Everyone else does AI alone. Storm is the one your whole team is in.

The only AI workspace your team can actually be in together.

Invite colleagues into a conversation. You see who is reading, and where — a small mark next to the answer they are looking at, and a breathing one when they are typing. Anyone can leave a comment on a specific answer; comments stay attached to the turn, get resolved, and stay readable six months later when someone asks why the plan changed.

Twelve people at most. Beyond a handful of steering hands it is not a conversation any more, and a presentation is a different thing — that one is just a shared link.

An answer is something you can take apart.

Every chat product treats an answer as a delivery: it arrives, you read it, you ask the next thing. That works for questions with an answer. It works badly for questions with options — and those are most of the interesting ones.

So the rule underneath everything here is a different one: an answer is something you can inspect, steer and reuse — not something you receive. Every feature on this page is that sentence, applied.

Not because the model cannot. Because the shape prevents it.

This is not a list of things the big chat products do badly — they are extremely good at what they are. It is a list of things a scrolling conversation cannot hold, which is a different and more durable thing. The same limits apply to any chat window, including one of ours.

A conversation is a tree, and you can see it.

Go back to turn four, change one word, and the new attempt appears next to the original instead of replacing it. Ask the same question two ways and keep both. On the right of every conversation there is a small map of what you have tried, and clicking a point puts you back there.

This is one decision, not four features: every turn knows its parent. Branching, rewinding, comparing and deciding what the model gets to see all follow from that — which is also why none of them cost anything extra.

And when a conversation has no branches, there is no tree. Showing complexity that is not there is exactly what this product is against.

A map of the model’s knowledge, with a switch on every dot.

Beside every conversation sits the list of what feeds it: files you added, facts you stated, and pages from your project wiki. Each one is a dot you can switch off — and then it genuinely stops travelling along with your next question. That is the whole point: not an explanation of what the model used, but a steering wheel.

Turnleaf pages are the interesting one, and it goes both ways. A company that keeps its Turnleaf wiki up to date has already paid for the context of every Storm conversation — the model can talk about what your company already wrote down, instead of guessing at it.

And when a conversation produces something worth keeping, one click writes it back as a draft page in the project wiki — with a note in its history saying which conversation it came from. A draft and not a published page: an answer is not documentation until a person has said so. What you no longer do is copy and paste it into a second tool.

When it does not all fit, the sources that were left out are struck through with the reason. A map that shows six sources while four went along is worse than no map: then you think you are steering while you are looking at a picture.

Two answers, two angles, and you pick.

One button gives you two independent answers to the same question — deliberately from different starting points, not the same reasoning in other words. You choose. The other one stays, in case you were wrong.

It costs twice as much, and it says so on the button. We would rather you use it when it matters than discover the bill afterwards.

Four things a chat does not tell you.

X-ray, or heatmap. Never both.

X-ray answers can I believe this? Every part of the answer is underlined by where it came from: your material, derived from it, or guessed. Guessing is allowed — hiding it is not. It is not a fact check against the outside world, and it does not pretend to be: the model is saying this about itself.

Heatmap answers what should I look at? The sentence that carries the point, the things you actually have to do, and the places that could bite you. So you can scan a long answer instead of reading it.

Two lenses, one switch — never both at once. Two colour systems over the same paragraph and neither says anything any more. And a lens that would mark most of the answer switches itself off: marking everything marks nothing.

Ask for a comparison, get a sortable table.

An answer that is a comparison becomes a table you can sort by any column. A process becomes numbered steps. A plan becomes a timeline. No mode to switch into and no second window — the answer arrives in the shape it wanted to be, and it says which shape that was so you can change your mind.

The text always survives. A shape that cannot be recognised falls back to well-set prose: the form is a bonus, never a condition. That is also why the model writes ordinary markdown rather than a rigid structure — a missing comma should never cost you an answer.

Prepaid credit, and a ceiling per conversation if you want one.

Storm runs on the same credit as the rest of Turnado: paid in advance, never billed afterwards, and never a charge you did not see coming. Every answer shows its own price, and you can cap a single conversation at €1.00, €5.00, €20.00, €50.00 — after that it stops and asks, instead of quietly continuing.

A ceiling is not a second meter. Your account credit and the pace limits sit underneath as they always do; this is a voluntary line inside them, for the conversation that turns out to be bigger than you thought.

When the thinking is done, hand it to Ops.

One branch of a conversation becomes a numbered analysis document, ready for the Turnado Ops intake — which turns it into a backlog with work items, sprints and agents that build it.

It hands you the document; it does not submit it. A conversation explores and a specification decides, and those are not the same document. Read it, cut what was thinking out loud, then submit it yourself.

What is not there yet.

Storm is new. There is no queue for very large jobs yet — "analyse these forty documents, message me when you are done" is on the list and not in the product. There is no shared template library across organisations, and no export beyond copy-paste.

Live presence works within one server. Run two of them side by side and half the people will not see the other half standing there — comments and answers still arrive, but the little marks do not. That is written down rather than hidden, and it is the first thing to fix when the load makes a second server worth it.

And the answer arrives in one piece rather than word by word. That is a choice rather than a gap: typing dots suggest a conversation partner, and this is a workbench.