The AI Trader

What The AI Trader Is

An AI Trader is an autonomous trader you attach to your desk — the signals you already scan with. It reads the same live signal data you see on the board, interprets it through a strategy you author in plain English, and for each setup that qualifies it writes a full trade plan: entry, stop, laddered targets, scale-outs, and an invalidation condition. It then manages the position turn by turn, logs its reasoning as it goes, and writes a journal entry when the trade closes. You choose how far it goes: recommendations only, real Hyperliquid testnet orders to prove a strategy risk-free, or real mainnet orders once you arm a non-custodial wallet — and it can never withdraw your funds.

How It Works — At A Glance

Think of it as a junior trader you hired. You give it the screens (your desk), the playbook (your strategy), and a risk budget (its mode). Then it runs this loop:

• YOU SET UP, ONCE — build a desk (pick signals; they score every market into your board), create a trader on it, describe your strategy in plain English, and pick its mode: Analyst = talk only · Execute = real orders on the trading account you select — a main account or a sub-account, testnet or mainnet; the account decides the venue. Turn it on. Execute also needs that account connected and armed in Settings › Wallets.
• EVERY CANDLE CLOSE IT TAKES ONE TURN — Read first: it keeps a running read of the market — its own written game plan — and starts every turn by revising it against what is actually happening: the session and the calendar, what is moving by sector and theme, breaking news on the tickers it watches, your regime readings. Then look: score the board with your signals, the exact numbers you see. Then decide: only what is genuinely in play AND matches one of your strategy's setups becomes a trade plan (entry, stop, targets — real orders in Testnet/Live); a nearly-complete one goes on the watchlist or gets armed; open positions get managed — hold, move the stop, take profit, or exit. Log: the turn's reasoning is written down, always.
• IT KNOWS WHY THINGS MOVE — every market carries its sector and themes (memes, DeFi, semis, crypto-linked stocks…), and when a headline names a ticker on its board, the trader sees it next to the price action — so it can tell a news-driven move from pure flow and say which one your setup is trading. Its read carries from turn to turn with a timestamped history, so it tracks themes over hours, not candles.
• EVERY MINUTE IN BETWEEN, FOR FREE — the armed watcher checks each armed setup's trigger; the moment one completes, its pre-written plan fires through every normal safety check. Stops always rest on the exchange itself, so they hold even if vyx is offline.
• ONCE A DAY — the review: it reads its own measured results (win rate, which setups actually make money, whether they even fired), updates its lessons, and may leave you one suggestion — which only you can apply.
• ON TOP, IF YOU TURN IT ON — the Brief: a short market view (risk-on or off, why, what would change it), rewritten only when the market actually changes. Every AI trader on the desk reads it before deciding.
• WHERE YOU FIT — you own the desk and the strategy; it never changes either by itself. You read the Desk panel: performance, trades, the log. The kill switch is always one tap away.

Create A Trader

Traders live on the rail on the right side of the desk. Choose Compose a trader — creating one is a conversation: the Composer AI chat opens on the left, reads your desk (the signals you scan with and what they read together), proposes a strategy direction, and mints the trader once you agree. A trader always trades your desk's active signals on your desk's interval — exactly the board in front of you — so there is no signal, timeframe, or persona to choose. Each desk holds one trader, and it takes its name from the desk: there is nothing to name, and renaming a desk relabels its trader everywhere at once.

