PRE-LAUNCHTwenty founding seats. Mapping calls are open.Request access →
Cook Trading

Tell it how you trade.
It builds the automation.

Describe a trade the way you would to a colleague. Cook writes it down as a strategy, tests it on years of real history, and — once you approve it — runs it on paper, telling you why it acted on every bar and why it sat out the rest.

Request access

Twenty founding seats.

An animation of the loop, in five steps. On a BTCUSDT 15-minute chart you open the rail’s Lines group, pick the trend line, and set its two points on two swing highs — 63,345.93 and 63,223.74 — giving a line of descending resistance. The agent drafts a strategy from it, names the four assumptions it had to invent — exit, size, session and fill model — and flags the one clause it cannot resolve on its own: whether “under 68” reads RSI at the signal close or at the prior close. It writes its reading into the file rather than hiding it. On approval the strategy is frozen to an eleven-line artifact with a hash. It is then replayed over 92 bars, producing 3 entries, 3 exits, 8 holds and 78 refusals, each refusal named from a closed list of 7 reason codes. The fifth step hands the frozen strategy to a simulated six-step Flow: schedule, load, evaluate, route, paper order, and journal. The trading Flow is labelled as not in the build.

It reads the chart
you marked up.

“Break down this resistance I marked” works, because the agent sees the line. Every drawing and note enters the snapshot alongside what you typed.

The Cook workspace. On the left, the chart pane and its drawing rail: 92 fifteen-minute BTCUSDT bars, and a trendline the trader drew between two swing highs, still selected with both handles showing. Docked on the right, the copilot. The trader asks Cook to break down the resistance they marked; a tool row records that Cook read the chart — one line, 2 points; and Cook answers that the line joins two highs, that price turned away from it 4 times across the 46 bars it spans, and that only 2 bars have closed above it.

It shows you every
guess it made.

You said one sentence. A strategy needs a dozen decisions. Cook writes them all down and marks the ones you never actually specified.

A Cook chat transcript, drawn at the size it appears in the desktop app and You send: “Long when a 15m candle closes back above this line. Stop under the reclaim bar. Risk 1%.” Cook thinks for 0.8 seconds, runs draft_strategy in 1.4 seconds, and answers that it wrote the strategy down as Reclaim v1 and needed four decisions you never made. A review card restates the strategy in plain English — Buy when a 15-minute candle closes back above the line you drew, with RSI under 68 at that close. Risk 1% of the account. Get out when price crosses back the other way. — quotes the levels it drew on your chart, entry 63,314.40 and stop 63,143.34 · 27 bp, and puts each guess in a pre-filled field you can edit before you approve: how to get out, opposite cross; how much to risk, 1% of the account; when to trade, all hours; what price you get, next bar open, 1 bp. Below them it refuses the one decision it would not make for you: “under 68” could mean at the moment the candle closes, or at the close before it, which are two different strategies. It took the first and wrote that into the file so you can change it.

Tested on years
it never saw.

Anyone can build a strategy that looks perfect on the years it was built from. Cook builds on one stretch and tests on a later one it has never seen. Every number is after fees, funding and the price you would really have got. The losing months stay in the chart.

The same strategy over two twelve-month windows. On the window it was built from, July 2023 to June 2024, it made +18.4% after all costs, won 54.1% of its trades, won $1.31 for every dollar lost, fell 14.2% from a high at worst, and had a longest losing stretch of 4 months. On the later window it had never seen, July 2025 to June 2026, it made +11.2%, won 51.7% of its trades, won $1.18 for every dollar lost, fell 17.9% from a high at worst, and had a longest losing stretch of 5 months — every figure worse than on the window it was built from. This is an illustrative example, not a real result and not a claim about any account.

No AI runs your strategy.It only helps you write it. What runs is the file you approved.

Nothing runs until it is proven.Not even on paper. A strategy without a backtest cannot be armed.

Cook never holds your money.There is no account here. If live trading ever opens, the keys stay on your machine.

You approve v1.
It stops being a draft.

Frozen, hashed, immutable. Deployed as a deterministic Flow: the runtime evaluates every bar and executes without consulting any model. What runs is the artifact you approved — auditable, reproducible, incapable of hallucinating.

Flow engine verified on a branch, not merged. Trading automation is roadmap phase 4 — nothing on this canvas places a real order.

