Introduction
Velox is a multi-exchange consensus trading terminal. Three AI models — Claude, Gemini, and Grok — analyze eighty perpetual futures contracts every ten minutes across six venues. The consensus engine opens virtual trades when at least two agree; an in-app co-pilot grounds its proposals in those signals against your real portfolio.
The one-sentence explanation
Most signal tools try to be right. Velox tries to be wrong about the same thing less often — by requiring that two independently-reasoning models agree before risking a single dollar, and by putting you, not the AI, on the trigger.
Two halves, one terminal
Velox runs two independent trading worlds that deliberately never write to each other's state:
- Signal engine — four cron jobs (scanner, monitor, evaluator, optimizer) analyze the market, open virtual trades in a dual portfolio, and self-optimize against shadow P&L. No real exchange is ever touched.
- Trading terminal — sign in once, link exchange API keys (Binance, Hyperliquid, Aster today; Bybit / OKX / Lighter read-only), and trade from an order ticket with a positions drawer. An AI co-pilot can read your real book and propose trades grounded in Velox signals — it cannot place them.
Signal flow is one-way: cron signals land in a co-pilot context, the co-pilot cites them when you ask, and you are always the gatekeeper who clicks Submit.
Who this is for
- Discretionary traders who want a second and third opinion before pulling the trigger — and a single terminal to execute across venues
- Systematic traders who want a live demonstration of multi-model consensus before building their own
- Researchers curious whether adversarial model disagreement filters out the kind of bad ideas that single-model LLMs routinely produce
What Velox is not
- A black box. You can watch every model's reasoning in the terminal in real time.
- An autonomous bot that trades your account. The AI proposes, you dispose —
propose_ordertools populate the ticket for review; the server never exposes order placement as an AI-callable tool. - Custodial. Hyperliquid and Aster are traded through agent wallets whose signing policy allows place / cancel / modify, never withdrawals.
- Financial advice. See the risk notice before using any output for real-money decisions.
Trade modes
Every real-order path is gated by a TRADE_MODE setting enforced at the adapter boundary, not just the UI:
| Mode | Behavior |
|---|---|
| paper (default) | Orders are synthesized locally at last trade price; nothing hits the network. |
| testnet | Orders hit each exchange's testnet. Useful for end-to-end verification. |
| live | Orders hit mainnet. Requires a deliberate, separate deploy flip plus a second gate (CREDENTIAL_SCOPE_LIVE_ALLOWED). |
How the consensus engine works
Every 10 minutes, a scanner pulls OHLCV, funding rates, open interest, and liquidation data for eighty perpetuals. The same normalized brief is sent to each of three models. Each model returns one of:
LONG— expects price up, produces entry, stop-loss, take-profitSHORT— expects price downNO-TRADE— insufficient conviction; abstain
The consensus tiers:
| Agreement | Tier | Behavior |
|---|---|---|
| 2 / 3 | Silver | Open position at half-weight |
| 3 / 3 | Gold | Open position at full Kelly-sized weight |
| < 2 | — | No trade |
Disagreements are logged with the full reasoning trace. Over time, the optimizer reads closed-trade P&L and adjusts each model's confidence weight — models that are often wrong when the others disagree with them get downweighted.
Start here
- Quickstart — get the terminal running in under a minute
- Architecture — the two trading worlds, the cron pipeline, the exchange adapters, the Redis/Postgres split
- API reference — endpoints for programmatic access
- Security — credential encryption, agent wallets, and what we never hold
- Risk notice — what you should read before trusting any output
Live demo
If you just want to see it work before reading further: watch a live scan →. Real Binance prices, simulated AI consensus, no account needed.