How it works

Understanding this page is optional to use taker, but it explains why the product behaves the way it does — particularly why numbers never come from the chat, and why the model cannot trade on your behalf even if it wanted to.

A sentence becomes a proposal

When you type an order, the language model does exactly one job: it turns your words into a structured intent — market, side, size, leverage, order type. It does not calculate prices, and it never contacts the venue.

The server then prices that intent itself: current market price, your size, slippage, fees, margin required, liquidation estimate. Those numbers are computed server-side, stored, and rendered on a confirmation card.

your words  →  model  →  intent  →  server prices it  →  card
                                                          ↓
                                              you confirm  →  venue

Why this matters

The numbers on the card are never produced by the language model. They are computed from live market data by the same code path that will execute the order. A model that hallucinated a fee or a size could not change what you are shown, because it is not the thing doing the showing.

Confirming is a separate request

Confirmation is not the model finishing its turn. It is a distinct, authenticated request from your browser carrying the intent’s id and a token that binds it to the exact numbers you saw.

Before anything reaches Hyperliquid, that request passes the policy engine — one function, one call site, no exceptions — which checks the order against your own configured limits. See Your risk limits.

If the market moved enough that the card is stale, the order is refused rather than filled at a price you did not agree to.

The model never holds a key

taker’s architecture puts the model on one side of a wall and the trading key on the other. The model produces proposals; a separate path executes them. There is no tool available to the model that places an order.

This is why taker can safely let a language model near a trading account at all, and why “the AI went rogue” is not a failure mode the design admits. The worst a confused model can do is propose something wrong — which you then see, priced exactly, on a card you have to confirm.

Where each surface fits

Surface	Best for
Chat	Expressing intent quickly, risk sizing, multi-step requests, questions
Trade page	Watching the book, precise limit prices, fast repeat entries
Portfolio	Reviewing equity, PnL history, fills, and open orders

They are views onto the same account and the same execution path — not separate modes.