The Cook Flows canvas for Reclaim v1, read top to bottom. Across the top runs the lock: a hatched frontier with the frozen artifact riding it — Reclaim v1.csl, immutable, carrying its hash — and the note that no model passes this line. Everything below the frontier is the deterministic engine, in five steps: a schedule trigger fires when a fifteen-minute bar closes; a pure function evaluates the rule with no model in the loop; a router asks whether the signal is valid; on yes it places a paper bracket order through the Hyperliquid connection; on no signal it writes the refusal to the append-only proof log. To the right, the step panel shows the paper order selected, with its piece, action and connection bound, and three inputs marked amber because Cook chose them and you have not reviewed them yet. Trading automation is not in the build: this is a simulated flow, the Flow engine runs on a branch that is not merged, and live operation is phase 4 of the roadmap.

It tells you why it traded.
And why it didn't.

Across 92 bars it took 3 entries and refused 78, each refusal named from a closed list of 7 reason codes. You can pause it or kill it at any bar.

Flow engine verified on a branch, not merged. Live operation is roadmap phase 4 — nothing on this surface places a real order.

A Cook Flow monitor for Reclaim v1, drawn at the size it appears in the desktop app. Along the top, the decision trace: one mark for each of the 92 bars the Flow has evaluated — 3 entries, 3 exits, 8 holds and 78 refusals — with three of those bars marked and joined by a line to the narration underneath. Bar 0062 at 10:45, refused: It touched your line and closed under it. Reason code NO_CLOSE_ABOVE. Bar 0064 at 11:15, entered: Closed above your line at 63,314.40. Long, 1% of equity. Reason code ENTER. Bar 0068 at 12:15, refused: The cooldown you set after an exit is still on — one bar left. Reason code COOLDOWN_ACTIVE. Under the transcript, the closed list of 7 reason codes the runtime can emit, each with the number of bars it refused: BELOW_LEVEL, 53; COOLDOWN_ACTIVE, 9; SESSION_CLOSED, 7; NO_CLOSE_ABOVE, 3; INSUFFICIENT_HISTORY, 4; RISK_CAP, 1; DATA_GAP, 1; 78 refusals in total. The Flow engine runs on a branch that is not merged and live operation is phase 4 of the roadmap.

Every patch goes back
through the proof.

Cook proposes weekly patches. You approve or discard. Every refinement is a new version that gets proven again before it can run.

The version chain for Reclaim, running left to right with a proof gate between each version. Version 1: approved, 92 bars, 3 entries — proven and armed. It reaches version 2 only by passing a proof gate: version 2 raised the cooldown to 3 bars and was re-proven, and it too is armed. Version 3, a proposed session filter, is waiting on your call: the gate in front of it has not been passed, so nothing of version 3 runs.

Bring one strategy.
We map it in thirty minutes.

You leave with the full blotter — every entry, exit, fee and slippage assumption — as a CSV you keep. Whether or not you buy.

The mapping callFree
  • Screen-shared, no deck
  • Walk-forward on the call, net of costs
  • Blotter CSV, yours to keep
  • Your strategy on the forward ledger
Book the call
Founding access20 seats
  • Everything in the mapping call
  • Live monitoring on paper
  • The agent's narration, bar by bar
  • Weekly patches, each re-proven
  • Priced on the call
Request access

Questions, answered.

An expert trader works with a trading copilot. You transmit the strategy — speaking, writing, or marking up the chart — and Cook turns it into an executable artifact, proves it against history, and operates it on paper. You start by reviewing every step and end up only supervising.

The terminal, the paper broker and the strategy runtime are built and running. The copilot and the backtester are what we are shipping now, and a flow that trades on its own comes after them. We are pre-launch: this page carries no user count, no testimonials and no returns, because there are none. Ask on the call and we will tell you exactly where the line is on the day you ask.

Paper only. Live trading with your own exchange keys is the eventual horizon and is deliberately outside the MVP. Cook never holds funds — if live mode ever opens, the keys belong to you and stay on your machine.

No. The model proposes, explains, drafts and translates; it never places orders. An approved strategy is deployed as a deterministic Flow whose runtime evaluates every bar without consulting any model. What runs is the frozen artifact you approved — auditable, reproducible, incapable of hallucinating.

Yes, and this is a hard commitment. Strategies, the full trade blotter and your signal history, as CSV — unwatermarked, unlocked, readable without us. You get the blotter from the first mapping call whether or not you ever pay. If export ever gets hard, treat that as information about the numbers.

Say it. Prove it. Run it.

One strategy, from the sentence you said to the bar it traded. Twenty founding seats.