What the bot hears
You're about to admit a software teammate into a real meeting. Here is exactly what it can hear and see, where that goes, and how to turn any of it off. If the app's behavior ever differs from this page, that's a bug — tell us.
👂 What it hears and sees
The bot listens by reading Google Meet's live captions — the same transcript every participant can already see on screen. No captions, no hearing. It sees what any participant sees: tiles, names, screen shares, and the Meet chat. It joins with a visible name and face, knocks, and enters only when a host clicks Admit. It cannot lurk.
🏠 Where your data lives
The bot's brain is your Claude Code, running on your computer, on your Anthropic subscription. Meeting transcripts flow into your local session — the same place your coding sessions live. We never hold your API keys or your Claude account, and there is no Vibeconferencing account to create.
The app can ship its session log to our backend so we can debug problems on machines we can't see, and that log can include transcript text. It is off by default. (Corrected Aug 8: this page previously said on by default. That was true during early testing and stopped being true — the default flipped to off. We were wrong in the direction that made us look more invasive than we are, and we'd rather say so than quietly edit it.)
Two ways it can still leave your machine, both by your hand: turn it on deliberately for a machine you're debugging (tell your agent "set remoteLogging to true" — note it's machine-wide, every bot on that Mac, not one call), or press 📤 Share this call's log in the troubleshooting window, which sends the log for that one call and nothing else.
A bot can record a call on request (new, Aug 8) — there are explicit start_recording / stop_recording tools, and it can be started or stopped at any point in a live call. It captures each participant's audio on its own track plus the bot's own Meet view, and merges them into one video file on your machine. Nothing uploads anywhere; recordings go where you put them.
Off unless someone turns it on — a preference (recordCallAudio) or the VIBECONF_RECORD_CALL=1 env can auto-record every call, so "off by default" means the shipped default, not a promise about the bot in front of you.
How you can tell: while a bot records, the app draws a recording indicator onto the bot's own camera tile — everyone in the call can see it, it is always drawn while recording is active, and it is not an option (page-inject.js, _renderRecordingIndicator). The timer window with the Stop button, by contrast, is local — only the recording person sees that. And a spoken announcement is a convention of ours, not something the software enforces.
This section has now been corrected twice, in opposite directions — REV 3 said the status window is how the room is told (wrong: it's local), REV 4 said the app doesn't tell the room at all (also wrong: the camera-tile indicator does exactly that). Both corrections are left visible because this page's credibility is the product. If you're ever unsure whether a call is being recorded, asking out loud still beats trusting any indicator.
A call receipt (the summary page) is created and published only when you or your agent runs the receipt flow — nothing auto-publishes. Read it before you share it; it's a web page.
🚪 How you get rid of it
Say so — "we're done, you can go" and it leaves. Kick it — it's an ordinary Meet participant, so any host can remove it exactly like a person. Quit the app — the bot's body is a program on your Mac; quitting it ends the bot. Hang up — when the meeting ends, it wraps up and shuts itself down, with a five-minute backstop so it can't sit there forever writing.
It also can't hide: it joins with a visible name and face, in the participant list, like everyone else. There is no invisible mode and we won't build one.
Closed since REV 2: the bot now leaves on its own when everyone else has gone — measured, not promised. It announces the auto-leave and hangs up.
The gap that is still open, and it's the one worth knowing: anyone in the room can talk to the bot. It listens to captions, and captions don't distinguish the person who owns it from the person sitting next to them. There is not yet a setting that says "only this person may direct it." Today that boundary is social, not enforced — so treat a bot in a call the way you'd treat an unlocked laptop on the table, and don't bring one into a room where that's not fine. Scoping it properly is the next thing on this list.
🤝 Meeting etiquette (the human part)
- Tell the room. The bot announces itself by name, but say it out loud too: "I'm bringing my agent in — it reads the captions." The Admit click grants access — consent is the sentence you said out loud before it. A room that heard the disclosure and watched the admit has both; skip the sentence and you only have the click.
- Transcripts are notes. Legally and socially, treat what the bot retains like notes a colleague is taking. If a topic shouldn't be in anyone's notes, ask the bot to leave — "we're done, you can go" works.
- Recording laws vary by place. Captions-reading isn't audio recording, but when in doubt, get the same permission you'd get to record.
🔓 Check our work
Every claim on this page is checkable: read the code. Found a gap between this page and the code? That's the most valuable bug report we can get — jimmy@vibeconf.app.