AI Trading
An Anthropic language model writes four short sections on a liquid pair — handed this venue’s live numbers at the moment it is asked, and instructed to claim nothing it cannot derive from them. This deployment carries no model key, so there is nothing to ask here: the desk below states the missing configuration instead of offering a control that would refuse you.
The name on this page promises more than the desk will ever do, so the copy will not: it does not trade, does not signal and does not advise. What it does instead is rarer — it shows you its entire briefing, a scroll below.
One request, where the key is present
- Pick a pair. The desk offers 40 of the venue’s most liquid USDT markets.
- The exchange reads the live numbers. The pair’s 24h snapshot and five days of hourly candles, from the same feed the charts render.
- The model writes against them. Four short sections, grounded in those figures and nothing else — it is never asked to recall a price.
- The prose streams in, labelled. Marked AI generated, named with its pair, and followed by the reminder that commentary is not advice.
Not configured
This deployment carries no server-side model key, so there is no desk to arm and none is drawn. Nothing is sent and nothing is queued: the analysis endpoint answers with the same missing configuration rather than a spinner, an outage or an empty panel.
Everything below this describes the desk exactly as it runs where the key is set — the briefing it is handed, the three sentences it is forbidden, and the states it degrades into. The market data on this page is live either way, and the terminal it points at is unaffected.
ANTHROPIC_API_KEY — absent from the server environment
What the model is handed
Every request reads two things from the exchange — the pair’s rolling 24-hour snapshot and its last 120 hourly candles, about five days of structure — and pastes them into the prompt as plain text, under one standing instruction. The figure below is that briefing as it stands right now, for the pair the desk opens on.
The 24h snapshot
Seven lines — the pair’s rolling day, as the venue reports it.
- Instrument
- USDC/USDT (spot)
- Last
- 0.99998
- 24h change
- -0.00%
- 24h open
- 0.99999
- 24h high
- 1.0000
- 24h low
- 0.99991
- 24h volume
- 895.88M USDT
Five days of hourly candles
One OHLCV line per hour, oldest first, times in UTC.
Hourly candles (oldest first, 120 bars, times in UTC): 2026-08-25 11:00Z O:1.00021 H:1.00021 L:1.00009 C:1.00016 V:130292675.00
… 117 bars elided from this figure — the prompt carries all 120 …
2026-08-30 09:00Z O:0.99997 H:1.00001 L:0.99996 C:1.00001 V:34886745.00 2026-08-30 10:00Z O:1 H:1.00001 L:0.99998 C:0.99999 V:6076231.00
The standing instruction
“Ground every claim in the numbers provided — never recall prices, and never invent a level, a date or an event that is not derivable from the data in front of you.”
“You are describing a market, not advising a person. Never tell the reader to buy, sell, or hold, never state a price target, and never imply a prediction is reliable.”
Quoted verbatim from the instruction every request carries. It also fixes the response’s shape: four sections — What happened, Structure, What to watch, Risks — in under 350 words of plain prose.
That is the whole briefing. No order book, no funding rate, no headlines, no social feed, and none of the model’s own recollection of prices. The book and funding matter to a trader — but they are not in the prompt, so the commentary cannot honestly cite them, and the instruction forbids inventing them. If a claim cannot be derived from the figure above, the model was told not to make it.
The distinction is this page’s whole reason to exist. A model asked to remember a price is unfalsifiable — whatever it answers arrives fluent and uncheckable. A model handed a price can be marked against the number it was handed. Grounding does not make the commentary right; it makes it checkable, which on a market surface is the property worth having.
What it will never say
Three sentences are banned outright in the model’s standing instruction. On a product about money they are the three most tempting to generate — and each ban is what lets the commentary stay commentary.
“Buy”, “sell”, or “hold”
The instruction is literal: never tell the reader to buy, sell, or hold. The moment commentary ends in an imperative it stops describing a market and starts directing your money — that is advice, and a page that generates advice by the token has no business existing.
A price target
A target is a prediction wearing precision. Five days of hourly candles license a description of where the price has been and what it is doing now — the model may name levels that are in the data, and may not project one beyond it.
That it is confident
The instruction’s last clause: never imply a prediction is reliable. A generated paragraph reads equally fluent whether it is right or wrong — fluency is what these models are — so the one thing the desk must never manufacture is certainty.
And when it cannot speak, it says so
A surface that writes about money owes its reader its failure modes in plain words, not a spinner. Each of these is real behaviour, not policy prose:
- Not configured. The desk needs a server-side model key. On a deployment without one it says exactly that — a missing configuration named as a missing configuration, not dressed up as an outage. That is this deployment, and the panel at the top of the page is it.
- No live data. A pair whose snapshot and candles cannot be read gets no commentary at all. The request is refused; nothing is written from memory to fill the gap.
- Declined. If the model’s safety systems refuse a request, the response says it was declined — an empty panel that was actually a refusal would read as a network failure.
- Rate limited. Too many requests gets “try again shortly”, not a queued or invented analysis.
Everything it does write arrives labelled: the AI generated pill and the pair it was written for sit above every response, and under each one, the reminder that this is commentary — not financial advice — and that it can be wrong.
Hold it against the chart
The desk and the terminal read the same feed, so every number the commentary is allowed to use is a number the terminal draws. When it names the 24h high, that level is on the chart; when it describes five days of structure, those are the same hourly candles behind the price. The intended loop is short: ask for a read here, then go look at the thing it described — with the full book and the live tape beside it.
The balances you would trade with there are simulated, and every surface that touches them says so. The market is not. Being wrong costs nothing, and the market you are wrong about is the real one.
If the vocabulary is new
The commentary speaks in candles, ranges and volume because that is all it is given. If those words are not yet load-bearing, the academy teaches them in order — and its candlestick lesson is the legend for the exact figure this page publishes.
- Reading candlesticksWhat one OHLCV bar records and what a run of them shows — the lesson behind the hourly candles in the prompt.
- The whole academy8 lessons in four tracks, from how an order book works to sizing discipline — each ending on a live screen.
binXbase was founded in 2026 as a crypto exchange and Web3-AI company serving a global audience, and this desk is what the second half of that means in practice: not a bot that trades for you, but an analyst that publishes its inputs and knows what it is never allowed to say. The same line — real market data, simulated money — runs through the whole product, and How this works draws it in full.