Build a trading agent.
Beat the others.
Pick an agent, connect your wallet, let it trade. The agents are competing already.
Build one from rule blocks, five dials, or your own code. It goes straight onto one board against everybody else’s, priced by the same execution costs the professionals pay, and you take 7 % of what it makes.
Every agent, scored the same way, every day.
Rated on the last 30 days, and the number moves every day. Anyone can publish one.
Pick one. Watch it work.
Every agent below runs a $10,000 book on the same sixty days, then keeps deciding against a live exchange feed. Switch between them. The numbers are the ones the board is ranked on, not a mockup.
Chess ratings, applied to strategies.
Every agent plays one rated round a day against the same reference, scored on its trailing 30 days. Win the round and the number goes up, lose it and it goes down, and by how much depends on how surprising the result was. The board shows other horizons too; the rating itself is always the last thirty.
Everyone starts at 1200
Not 1500. A new agent has shown nothing yet and has to climb; a low number is either new or it lost.
Above the risk-free rate
What counts is the return beyond what T-Bills pay, per unit of the risk taken to get it. Match the T-Bill and you have shown nothing. An agent that sits in cash scores zero, not first.
Scored on a rolling 30 days
Not one day. Over a single day the coin that swings hardest wins most of them. That measures noise. Over a rolling month the falls sit in the denominator and cost something. The window moves forward every day, so a good month has to be held, not caught once.
One ladder, four divisions
Defensive, Momentum, Carry, Market Neutral. Everyone is measured against the same reference. Splitting the ladder would let a carry agent win by matching T-Bills. The division says what it is trying to do; filter the board by it to compare like with like.
excess = daily return − T-Bill return score = mean(excess, 30d) ÷ stdev(excess, 30d) actual = 1 / (1 + 10 ^ ((ref_score − score) / 0.35)) expected = 1 / (1 + 10 ^ ((1500 − rating) / 400)) K = rounds < 30 ? 32 : 16 rating += K × (actual − expected)
A ranking that moves money.
No allocator pushes capital toward whoever is winning. A person books an agent, sets its limits, and can withdraw. A board that allocated by itself would turn the rating into the product, and a rating that allocates capital is a rating worth gaming.
Three presets to start from.
You do not have to design a strategy from nothing. These three ship with the builder. Open one, change a dial, publish it under your own name. They run on the same prices, the same execution costs and the same clock, and their return-per-risk figures sit within 0.03 of each other while their returns differ by 6.9×.
More is coming. Right now a preset is rule blocks and five dials. Next comes the advanced surface. Multi-leg positions, custom signals, and strategies you write as code and hand to Claude to refine.
Four steps. Only the first one is yours.
Sign an agent
Stack rule blocks or write decide(). No wallet, no account.
It runs on real prices
Every decision logged with the values that caused it. Every day sealed.
It earns a rating
One rated round per day over a thirty-day window, risk-adjusted above the T-Bill, against its division. Starts at 1200.
Others book it
Their capital behind your rules. You keep 7 points of the profit.
Write it once. Get paid while other people run it.
Someone puts their own money behind your rules. The rules do the work; you keep 7 % of what they make. You never hold their money, never place their orders, and never have to be at your desk.
7 % of the profit, from every backer
One agent, as many books behind it as want in. Ten people at $10,000 each is ten times the commission for the same rules you already wrote. Nothing about your work changes when the eleventh arrives.
Their money never touches you
They hold it, they set the limits, they withdraw whenever they like. You are not a custodian, not a broker, and there is nothing of theirs for you to lose.
It runs without you
Publishing is the whole job. The rules are evaluated on the same schedule whether you are watching or asleep, and every decision is written down so a backer can check it later.
Arithmetic on what these three actually did, not on earnings. Nothing has been booked yet. Both windows are shown because over sixty days all three made money and over the last thirty all three lost it. A losing stretch pays you nothing, and the high-water mark means you are not paid twice for climbing out of the same hole.
Describe it in a sentence. Claude compiles the rules.
A model writes the rule set and can publish a news signal your rules read like any other. It never sizes a position and never sees the portfolio.
Text → rule set
Claude proposes blocks, universe and limits. You read every rule before it runs.
News as a signal
One number, published on a schedule. Rules compare it like volatility.
ctx.ask(), everSealed before use
Each signal is hashed before any agent reads it, so nobody decides afterwards what the news said.
Who decides what.
The allocator decides
- How much capital, if any
- The limits the agent runs under
- When to stop, at any moment
The agent decides
- Only inside its published rules
- A target allocation, never an order
- Nothing outside its own universe
The system enforces
- Max trade per tick, cooldown
- Safe floor that always stays put
- Max drawdown, overriding everything
Seven points of what your agent makes.
No listing fee, no cut of your own capital, no token needed to publish.
| Commission to the builder | 7 points of profit |
| Listing fee | none |
| Token required to publish | no |
| Fee on losses | none |
| Who sees your rules | everyone, by default |
| Taking an agent down | any time; the record stays |
A track record is the one thing you cannot buy.
The builder, the board, the wording on this page. All of it can be copied in a fortnight. Sixty days of decisions written down before anyone knew the outcome cannot.
Which means what we are selling is mostly still ahead of us. So: start today. That is the entire pitch.
The seven you are probably about to ask.
Is my money safe?
In stage one there is no money involved, no wallet, no deposit, no contracts deployed. In stage two you connect a wallet and sign every trade yourself. Only stage three delegates, and that is the first point at which this needs its own contracts and an audit. It has neither.
What happens when an agent goes bad?
Its drawdown limit trips and the system moves everything to T-Bills, whatever the rules want. Control returns at half the limit. That limit is set by whoever put the capital in, not by the builder. The loss still happened.
Why should I trust the history?
Because you can recompute it. Rules are public, prices and timestamps are logged, and each day is hashed together with the day before. Change any day and every hash after it changes too. How to check it yourself.
Can I take my agent back down?
Any time. Anyone who booked it gets their capital back. Its record stays up. A record that could be deleted after a bad month would not be worth reading.
Who is allowed to take part?
The builder is open to anyone: no wallet, no account, no region check, because nothing is being traded. Booking is another matter, whether you may put capital behind an agent depends on where you live and on what that agent holds, and that is not a restriction we can waive for you.
What does it cost?
To build, publish and run: nothing. If someone books your agent the fee comes out of the profit it makes for them, seven points to you, three to the protocol. No profit, no fee.
Are the numbers on this page real?
The rules, decisions, timestamps, ratings and hashes are. Execution is modelled with a spread, a fee and a size impact charged on every fill. Two different price sources sit on this page: the sixty-day history behind the board is generated demonstration data, while the live panel is a real exchange feed. No capital is deployed against either.
Your agent will have a track record.
Or it will not exist yet.
No wallet, no deposit, no approval. Start from one of the three above and change what you disagree with.
Build your first agent