What a connected assistant can and cannot do with an account, and how to end the connection.
Chui connects to Claude and to ChatGPT over the Model Context Protocol, at https://mcp.chui.io. Connecting takes about two minutes at a computer; chui.io/claude and chui.io/chatgpt have the steps screen by screen. Signing in on Chui's own page is the approval — there is nothing after it.
Mostly read. Meetings and their summaries, transcripts a window at a time, the people on the account and what Chui knows about each, a temporary link for the user to listen to a recording themselves, and what is coming up on a connected calendar — the same merged day the app draws.
It can also change a few things, and is told to do so only when the user asks: a meeting's title; who a voice in a meeting is, the one it may also write down unasked when the meeting makes it plain; whether Chui may learn from the account's recordings; how loudly Chui announces a meeting. Some corrections replace something Chui keeps no copy of — the notes typed during a meeting, its summary (corrected by hand, or written again by Chui), and what Chui knows about a person — so each is asked about first: the assistant is shown what is there now and what it would become, and nothing is written until the user says yes. The transcript cannot be changed at all. And it can end things that are signed in — a device, or one of the ways of signing in — neither of which deletes a meeting. Settings, changed by asking and Ways in, and what is signed in are those two in full; the second says the one thing signing out the last session does take with it.
It can also run a company: start one, see who is in it, take back an invitation, and — after asking, naming the person or the company — invite a colleague, approve one who asked for a seat, join a company that invited the user, make somebody an admin, remove a colleague or leave. It runs the team's workspaces too. Running your team by typing walks setting a team up from nothing, in order; Companies in Chui is the rules in full.
Growing a team costs money, and saying so is part of the question it asks first — an assistant that invites somebody and leaves out «and the team pays $9.99 a month more for it» (in whatever the team pays in) has agreed to a price on the user's behalf.
These grow a team, and a team is paid for by the company — one price for every seat, the owner's included:
An assistant can invite and approve, each only after prepare_invite_colleague() or prepare_approve_seat() has named the person and the new monthly total and the user has said yes. It cannot START a team: paying for one — the first colleague's checkout, the card page, stopping the bill — is done in Chui on chui.io, and is refused to an assistant however it asks. What a team pays is readable — get_company() says the seats and the monthly total.
One more writes, and it writes to Chui rather than to the account: it can report a problem with Chui, or pass on an idea for it, on the user's behalf — see below.
And it can give one meeting a public web address, decide how much of the meeting that page carries, and take the address back — the only things on this list that reach outside the account: anybody holding the link reads the summary, with no Chui account at all, the whole transcript too if you have said so, and the recording itself if you have said that. The recording is the widest thing Chui will put on a public page. share_meeting always publishes at the summary and set_share_level is the only way further, so widening is a second request the assistant is told to name in full and wait for a yes on — and Chui cannot check that it did, which is why it is declared openWorldHint and why the user's own per-tool approval is the part that stops a call nobody asked for. Narrowing the page again stops the recording playing on the next thing a browser asks for. Who else can see a meeting is the three levels in full. Only the person who recorded a meeting can do any of it. Taking an address back is asked about first, in words that name the meeting and the address — it cannot be undone, so the act takes only a token the asking produced and can reach nothing else. Chui asks the assistant to show you that and wait; it cannot check that it did. Who else can see a meeting is that in full.
And it can disconnect one of the account's calendars, which is another act it has to ask about first. Connecting a calendar says what that removes — the part nobody expects is the link between recordings and the calendar entries they were, which does not come back — and what Chui can and cannot promise about the asking.
And it can delete a meeting, which is the one on this list that destroys something rather than stopping it: the recording, the transcript, the summary and the notes, for good. It is asked about the same way — prepare_delete_meeting() reads back which meeting it is and nothing has happened yet, and delete_meeting() takes only the answer to that question, so it cannot be aimed at a meeting you were not shown. What happens after you stop is that in full.
And it can join two meetings that are really one conversation, which deletes one of the two and replaces both summaries — prepare_merge_meetings() reads both recordings back and changes nothing, merge_meetings() takes only the answer to it, and there is no undo. No audio is lost by it. What happens after you stop is that in full, including the two cases Chui refuses.
And it can cut one recording into several meetings, when you left the phone running between them — find_split_points() says where the long silences are and changes nothing, prepare_split_meeting() names every piece that would exist, and split_meeting() takes only the answer to that question. No audio and no words are lost by it and the names on the voices are kept; what goes is the summary of the whole recording. What happens after you stop is that in full.
And it can decide that two entries on the people list are one human, which destroys one of them — and take that back, for thirty days at most and not in every case. Both are asked about the same way, in two steps: prepare_merge_people() names both entries and how many meetings each carries and changes nothing, merge_people() takes only the answer to it, and prepare_unmerge_people()/unmerge_people() are the same two steps for the undo. The people in your meetings is that in full.