• It trades your desk: the same signals, settings, and interval you see — edit the desk and you change what the trader trades.
• Decision source: an AI analyst that reads the board and authors every trade, or a Mechanical rule that enters straight off your signals with no model (see below).
• Mode: Analyst only recommends; Testnet places real Hyperliquid testnet orders; Live places real mainnet orders (see Three Modes below).
• Sizing: for testnet and live, set the per-order size in dollars and the most open positions it may hold at once. Positions are sized by volatility parity — a calm market gets a bigger position and a jumpy one a smaller position, so each contributes similar risk — with your per-order size as the average and your caps as the hard ceiling.
• Model: for an AI-analyst trader, choose which model does the reasoning.
• Alerts: choose how a notable decision reaches you — a chime, a spoken voice line, or a desktop notification — each tunable for how often it fires. Voice is off by default; hit test to hear any of them before turning it on. Alerts play in whatever tab has the desk open.
• Edit anytime: open a trader's settings with Edit on its card — even while it is running. Saved changes take effect on its next check, so you never have to turn it off to reconfigure it.
• Honest self-review: when a trade closes, the trader's journal is graded with the measured result from its trade ledger — never its own estimate — so the lessons it learns from are built on real numbers.
• Disciplined exits: on a losing position an exit call closes immediately, but on a winning one the first call de-risks half and the exit only completes if the trader confirms it on its next check — one wobbly reading no longer dumps a whole winner.
• Entry quality gates: the trader will not open fresh positions on books too thin to fill fairly, and during historically weak hours (late US evening and overnight) it only takes the highest-conviction, fully-valid setups. Exits are never gated.
• Attempt memory: the trader remembers what happened to its recent orders — filled, refused, or blocked and why — so it stops re-submitting an entry the venue already refused.
• Trades exactly what you see: the trader now honors your desk's liquidity slider. Symbols too thin to trade at your floor are hidden from it just as they are hidden from your board, so it only takes setups on markets deep enough to actually fill — set the slider to All to let it consider everything.

Write — And Edit — The Strategy In Plain English

For an AI-analyst trader you author its strategy by talking to the Composer AI chat on the left — describe the setups you would take, how to enter and manage them, and what to avoid. It drafts a structured strategy phrased in your desk's own signals: a one-line summary, the setups it will look for, and guidance for planning, management, sizing, and no-trade conditions. The chat can see your desk's trader in full, so you can come back any time and just tell it what to change — tighten a setup, add a no-trade rule, or rework sizing and management. It shows the revision for you to Accept before it overwrites your strategy, and it only ever authors the strategy: it never enables, arms, or switches a trader's mode — those stay your controls. An AI-analyst trader needs a strategy before it can run autonomously — a mechanical trader runs from its rule instead.

• Example: "Fade funding extremes on liquid majors only. Enter on a book flip, stop beyond the sweep, scale out in thirds. Stand down when spreads blow out."
• The strategy is yours — composed from your words and your signals, not a downloaded template.
• Come back later and just say what to change — "halve the size in thin books", "drop the breakout setup" — then Accept the revision; the trader keeps everything you did not touch.
• Setups reference the signals actually on your desk, so editing the desk changes what the trader can trade.

Prefer A Fixed Rule? Go Mechanical

Not every trader needs an AI author. Switch a trader's decision source to Mechanical and it enters on a deterministic rule over your desk's signals — no model, no prose, fully repeatable. You choose the rule and the risk geometry; it sizes the position, rests a stop, takes its target, and manages the trade on the same rails as an AI trader — on testnet and live it signs real Hyperliquid orders.

• Entry rule: act when all your signals agree, when at least K of N agree, or when their blended total clears a threshold.
• Risk geometry: a stop at a multiple of ATR, a target in R, and an optional move to break-even — set once, applied every trade.
• Optional exit-on-decay: close when your signals stop agreeing, or leave the stop and target to do the work.
• Optional fade: flip the trade to the opposite side — a buy setup opens a short — for signals that work better faded than followed. Our backtesting found order-flow buy extremes tend to mark short-term tops on the hourly timeframe, so a fade desk trades against them. Run it in simulation first to see its track record before going live.

The Trader Loop

Once enabled, the trader runs around the clock on VYX servers — no browser or open tab required, and nothing ever runs in your browser. It takes a turn each time a candle closes on its timeframe — a 5-minute trader every 5-minute close — so its reads line up with the bars you see. Want a read right now? Analyze asks the server loop to take a turn immediately. On each turn it works like a discretionary trader:

