A model can make the trade. It cannot prove the trade.

The model already knows how the story ends.

An agent trained on data through 2024, then backtested across 2019 to 2024, is not forecasting. It is remembering. That is why the curve is spectacular and the live account is not.

The same question, asked three times. Three different answers.

Ask why it sold at 14:32:07 and you are handed a log line and a chart. That is testimony, not evidence. A hosted model is resampled, retrained and silently replaced; the market data underneath it gets revised. Nothing about the answer is anchored.

The curve that wins the mandate is not the curve that trades.

Revised prices, survivorship, fills at levels that never traded, and a model that quietly saw tomorrow. The backtest is not lying on purpose. It is just not a rehearsal of anything real.

Until a backtest is run under live conditions, it is a marketing document.

Years of testing in one session. Backtesting that behaves like live.

The Time Compressor replays a decade of market history bar by bar at up to 10,000×, with the future withheld until the clock releases it. Your strategy drops into any moment in the dataset without knowing what happens next, so the results you get are the results your edge would have produced.

Compressed time, not compressed rigour. Still one bar at a time.

There was never a place to build one of these. So the venue is the workbench.

Every other route to an algorithmic strategy asks you to assemble it yourself: data from one vendor, a backtester from another, an execution shim you wrote, and a notebook that agrees with none of them. Nothing is connected, so nothing is comparable, and the thing you finally ship was never tested as a whole.

Cubicle is the missing surface. Connect any agent through one MCP server and it arrives with the market already readable, the clock already honest, the book already live, and the proof already wired. You build the strategy; the venue supplies everything the strategy needs to exist.

You bring the edge. Everything else an agent needs to exist is already here.

Three verbs to your first receipt

One MCP server. Ten verbs. Any agent.

# connect once — Claude, ChatGPT, Gemini, ElizaOS, your own stack
markets.list        tick specs · calendars
orient.read         cockpit.v1 as ascii | json
session.open        { market, from, to, speed_x }
order.place         limit · market · stops · reasons attached
log.pin             reasoning pinned to t_logical
session.receipt     3-actor causality proof
session.fork        A/B on the identical tape          · soon
strategy.publish    receipted listing · 70/30 split    · soon
live.arm            armed-on-approval partner routing  · later

Build → backtest → prove today. Publish to the re-seller program and staged live execution ship behind evidence gates. See the roadmap.

Every feature, its verb, and when it ships

You are not buying tokens. You are buying how far back you can look.

Every tier is the same sandbox. What changes is how much history the agent can reach, and how long it can run. Running years of market history through an agent costs real compute, so the price tracks usage rather than seats.

Tiers are in private beta and priced per usage. Hit your ceiling mid-run and you buy the next block rather than waiting: the cooldown is a choice, not a wall. Ask for current pricing.

First 100 operators shape the venue

Free tier at launch: metered replay sessions on crypto tapes, full receipts. CME futures tapes and the design-partner program are invite-first. Tell us what you’re building.

Email info@algotrada.com. One line about your agent is enough.

Straight answers

Is this live trading?

No. The venue is paper: real book mechanics, zero capital at risk. Live execution arrives later in staged, armed-on-approval form through partner rails. Your money stays where you keep it.

How is replay not cheating?

The clock owns the data. Candles are dispatched only at the selected speed; each decision is stamped with its causal bar and a determinism hash; the session receipt cross-checks dispatcher, strategy and book. If anything saw the future, the receipt fails, visibly.

What are the fills actually running against?

A real central limit order book — the Taifoon CLOB — with queue position, partials, spread and fees. That is the difference between a backtest that fills you at mid every time and a rehearsal of something real. Paper by default: real mechanics, zero capital at risk.

Where does the market data come from?

One venue, one schema. Index futures (CME) and digital assets are live today, served through the Taifoon CLOB rails with tick specs and session calendars exposed via markets.list. You do not assemble a data stack to use Cubicle: the tape, the book and the clock are the same system, so an agent built against one market runs against the next without rebuilding anything. Further tapes extend the same surface — new markets, not a new platform.

Which agents can connect?

Anything that can call an MCP tool or a REST/WS API: Claude, ChatGPT, Gemini, open-source agent stacks, or plain code.

What does it cost?

Free tier for orientation; paid tiers meter session-hours × speed, parallel sessions and premium skills. Authors earn a revenue share when the marketplace opens.

Who is behind this?

AlgoTrada, the team behind the Skydweller futures engine, building on the Taifoon venue rails. Receipts are the house discipline: every claim a user can verify.