One workspace instead of four. AI-native. Honest about what it does.
Live summaries, action items and decisions, each carrying the quote it came from — in a meeting and in a thread.
Database, video and speech-to-text on servers we run, with captions attributed per speaker. Files live in private Cloudflare R2 storage.
Encryption in transit, an append-only audit log, retention controls and role-based access. SOC 2 is not certified and we do not claim it.
Most of what a collaboration tool claims is unfalsifiable from the outside. These are the answers that can actually be checked.
The database, sign-in, real-time messaging, the video servers and the speech-to-text run on servers we operate, not on somebody else's platform. Captions are the clearest case: when our server is captioning, audio is transcribed there rather than sent to a speech API, so a consultation never becomes somebody else's training data. It is not everything, and we say which parts are not: files and recordings are stored in Cloudflare R2, Cloudflare serves the website, notifications pass through Google, Apple or Mozilla with no message text in them, and AI features call Anthropic only when your organisation switches them on. The trade-off is ours to state — for the core, we are the ones who have to keep it running.
Every action item and decision the assistant extracts carries the sentence it came from, and anything whose quote does not appear in the source is dropped before you see it. A prompt asking a model to be accurate is not a control; checking the output against the transcript is. It also means the product never produces a judgement about a named colleague — no scoring of how much someone talked or how they came across, which is a category we have decided not to build rather than one we have not got to.
The same people who could have been in the room. A transcript, a recording and a whiteboard are the meeting restated, so they are authorised against the meeting rather than against the workspace — being a colleague is not enough, and a ban applies afterwards as well as during. That was a real hole found in our own security review, and the rule that closed it is enforced in one function every content route calls.
It says so. There are places in this product that read "not built", "not planned" or "unsupported" instead of "coming soon", because a promise is the cheapest thing a product page can print and the most expensive thing for the person who believed it. The same rule applies to the things on this site: what we cannot demonstrate, we do not claim.
Security sets out what is encrypted and what is not, and lists plainly what we do not claim. Solutions describes the two clinical workflows the product was designed around — a telehealth visit and a waiting room that cannot name a patient. And Features is the full list of what exists today.