• Perceive: score the board with your signals — the same rows and values your heatmap shows.
• Focus: attention goes where a trader's would — its open positions, its watchlist, setups firing or one signal away, fresh extremes waking up, and the biggest movers. It never wades through a fixed leaderboard: a quiet market reads quiet, and standing down is a valid read.
• Scan: read that working set through the strategy. No qualifying setup means "nothing here" and the turn ends — no forced trades, no churn. Something merely interesting goes on the watchlist for a closer look next turn.
• Plan: for each fresh setup, author a trade plan — direction, entry, stop, target ladder, scale-outs, invalidation, and size within your caps.
• Manage: for each open position, re-evaluate against fresh data — hold, move the stop, take a partial, add, tighten to break-even, or exit if the thesis breaks.
• Record: append the turn's reasoning to the decision log and the trade's running thought-log. Every trade is also stamped at entry with the setup it expresses and the conditions around it — which trading sessions were open, how strong and how rare the signal was — so performance can be graded per setup, honestly.
• Reflect: when a trade closes, write a journal entry and update what it has learned.

What The Trader Reads Beyond Your Signals

Your signals decide what qualifies as a setup. Around them, the trader keeps the same context a human desk keeps in the corner of its eye — computed exactly, at no AI cost, fresh every turn:

• Price structure, per market: where price sits in its recent range, how wide that range is, and whether volatility is contracting or expanding — neutral facts it reads through your strategy, whether you trade breakouts, pullbacks, or fades.
• Who moves together: markets are grouped by how they actually move once the market-wide tide is stripped out. It can tell a coin moving on its own story from one just riding its group — and it knows three positions inside one group are really one bet, not diversification.
• BTC sensitivity: how much of a move is the whole market, and how much is the coin itself.
• The clock: which global sessions are awake (Sydney, Tokyo, London, New York), the weekend, and the high-impact economic calendar. Thin hours argue for standing down; minutes before a big print is no time to open fresh risk unless your strategy explicitly trades news.
• The higher-timeframe tape: each market's 1-hour and 4-hour price trend, so a 5-minute setup is taken with — or knowingly against — the bigger flow.

Armed Setups — Patience, Engineered

A good trader doesn't stare at the screen waiting for the exact minute a setup completes — they define the moment and let the desk wake them. Your AI trader works the same way. When it sees a setup that is close but not ready — three of your four signals agree, say, and it is waiting on the fourth — it can arm the trade: write the complete plan now (entry, stop, targets, invalidation) and state the exact condition that should fire it. From then on vyx watches that condition every minute, between the trader's normal turns, at no AI cost. The moment the condition is met, the plan fires through every normal safety check — sized, capped, and gated exactly as if the trader had placed it on a live turn — and the decision log marks the entry with an amber armed badge: the machine pulled the trigger the AI loaded.

• Up to five armed setups at a time, each with a time-to-live (default 4 hours, up to 24) — a condition that never comes quietly expires, with a dim note in the log.
• It fires only if the setup is still fully valid at that minute — never on a stale idea — and a market the trader already holds can't re-fire its own entry.
• News-aware: inside the half hour before a high-impact economic print it holds its fire, and re-checks after the release.
• You can see what's armed: each live setup shows on the trader's card and desk panel — the market, the side, the setup and its condition, and how long it has left. Change your mind? Cancel any one of them from the desk panel; the trader simply won't fire on it (it may re-arm later if the case genuinely reforms).
• Turning the trader off stands down everything armed with it.

The Desk Analyst — A Top-Down Market Brief

Turn on the Desk Analyst and your desk gains the layer a trading floor calls the morning meeting: a top-down Market Brief every AI trader on the desk reads before it acts. The analyst watches the same market your desk does — regime state, firing breadth, higher-timeframe BTC trend, which global trading sessions are awake, and the high-impact economic calendar — and rewrites its brief only when something actually changes: a regime flips, London or New York opens, a high-impact print lands, BTC makes a real move, or the view simply goes stale. The brief is context, never command — each trader still trades its own strategy, and it can disagree with the desk view when its setup says so. Find it on the desk panel: a one-line Brief row showing the desk's bias and how old the view is; expand it for the full read, what would change it, and what the analyst is watching.

