Cursor for
trading strategies.
A trader describes a setup in plain English. Plutus compiles it into rules they can read and edit, tests it against years of candles, runs it on the live tape without money, and only then trades it from the trader’s own wallet through a key that cannot withdraw. Live on Hyperliquid.
Built inside Vitwit, an infrastructure company operating since 2018: $39.2M staked across networks, zero slashing events.
PRE-SEED · $1.5M231,000 traders. Two ways to automate, both bad.
Hyperliquid alone had 231K active traders in March 2026, and 74 to 89% of retail accounts lose money, because a read that is not tested, sized and executed identically is not a strategy. Here is what a trader chooses between today.
Write the strategy as code, not as a sentence
Build a backtest whose result you can trust
Run execution that survives 3am
Months of work before the first trade
A signal group, a bot, a copy-trade
No way to test it before you buy
No way to verify the record after
Usually your API keys, sometimes your funds
Plutus is a third option: the trader brings the idea and the capital, the software does the translation, testing, execution and limits, and shows its work at each step.
This was not buildable two years ago.
Turning plain English into typed, executable strategy logic is a task language models only recently got reliable at. That reliability is what removes the engineer from the loop, and it is the whole premise.
Perp DEXs are 26% of the futures market. On Hyperliquid a trader signs one EIP-712 approval that lets an agent key place and cancel orders and nothing else. The audience and the safe rail arrived together.
A Hyperliquid record can be checked against the chain by a stranger who does not trust us and does not have to ask us for anything. Without that, a track record is a screenshot.
The sentence becomes a spec the trader can read.
A language model can misread a sentence. That is exactly why the spec is rendered, editable, and never the last word before capital moves.
Thirty-plus indicators across trend, momentum, funding and open interest. Multi-condition regime logic, session windows and cooldowns. Scale-in, partial take-profit and volatility-target sizing. Seven timeframes, 1m to 1d.
One evaluator runs that spec in backtest, paper and live, so the thing that is tested is the thing that trades. That is the claim most of this category cannot make.
What has to be true before it trades.
We chose these thresholds. Nothing yet demonstrates they select strategies that survive live, and backtests do not model fees or slippage yet, which is the gap that matters most. We publish predicted versus realised every month so this is settled with evidence rather than assertion.
The trader never hands over their keys.
The barrier to automating is not features, it is that every existing option asks for API keys or custody. Hyperliquid’s approveAgent lets the trader sign once, granting an agent key permission to place and cancel orders and nothing else.
Across 703 live fills we have held $0 of user funds, because no code path exists that could. The same fact is the regulatory posture: no custody, no pooled capital, no discretion, and the strategy belongs to the trader. We sell software that executes their instructions.
Hyperliquid is the beachhead.
One venue, bottoms up
Both inputs are guesses and we have never charged anyone. Read it as the beachhead floor, one venue at 3%, not the ceiling. The surface is the same product across perps, prediction and equities.
Perps. Orders, brackets, agent wallets, risk limits.
Perps. Robinhood Chain’s official partner venue.
Event contracts. A thesis in English suits them well.
Builder DEXs put equities on the same margin rails.
Venue is one field on the spec, so a new venue multiplies the same product instead of forking it. The honest caveat: our first venue took five weeks and the second is not live yet, so we do not yet claim a repeatable integration cost.
Plain English is not the moat.
Natural language strategy building shipped everywhere this year, including a 22,000-star open-source repo. Green marks whoever actually wins the row, including when it is not us.
| Plutus | Vibe-Trading · OSS | Composer · SoFi | HyperAgent | Grid bots | |
|---|---|---|---|---|---|
| Plain English in | Yes | Yes | Yes | Presets only | Configs |
| Statistical gate, not just a backtest | Seven checks | Backtest only | Backtest only | None | None |
| Nothing to self-host | Hosted | You run Docker | Hosted | Hosted | Hosted |
| Funds never leave the wallet | Signed, trade-only | Your API keys | Broker custody | Trade-only keys | API keys |
| Record a stranger can verify | On-chain fills | Local log only | Platform-reported | Own claims | Self-reported |
Every rival wins at least one row and none wins all five, so the intersection is a wedge today. The one row a rival cannot copy, the record a stranger can verify, is also the seed of the moat: kept and labelled across every run it becomes a dataset that a self-hosted tool keeps private and a single-venue product never sees. That is what compounds.
Three ways this could earn. None tested.
No revenue and no price in market, so this is a decision tree rather than a price list.
Predictable and simple to sell, and the band is already visible: a competitor charges near $56 a month.
Caps revenue per trader no matter how much capital they run, and a rival set that ceiling first.
