Rewritten on 22 August 2026. The original version of this post described an AI that returned an entry, a stop and targets, and a 📐 RR @ market button that recomputed reward-to-risk at the live price. Both are gone. NextScalp stopped shipping trade plans on 21 August 2026 - the reasoning is in Why NextScalp Stopped Shipping Trade Plans - and the AI layer followed a day later. This is what the co-pilot does now.
Most "AI trading" tools point a language model at a chart and let it narrate. NextScalp does the opposite. The automated alert stream is 100% deterministic - real price action, measured levels, real volume, no model in the live path. The AI co-pilot is a separate, on-demand layer you summon yourself when you want a second opinion on a specific pair.
It reads structure. It does not tell you what to trade, and since August 2026 it structurally cannot: there is no field in its output for a direction, an entry, a stop, a target, a reward-to-risk figure or a size.
Why the AI is on-demand, never automatic
The hard line in NextScalp is simple: a deterministic alert carries only facts, and an LLM does
not produce facts - it produces interpretation. So the model is never wired into the automated
delivery path. It runs only when you ask for it: when you type /ai, upload a chart to /review,
or tap one of the inline analysis buttons on a signal. Nothing the model says is broadcast to anyone
else, and nothing it says can fire an alert.
That separation is the whole point. The screener tells you what is happening with math you can verify. The co-pilot helps you read what that arrangement means - a structured opinion, clearly labelled as an opinion, summoned on demand.
The model proposes a price; the engine names the level
A model left to free-associate will happily invent a clean-looking level that does not exist, or attach the wrong name to a real one - calling the point of control a value-area high, for instance. Prompting against that helps and does not solve it. So the co-pilot is built so that the failure cannot reach you.
The model is handed the levels the engine actually computed for that symbol - volume-profile POC, the value-area bounds, VWAP, 5-minute swing highs and lows, daily pivots and open, live order-book walls - and it may quote only those. When it wants to put a level in front of you it emits a price and a note, never a name. The server then matches that price against its own list:
- A match inside the timeframe tolerance - tighter on fast timeframes, looser on slow ones, because a number that is "close enough" on the 4h is nonsense on the 1m - is printed with the engine's own name for that level. The model cannot mislabel it, because the model never labelled it.
- No match means the number was invented. It is dropped from the message, logged, and you get a short caveat line instead. You never read the invented price at all.
The same check runs over the prose: if the text names POC, VAH, VAL, VWAP or a pivot and puts a number next to it that disagrees with the engine's value, the read still renders but carries an explicit "treat these numbers as suspect" note rather than being quietly trusted.
The two ways you summon it
/ai - on-demand structure read
/ai <TICKER> [TF] reads a pair on a timeframe you choose (1m / 5m / 15m / 1h / 4h) and returns:
- the regime and where structure currently leans, with the reasoning behind it;
- a structure read - where price sits relative to the levels that matter and what that arrangement means;
- the levels in play, each with a one-line note on why it matters right now;
- what the data confirms, what it has broken, and what contradicts the read - including the evidence that argues against it;
- up to two conditional observations - "acceptance above this level keeps the range broken" - each with what would negate it;
- the material risks.
The 1m read is the pre-entry special case: RSI and ATR are deliberately suppressed because they are noise-dominated on that timeframe, while higher-timeframe context flows in from prior analyses of the same symbol so the 1m read stays anchored to the bigger picture.
Two conveniences make it cheap to lean on:
- One-tap analysis on every signal. Each delivered alert carries inline 5m / 15m / 1h buttons that run the same read straight from the message - no retyping the ticker. The read arrives as its own message; the alert itself is never rewritten.
- Multi-timeframe delta. Re-run on another timeframe and the bot feeds the earlier reads in as context, so the new one also reports what changed relative to what you already saw, instead of restating it. Repeat analysis of the same symbol is free and does not cost quota.
The exact syntax, default timeframe and per-run token cost live in the /ai command
reference.
/review - chart vision on your own screenshot
/review <TICKER> [TF] is the same engine pointed at an image. You upload your own chart screenshot on
any supported timeframe; Claude reads the picture, combines it with live data, and returns the same
structure read. The division of labour is strict and worth knowing: the shape comes from your image,
every number comes from exchange data. The model is told never to read a price off a picture, and the
level matching above is what enforces it. The /review command reference covers how to
attach the screenshot and which timeframes it accepts.
What it deliberately does not do
Honest framing means being clear about the limits, too:
- It never hands you a trade. No direction to take, no entry, no stop, no target, no reward-to-risk, no size. There is no field for any of it, the prompt bans the vocabulary outright, and a sentence that smuggles one in anyway is removed before the message is sent. Prevention first, redaction as the backstop.
- It is a second opinion, not a verdict. Every AI output is labelled as interpretation. It does not override the deterministic gate and it cannot promote a signal the engine suppressed.
- It will tell you it sees nothing. When the grounded read finds no readable structure, it says so rather than manufacturing a narrative to look useful. A blank is information.
- It never updates silently.
/aiand/reviewrun only when you trigger them. There is no background model quietly changing its mind in your chat. - It degrades gracefully instead of refusing. Each reply shows which model answered and how much quota remains; when the premium budget runs low the bot steps down to a faster model rather than blocking - you always get an answer.
Sizing is yours, not the model's
The original version of this post ended with a button that told you whether a plan was still worth
chasing. That button went with the plans. What replaced it sits on the other side of the line:
/size takes your entry and your invalidation and returns a position size from
the capital and risk you stored in settings. The bot supplies the arithmetic, and a warning when that
size is a large share of the pair's daily volume - never the levels. It is free on every tier.
How it fits the rest of NextScalp
The co-pilot sits on top of the deterministic machinery, not instead of it. The alert you are analysing already passed the scoring gate - scored against higher-timeframe alignment and volume - and arrives as a factual market-structure event. The AI layer then lets you interrogate that specific event: read it on three timeframes, or point it at your own screenshot. For a deterministic pre-entry view of the same symbol - levels, walls and value area with no model involved - the Focus dashboard is the companion tool.
/ai and /review are Premium commands, available during the 7-day free trial; Free users see the
alert text without the inline buttons. What the free tier ships on its own is covered in
free crypto market tools in Telegram.
The discipline is the same one that runs through the whole product: the screener never lies about what is happening, and the co-pilot never pretends its interpretation is a fact - nor dresses it up as an instruction. Together they let you watch the whole market and still make the final call yourself.
Want a grounded AI second opinion on every signal, plus 16 other deterministic alerts screened for you in real time? See how NextScalp Bot works, or try it free for 7 days.