• Per desk, off by default: turn it on from the Brief row on the desk panel. Refreshes are capped per day, so cost stays trivial.
• Event-driven, never a timer — a quiet market writes no new brief, and the row shows the age of the current one.
• Every trader on the desk reads the brief with its age attached, so a stale view is discounted, not obeyed.
• It also weighs what your own traders learned: each trader's latest daily-review takeaway feeds the analyst, so the desk brief reflects what's actually been working here — history informing the read, never a trade call.
• Refresh now asks the analyst to re-read the market immediately — the brief updates within seconds.
• If the Brief can't run — the analyst is off, the desk has no enabled AI trader, or that trader is held at the gate — Refresh now says so right on the row, with the exact reason.

Trade Plans

A plan is the unit of a trade — one open plan per symbol per trader, carried as living state from proposal to close.

• A thesis in the strategy's terms: why this setup, right now.
• An entry (market, limit, or trigger), a required stop, and a ladder of targets with per-level scale-out sizes.
• An invalidation condition that exits the trade early when the thesis breaks — before the stop is hit.
• The AI proposes; deterministic guardrails validate. Plans are checked against price sanity and your sizing caps, and anything out of bounds is clamped or rejected. The strategy's words can never override your limits.
• Stops, targets, and trailing are enforced mechanically each turn — the bookkeeping never depends on the AI.
• You can see the plan on the chart: while the trader holds a position on the charted symbol, its entry, stop, and every remaining target draw right on the price axis, and the same stop and target ladder show in the Desk panel. A solid stop line rests on the exchange as a real protective order; the dashed target lines are managed by vyx — closed at market as each level is reached — so nothing sits waiting on the venue for them.

Replay Any Trade

Click any position in the trade panel — open or closed — and its whole life opens inline: the setup that triggered it, every hold and adjustment while it was on, and how it ended. It plays on the chart you are already reading, not in a separate window.

• Entry: the thesis and the signals that fired it, at the price and stop it went on.
• While it was open: the quiet holds collapse to a single line, and each real change — a tightened stop, an early exit, a scaled target — reads on its own, with the reason the trader gave.
• The live read leads with how far the trade is from its stop, in R — a drawdown seen as risk, not a bare dollar figure.
• The close: the venue-true exit, the realized R and PnL, and why it ended — target, stop, or invalidation — plus the trader's written reflection.
• Hover any moment and the chart's crosshair snaps to it. A manual trade with no AI behind it shows its venue fills.

The Logbook

Every trader is fully auditable. On an automated desk everything lives in the one Desk panel (below); on a manual desk with an advising trader, the Logbook sits on the desk rail directly beneath the trader card — either way it is always on screen, nothing to open, showing what the trader is doing and why. Each desk holds one trader, so you are always reading exactly the trader that trades the board in front of you.

• The forward record: the equity curve built only from what the trader actually did.
• The decision log: one entry per turn — what it saw, why, what it did, and the outcome. Recorded server-side, so it is the same log on every device; skipped turns show too, with their reason, and every entry is badged by its author — mech, ai, or an amber armed for an entry the armed engine fired; daily self-reviews carry a violet review tag.
• The thought-log: the server loop's per-turn reasoning while a trade is open, threaded to the trade.
• The journal: a reflection per closed trade — what the setup was, how it played out, the R outcome, and the lesson.
• Run cost: what the trader costs in AI compute — a running total, average per run, and a per-run cost chart, alongside each run's cost in the thought-log. Shown next to its measured P&L, on desktop and mobile.
• Every market the trader names is a live link: hover to preview it on the chart, click to jump the chart and board straight to it at the exact candle it was flagged. Each entry is stamped with its date and time.

Market Context — What The Trader Is Seeing

On a manual desk with an advising trader, a Market panel sits on the rail between the trader card and its Logbook — the same read of the market the trader itself works from, so a quiet board never looks like a dead app. It shows which global sessions are awake and the next open or close, the higher-timeframe trend of the majors (BTC, ETH, SOL on the 1-hour and 4-hour with the 4-hour move), the next high-impact economic release, your desk's Market Brief when the Desk Analyst is on, and the trader's latest read — the one-line take from its most recent turn, with how long ago it thought it. It is context, not a call: it tells you what the trader is looking at, never what to do.

