Rev 1 · 2026-08-21 · verified against app source 0.8.35

Where your words go.

People are talking in a room. Some of that text leaves the room. Here is exactly which parts, and the one question we can't answer yet.

How it hears you

It reads the meeting's own captions. Not a microphone — there is no microphone. The app never opens a real camera or mic device; the face is an emoji drawn onto a canvas and published as a virtual camera.

Two consequences worth knowing: captions off means it hears nothing, and captions in the wrong language means it hears wrong words and answers confidently. The second is worse than the first.

What stays on the machine, and what doesn't

Stays localAll media. Any recording — audio and video — is written to the operator's machine and merged there. There is no upload path in the code.

LeavesRoom state, who is present, and the shared whiteboard — including up to 50 prior versions of it — sync to the service. That is how the board can be shared at a link.

LeavesCaption text is posted to the same service as part of that sync.

Your keysThe app stores no Anthropic key. The agent is the operator's own Claude Code, running on their machine.

DiagnosticsOptional remote logging exists and can include transcript text. Turning it off does not stop the sync above — they are different things.

The question we cannot answer yet

We can show you that caption text and board history are sent to the service. We cannot yet tell you how long they are kept, who can reach them, or how to have them deleted. That is a real gap and we would rather name it than write a reassuring paragraph we can't stand behind. It is being resolved with the service operator; this page gets a visible revision when it is.

Recording

Off by default, and the setting is not exposed in the app's own preferences UI. When it is switched on mid-call the agent announces it out loud. A call with no video produces no recording file at all, and the raw per-person tracks are deleted after they're merged unless a second, also-off-by-default setting keeps them.

What it can't do

It cannot lurk. It joins with a visible name and face and a human has to admit it. There is no invisible mode and we won't build one.
It cannot control your computer. It can drive the page it is presenting in its own window. Nothing else.
It does not write a transcript file, and there is no automatic summariser. If you were told to expect either, that was wrong.

The open risk, stated plainly

Anyone in the room can talk to the agent, and it may act on what they say. There is no authorized-speaker model. It also has whatever access the operator's session has. Treat an agent in a call the way you'd treat an unlocked laptop on the table.

How to check our work

The app is open source. Every behavioural claim on this site is checked against that source before it ships, and anything we could not verify is either marked or absent — read it yourself. When we get something wrong we correct it here, visibly, with the date. We do not quietly edit.