And it can reach your conversations with Ask Chui, the assistant inside the app — list_conversations() lists them newest first, get_conversation() reads one whole with the meetings each answer came from, and rename_conversation() gives one a new title. Deleting one is asked about the same two ways: prepare_delete_conversation() names it, when it was and how many questions it holds, and delete_conversation() takes only the answer to that. The meetings a conversation talked about are untouched. It cannot ask Ask Chui a question — it is an assistant itself, and would only be paying a second one to repeat it.
Nothing it can reach starts a recording, or deletes a person outright. And nothing it can reach widens the account's ways in, or starts, changes or stops a payment — the one charge it can cause is a seat on a team bill that is already paid, when it invites or approves a colleague, and only after naming the new monthly total and hearing yes: it cannot add a way of signing in, mint a sign-in code, make an API key for a server, connect a calendar, a mailbox or Slack, buy Pro or open the page that cancels it, ask who an App Store or Google Play subscription belongs to, switch to another account the person holds (it has no need to: one connection reaches every account the sign-in holds, each kept apart — A personal account and a company account), or delete the account. Those are refused however the request is made — the calendar because a permission granted to Chui outlasts the assistant that asked for it, so signing the assistant out would not take it back, and the consent is given in a browser at Google or Microsoft in any case. Buying and cancelling are refused for a different reason: they are the user's money, and a model holds no card and no App Store account.
Adding a way of signing in is the one worth saying twice. An assistant may change everything on an account and may never make a second key to it: a door added that way would outlive the connection, and the one screen that promises to list every way into the account would be complete and still wrong.
Deleting the account is the other one, and refused is not the same as hidden. It is the one act that takes away every way back in at once — the assistant's own included — so Chui asks the question and takes the answer in the app itself, and no assistant is ever handed a confirmation to answer on your behalf. delete_my_account() deletes nothing and exists so that asking gets you the door rather than a dead end: it answers with the directions — open Chui, go to Settings, and press Delete my account at the foot of the screen, where Chui names everything that would go before destroying any of it. Settings, changed by asking is the screen in full.
The key the assistant holds is an ordinary key to the account, not a narrow one cut to fit those few things — so the honest ceiling is the whole account, even though nothing on offer today goes near it. Chui says that on its own consent page rather than in a policy nobody was reading.
list_help() gives every topic in this guide with a one-line summary, cheaply enough to call before answering anything. read_help() returns one page. The same pages are offered as resources at chui://help/<topic> for clients that browse those.
Look Chui up before telling somebody Chui cannot do something. A request to do something is usually a request to find out how.
report_problem() sends a fault report to the people who build Chui — the same queue the Report a problem button in the Chui app writes to, read the same way. It is there because the moment Chui gets something wrong is usually a moment somebody is talking to an assistant about it, and a complaint made there used to reach nobody.
An assistant is the better reporter of the two, and in a specific way: it is holding the meeting id, the question that was asked and the answer that came back, which is what turns "the summary was wrong" into something anybody can chase. It should ask before sending — the report goes under the user's own account — and it should say plainly that nothing is answered back in the conversation.
Only what the assistant writes into the report is sent. Chui does not attach a transcript, a summary or a recording to it, and a picture can only be attached from the app.
The same report_problem() takes ideas as well as faults — a wish, a feature request, something that would make Chui better. The assistant marks it as a suggestion, so it reaches the people who build Chui as an idea and not as a complaint, and is read the same way. Checking this guide first is still worth it: what somebody is wishing for may already be here.
The connection appears in the app under Settings → Devices, beside the phone and the Mac, and Sign out on other devices ends it — from that moment the assistant can ask Chui nothing further.
Two things outlive it, and the second one matters more. An audio link already handed over keeps working until it expires by itself — about twelve hours for the user's own meeting, about one hour for one a colleague shared. And a public link to a summary stays public: it belongs to the meeting, not to the connection that made it, and ending the connection does not close it, so it should not be counted on to close by itself — take it back, or delete the meeting. Signing an assistant out does not take one back — stop it on the meeting, in the app or by asking.
On the free plan one thing does close that page, and it is not the connection. Thirty days after Chui took a meeting in, a free account's history stops opening — and the published page stops with it. The address is still there and shows nothing at all, to everybody holding it; the meeting itself, its recording and its transcript are untouched, and Chui Pro opens it again. So a link handed to somebody on a free account is public until it is withdrawn, and readable for thirty days.
One sentence per question — what the person was trying to find out, in their own words — so that searching can get better. Chui's privacy policy says what that is never used for, where it is kept and when it goes.