• Quiet is information, not emptiness: a still weekend board reads "board quiet" in the trader's own latest note instead of showing you nothing.
• Every line appears only when there is something to say — a desk with no analyst simply shows fewer lines, never an empty panel.
• The latest read updates every turn the trader takes, so while it is on you can always see its current thinking at a glance.

The Desk Panel — An Automated Desk At A Glance

When a testnet or live trader runs a desk, the trade panel becomes one quiet Desk panel, ordered the way you actually read it. Performance first: the wallet's equity, open and realized PnL, the trader's equity curve, and its record in one card — tap the record for the deeper stats (win rate with its confidence range, lessons learned, what the trader costs to run). Then the trades: open positions with their stops and target ladders, plans not yet on the venue (tagged), and its recent closed trades. Then the activity: the desk's Market Brief, open by default, followed by the near-miss setups it is tracking and its live decision log. A single status line appears only when something needs you — not armed, or the kill switch — with the one action that fixes it; a healthy desk shows just a green dot. The trader's strategy and settings live behind the gear on the trader's name. The desk is marked AUTO on its cards and its wallet is view-only: watching is unlimited, but manual orders stay off while the trader is on. Want the wheel? Off is Take over — it pauses the trader and hands its wallet back to you; the kill switch is always one tap away.

• Click any trade — open or closed — and it opens full-screen in the panel: the outcome, the plan it worked (entry, stop, every target and what filled), the setup and conviction the moment it entered, the trader's reasoning turn by turn over the trade's life, the actual venue fills, and the lesson it journaled. Back returns you exactly where you were.
• chart → inside a trade's story jumps the chart to that trade's window, so you can read every decision against the price action that caused it.

Learnings

The strategy keeps a short, curated list of one-line lessons the trader has drawn from its own closed trades — "thin books stopped me out; require a liquidity floor." It is a standing rulebook, not a trade diary: after each journal entry the trader works the list it already has, reinforcing the lesson this trade proved again or sharpening its wording, and only writing a genuinely new one when nothing there covers the situation. Restating a lesson in different words strengthens it rather than adding a second copy, stale lessons are dropped, and the list is hard-capped so it stays sharp. The trader rereads its learnings every turn before it plans. Changes to the strategy core or your desk are never self-applied — those always wait for you.

The Daily Review

Once a day the trader steps back and reviews itself the way a serious desk reviews a trader: with the ledger, not with feelings. It reads its own measured record — win rate, average win against average loss, and how each named setup in your strategy is actually performing — plus how often its setups fired at all this week (and which signal was the last one missing when they almost did), and how the things it merely watched turned out after it stopped watching. From that it curates its learnings — it is the one pass that sees the whole list at once, so it merges lessons that had drifted into near-copies of each other, drops what the ledger no longer supports, and adds at most one that the week actually earned — and at most once it leaves you a suggestion: a concrete, evidence-cited strategy change — "this setup fired zero times in seven days; the holdout was always the same signal — consider requiring three of four" — that only you can apply, in the Composer. Review turns appear in the decision log under a violet review tag. A trader that took no trades all week still gets its review — that is the point: it learns while flat.

• Runs once per day, costs one AI turn, and can only write — a review never places, changes, or closes anything.
• Numbers first: every claim cites the measured ledger, and with under about ten trades it says the sample is thin instead of pretending to know.
• Suggestions never self-apply. The trader refines its own lessons; the strategy's structure stays yours.
• When it leaves one, the suggestion shows as a quiet note on the desk panel — read it, act on it in the Composer, and dismiss it; the next day's review may leave a fresh one.

Manual Or Automated — And Each Trader Picks Its Own Account

Every desk answers one question: who places the trades. Simulated means the trader takes its own trades against a sim book — a real forward record, no money — and the desk stays yours to trade by hand. Automated means the trader places the orders itself, on the account it selects, so every desk can run its own book. Set it in Edit trader; Start and Stop control whether it is working. If you want something that only watches and reports rather than trades, that is the Analyst, added on the desk rail.