Scales with capital deployed rather than seats, and can cost the trader nothing.
Depends on rebate programmes we neither set nor control.
Best alignment with the trader, and by far the largest ceiling per account.
Moves us toward regulated advisory. Not touchable without counsel.
The first two need no licence, so both go live with the design partners. The third is the version we want and the one we will not touch until we are told we can. Marginal cost is compute in all three cases, so the gross margin is a software margin, though we have not yet measured cost per active trader.
How the first thousand traders arrive.
None of this is tested: we have run no campaign and hold no waitlist. It is a plan with one clear failure mode, stated below it.
Design partners
Ten Hyperliquid traders we onboard personally and then watch use it. Pricing on from day 60. The output we want is a list of where they stall, not revenue.
The public record
Every live strategy gets a shareable page backed by its on-chain fills. Traders argue about performance publicly already, constantly, and this hands them something falsifiable to argue with.
The venues
Builder rebates mean a venue earns when we route volume, so their incentive to distribute us is direct and their traders are exactly our audience.
The failure mode is symmetric: if the published records are bad, the same loop runs in reverse. That is the argument for keeping the gate strict even when it costs us a signup, and for never publishing a record we have tidied up first.
Eight years operating capital-at-risk infrastructure.
Founder and CTO of Vitwit. Cosmos ecosystem contributor since 2018.
Vitwit, founded 2018, is an engineering company with an existing bench, payroll and clients, so Plutus was built by a staffed team rather than between other jobs. Core development and infrastructure for Cosmos, Akash, Polygon, Ondo and 60+ others.
What transfers, and what does not
- ·Signed authority over other people’s assets is our day job. A validator is the same shape of problem: delegated keys, hard limits, real financial penalties for mistakes
- ·We build on the chains these venues live on, including dYdX, Injective and Osmosis, so execution is familiar ground
- ·What does not transfer: we have never sold to traders, and there is no quant researcher on the team yet. That hire is line one of the use of funds
What is actually true today.
- ·Rust engine with one evaluator shared by backtest, paper and live
- ·1.5M candles for BTC, ETH and SOL across seven timeframes, hourly back to Jun 2024
- ·Live end to end on Hyperliquid: agent wallets, orders, brackets, enforced limits
- ·55 strategies written, 42 evaluated, 6 cleared the gate
- ·703 live on-chain fills across 4 strategies, $60K notional, since Jun 17
- ·Gate-cleared strategies now trading live, zero custody incidents, zero funds held
- ·No external users. Seven accounts, all internal. No waitlist, no revenue
- ·The live track record is net-negative after fees, and backtests model funding but not fees or slippage
- ·The gate is not yet a hard precondition for deploying, the gap this round closes first
The sharpest version of that last column: our demo scalper showed +100% gross across 577 fills and roughly −159% once round-trip fees at 5× are applied. It looked like the best thing we had built. Modelling costs is day one of the plan for exactly that reason.
Fees and slippage into the backtest, and the gate wired as a hard precondition for deployment.
Hand-recruited from the Hyperliquid community, onboarded by us. Pricing on.
Predicted against realised, in public. Ten traders live 30 days, three paying.
$1.5M pre-seed. 24 months.
SAFE, valuation to discuss. Built so far on Vitwit revenue rather than venture money, and at our cost base this funds a team of seven for the full 24 months.
Use of funds
What this round has to prove
- ·The gate predicts: gated strategies beat ungated ones, net of costs, on live fills
- ·$1M ARR, roughly 1,200 paying traders at an assumed $75 blended
- ·Traders renew: retention past 90 days on live strategies
- ·Two more venues live, and a third-party custody audit
The real risk, and the only item here that is measurable rather than arguable. Publishing predicted against realised monthly is how we find out in public and early.
We sell discipline rather than returns, and we measure day-30 live retention rather than signups. The first ten partners tell us inside 90 days.
Non-custodial, no pooled funds, no advice, no discretion. Counsel engaged, and a third-party custody audit is funded inside this round.
Every run leaves a labelled record.
Each strategy writes one row that nobody else is positioned to collect: what the trader believed, what we predicted, and what actually settled on-chain. A self-hosted tool keeps its runs private and an exchange sees only its own venue, so neither can accumulate this.
Enough of those rows and the gate stops being seven thresholds we chose and starts being a model of which ideas survive contact with a market. That is the asset, and today we have 42 rows of it.
Then the marketplace
Once records are verifiable, a trader can publish a strategy and others run it with their own funds, in their own wallets. Not signals, not a fund. It is unbuildable until you can prove which strategies work, which is the part we are building first.
Almost none of them can code.
Fewer still can tell a good idea from a lucky one. That gap is the product.