How to Create an AI Trading Agent for Hyperliquid

Most 'AI trading bots' are a fixed rule with a chatbot attached. Signalview's AI agents work differently: you write the strategy in plain language, pick the tokens it may trade and the data it may read, and set hard risk limits. From then on, every check, the agent receives a fresh market brief for each token and decides one action per token: open, close, reduce or hold. The model applies your strategy to structured data it's given rather than guessing from nothing, and every decision passes through risk checks in code before an order is sent.

This guide walks through creating one, what data it actually sees, the guardrails it can't talk its way around, and how to run it safely. It describes our own product, so we'll be specific about its limits as well as its features.

Published September 30, 2026.

Step 1 — pick the tokens

An agent trades a universe of tokens, not one market. Pick any Hyperliquid perp — BTC, ETH, SOL, HYPE and the rest of the main book — plus TradFi perps on the xyz builder dex (equities, indices, gold, FX), which follow their own market hours. Each token gets its own brief and its own decision every check, so an agent can be long one market, short another and flat on the rest at the same time.

Fewer, liquid tokens are the better start. More tokens means more decisions per check, a larger prompt, and more ways for a mediocre strategy to find a trade it shouldn't take.

Step 2 — choose what it reads

Five inputs, each switchable per agent. Trend: direction and strength on the 1h, 4h and 1d timeframes from the same scoring engine behind Signalview's published signals, plus a plain-language alignment read such as 'higher timeframes long, 1h pulling back'. Regime: trending up, trending down or chop on the 4h, with ATR volatility. Micro: funding and its annualised rate, open interest, 24h volume and change, mark-versus-oracle premium, and the order book's spread and depth imbalance.

Macro: BTC's 4h trend and funding as crypto beta, plus 24h moves in S&P 500, gold, dollar-index and VIX perps, and the economic calendar. News: a live web search per token and for the macro backdrop, summarised into headlines, a sentiment score and an event-risk level with the time of the next scheduled catalyst. You can also attach published Signalview signals to specific tokens as extra evidence.

More inputs isn't automatically better. A trend follower may do best reading only trend and regime; a funding strategy needs micro. Turn off what your strategy doesn't use.

Step 3 — write the strategy

The strategy is plain text, and it's the most important thing you write. Presets give you a starting point: Trend follower (trade with 4h and 1d, stay flat in chop, stop about 1.5x the 4h ATR, target at least 2x the stop), Mean reversion (fade stretched moves only in choppy regimes when funding is crowded), Funding & positioning (trade crowded positioning confirmed by the book), and News momentum (trade fresh catalysts only when the 1h and 4h trend already agrees).

Write it the way you'd brief a disciplined human trader: when to enter, when to stay out, where the stop goes relative to volatility, when to take profit, and when to close early. Vague strategies produce vague trades. 'Buy strength' will fire constantly, and 'enter long only when 4h and 1d both point up and funding isn't extreme' won't.

Step 4 — set the limits the agent can't break

Risk limits are enforced by code on every decision, whatever the model outputs. Budget is the total margin the agent may commit (default $200). Max positions caps how many are open at once (default 3). Max position size caps any single position's margin as a share of the budget (default 40%). Max leverage defaults to 5x and can go up to each market's Hyperliquid maximum: 40x on BTC, 25x on ETH, 20x on SOL, less on smaller markets. A daily loss stop halts new entries after losing a set share of budget in a UTC day (default 10%), and a daily entry cap limits churn (default 10).

Then the automatic checks. Every open needs a stop-loss and take-profit, and the reward must be at least as large as the risk. If the stop would sit beyond 80% of the distance to liquidation at the requested leverage, the agent lowers the leverage until it doesn't, so the stop fires before liquidation can. An entry is rejected if the account doesn't have the free margin, or if the account already holds a position on that token that the agent didn't open. The agent never touches positions it didn't open. And with event risk on, new entries pause within an hour either side of a high-impact scheduled release (CPI, PCE, FOMC), while positions it already holds keep their stops and targets.

Step 5 — choose private or public, then run it

A private agent is visible only to you. A public agent gets its own page with its live positions, its decision journal (every check's reasoning, orders and rejections) and its track record, and other traders can follow it. More on that in How to Follow AI Trading Agents.

To run it, approve a scoped Hyperliquid agent key from your wallet. It can place and cancel perp orders but can never withdraw or transfer funds. Then add a Groq API key: agents run their decisions and news searches on your own Groq key, and Groq's free tier is enough for a small agent. You can start in paper mode first to watch its decisions against live prices without risking funds, then switch to live. By default it checks every 15 minutes, 24/7, whether or not your browser is open. The agent page shows whether it's running, when it last checked and when it checks next, and flags anything that needs you, such as a revoked key.

What to expect, honestly

An AI agent is an amplifier for the strategy you give it. It will apply your rules consistently, around the clock, without revenge trades or moved stops, and that discipline is the real value. It won't turn a strategy with no edge into one with an edge, and a language model reading a brief is not a crystal ball. Expect to spend the first days reading the journal: when it held, why it rejected what it rejected, and whether its reasoning matches what you meant.

Start small, in paper mode, on a few liquid tokens with modest leverage. Scale only after enough closed trades to tell skill from noise, which is more than a handful. For the broader view of what these systems can and can't do, see Do AI Trading Bots Actually Work?.

Risk note: perpetual futures are leveraged, high-risk instruments, and automation doesn't remove the risk of losing your entire margin. Signalview agents trade with a 0.1% builder fee on top of Hyperliquid's normal trading fees. Nothing here is investment advice.

Frequently asked questions

What data does a Signalview AI trading agent use?

For each token it trades, the agent gets a brief with trend on the 1h, 4h and 1d timeframes, regime and ATR volatility, funding, open interest, order-book spread and depth, macro context (BTC trend, S&P 500, gold, dollar index, VIX, the economic calendar) and a live news summary with sentiment and event risk. Each input can be switched off per agent.

Can the AI agent ignore my risk limits?

No. The model proposes actions, and code checks each one against the budget, position count, per-position cap, leverage cap, daily loss stop, daily entry limit, stop-loss and take-profit requirements, liquidation distance and free margin before any order is sent. Anything that breaks a limit is rejected and logged in the journal.

Do I need to pay for an AI model to run an agent?

An agent that makes its own decisions runs on your own Groq API key, and Groq's free tier covers a small agent. Following someone else's public agent doesn't need a Groq key, because the follower mirrors the creator's trades rather than making decisions itself.

Can a Signalview agent withdraw my funds?

No. The agent trades through a scoped Hyperliquid agent key that can place and cancel perp orders only. It can't withdraw or transfer, your funds stay in your own wallet, and you can revoke the key at any time.