• Manual — you trade, the analyst advises. It reads the board and posts its calls; it never touches an order, so it starts the moment you press Start.
• Automated — real Hyperliquid orders on the account you pick in the trader's Account row: any connected main account or one of its sub-accounts, testnet or mainnet. The account decides the venue — pick a testnet account to prove a strategy with zero money at risk, a mainnet one to trade for real (picking mainnet asks you to type the word first). Setting a desk to Automated does not set it running: Start is the deliberate act that lets it place orders, and its account must be armed as well — two switches, because arming is what the safety breaker takes away when your limits are crossed.
• Stopping hands its open trades back to you. Their stops rest on Hyperliquid and keep protecting them, but take-profits are worked by vyx rather than the exchange, so they stop with the trader — the confirm says so before you agree to it.
• Clear data resets a trader to blank — its strategy, setups, record, journal and decision log — while keeping its desk, account and sizing. It is refused while the trader holds an open trade, so a live position is never left with nothing managing it.
• Different traders, different books — give each desk its own account (or sub-account) and they trade side by side without touching each other's positions. Traders sharing one account take turns (the earliest keeps it; the other says so), and a trader never touches a position it didn't open — your manual positions are structurally off-limits.
• Wallets themselves are managed in Settings › Wallets: connect one or many, see each account's sub-accounts and equity, arm, and kill from one place.

Not Trading? Read The Gate

A started trader can still be held at the gate — and the desk always tells you where. An automated trader takes its turn only when the account can actually place its orders: the connected wallet's network must match the account the trader signs on, and the account must be armed. Until both are true the trader sits out, and the desk's status line names the one thing to fix. Everything downstream follows the same gate: the Market Brief serves the desk's trader, so if that trader can't run, Refresh now tells you why on the spot instead of silently doing nothing.

• The status line is specific — "set to testnet but the connected wallet is mainnet", "not armed" — and carries the control that fixes it.
• Held turns are logged: the decision log shows every skipped turn with its reason, so a quiet trader is never a mystery.
• In a hurry? Flip the desk to Manual — its analyst has no gate, starts reading the board immediately, and touches no orders.

How Live Trading Stays Safe

• Non-custodial: it signs on an agent wallet you approve — it can only place and close positions, never withdraw or move your funds.
• Stops rest on the exchange: every position's protective stop is placed on Hyperliquid itself, so it guards you even if vyx is offline between turns.
• Always-reachable kill switch: one control disarms every live trader instantly, and per-order size caps hold no matter what the strategy says.
• Auto-disarm safety breaker: if a day's losses or a losing streak crosses your limits, live trading shuts itself off automatically. An optional rate limit — off by default — can also halt the account after a set number of positions opened in an hour; it counts opens only, never exits, and starts fresh whenever you re-arm. The halt is never silent — the trip and its reason land in the trader's Logbook, the Live Execution panel shows why and when the account disarmed, and every held turn says so until you re-arm.
• Venue-true P&L: your profit and loss is measured from your real exchange fills — never from what the AI thinks the trade did.

One Focus Per Desk — And Safe Edits

Each trade desk has a single trading focus: Manual (you trade it), or an AI or Mechanical trader on Testnet or Live. Because the board you see is exactly what the trader sees, you always know what it is working from — and when a trader is running a desk with real money, a green LIVE marker sits in the top bar so it is never out of sight. When an automation is running a desk, changing the board no longer re-wires it behind your back.

• Your edits stage as a draft. Change the tradable floor, active signals, regime, min-score or interval on an automated desk and a banner shows exactly what moved — the trader keeps running on the old setup until you decide.
• Apply, when you mean it. Applying makes your changes the trader's new basis from its next check. On a Live, real-money trader this takes a deliberate confirm — never a stray click.
• Or branch off. "Continue in new desk" copies your changes into a fresh Manual desk so you can keep exploring, leaving the running automation completely untouched.
• A Manual desk has no lock — with no automation to protect, your edits just save as you go.

Good To Know

• Running traders 24/7 server-side is part of the paid plans: mechanical traders run on Trader ($29/mo); AI traders and analysts run on Autopilot ($79/mo) and up. Accounts from the beta keep everything free.
• Each account has a daily AI-turn budget, so a busy board cannot run up unbounded usage.
• Trader decisions are also readable programmatically via the API and MCP server.
• A track record is a research instrument, not a promise of returns. Nothing here is financial advice.