deposit into someone else's wallet, the biggest risk is no longer the market, it is the operator. You can be right about every trade and still lose everything if the platform disappears. This has happened again and again in crypto.
Second, discipline. An unconstrained bot, AI or not, will eventually do something reckless: oversize a position, add to a loser, or hold through a cascade. A model that can override its own rules is not disciplined, it is just fast.
A safe setup has to solve both at once.
How a non-custodial AI desk works
The pattern that actually works separates three roles:
- The owner wallet is user-controlled through the configured provider. External wallets are offered only when browser owner approval is enabled.
- A separate trade-only agent signs trades. On Hyperliquid it can place and manage orders but cannot authorize an external withdrawal.
- A risk engine sits between the AI and the market. Every order the AI proposes is checked, sized, and either approved, resized, or vetoed before it can ever reach the venue.
With this shape, the agent can still trade badly and cause losses, but it cannot authorize an external withdrawal. The risk engine limits trading behavior; it does not guarantee safety or profit.
Step by step
Here is the flow, using Hypertrade as the reference implementation.
- Start on testnet. Testnet uses fake funds, so you can watch the whole loop with zero risk. Do this first, always.
- Choose your funds. Hypertrade One creates a user-controlled trading account for real funds without asking for a private key.
- Grant trading permission. A separate provider-held agent is activated only after venue verification.
- Set your risk style. Chill, Normal, Aggressive, and Extreme preserve their deterministic limits.
- Let it run, and watch. Every proposal, veto, and fill is logged so you can see what the engine did.
When you are comfortable with the behavior on testnet, and only then, you move the same setup to mainnet with real funds and a small size.
What good looks like
A healthy AI desk is boring in the best way. It stands aside far more often than it trades. It takes small, defined risk. It respects a mandatory stop on every position. It never adds to a loser. And when the market is ugly, it does nothing rather than force a trade.
If a product promises you a number, guaranteed returns, "high win rate", a magic percentage, walk away. Nobody can promise market outcomes, and anyone who does is selling the wrong thing.
Where to go next
- Read what a perpetual future actually is if funding and liquidation are new to you.
- See why the risk engine matters more than the AI.
- Browse the live Hyperliquid markets you could point a model at.
Ready to try it with zero risk? Start free on testnet.