Ask the assistant to research something and you get a proper write-up rather than a chat reply: trend and structure off your own live data, what is actually happening to the asset with sources you can click, a trade setup with levels, and what would make the whole thesis wrong.
“Research SOL” is enough. Name an exchange if you want a specific market; otherwise it uses whatever is on screen.
It takes tens of seconds. You can minimize the window and keep working.
How it is built
It gathers. 200 daily and 200 hourly candles for the asset, straight from your own connector, with signals computed over them. If you have a web-search provider connected, it also searches for current material. If not, the report is built from market data alone and says so.
It writes. The report streams in, grounded in the market data your terminal already holds.
It comes back to the assistant. The report is a tool result, not a dead end, so the assistant summarises it in the chat and can act on it: draw the levels the report named, set an alert at the invalidation, add the pair to your watchlist.
Live data wins a disagreement. Where a web source conflicts with the price your connector is streaming, the report trusts the live data and notes the discrepancy. That is the failure mode a research tool most needs to get right: a confident report built on a stale quote is worse than no report.
What is in it
The report lands as a card in the conversation, collapsed, with the assistant’s own summary underneath it, which is often enough on its own. Expand it and each section is laid out rather than dumped as text.
Executive summary. The verdict, bullish, bearish or neutral, as a badge, then the reasoning.
Price action and structure. Trend, volume, moving averages and volatility, above a small chart of the last few months with the support and resistance the report named drawn across it. Those levels repeat as chips, so you can read them without reading the paragraph.
Catalysts and developments. What is actually happening, with numbered citations wired to the pages they came from. For stocks that covers earnings, revenue and guidance; for crypto, protocol and ecosystem events.
Market context. The wider backdrop: policy, rates, the dollar, inflation and employment where relevant.
Trade setup. A card rather than a paragraph: bias, entry zone, invalidation (your stop), targets, and reward against risk. When nothing looks attractive, it states the conditions that would create a setup instead of inventing levels, which is the honest answer most of the time.
Risk factors. What would make the thesis wrong. Read this one first if you already like the idea.
Sources. Numbered to match the citations. The model is forbidden from inventing or altering a URL, so a link that does not resolve is a bug worth reporting rather than something to expect.
Reports run 1,000 to 1,800 words and end with an explicit note that they are informational and not financial advice.
The little chart inside the report is a picture of the report, not your chart. To get the levels onto the real one, say “draw those levels” and the assistant puts support, resistance, entry, stop and target on it as drawings you can drag or delete.
What it costs
Research is the most expensive thing the AI does: a long generation plus web searches. On hosted Intelligence it draws from the same credit budget as everything else, with each web search costing a flat amount on top of the tokens. On your own provider key, it costs whatever your provider charges.
Ask for one per asset per session, not one per idle moment.