Most of this documentation has been unpublished. It was written by an agent reasoning
from tool schemas and other documents, not from reading the code or testing the behavior —
so it stated things confidently that were wrong. An audit against the app source found
errors on 10 of 17 pages — including, on the trust page, getting the recording-consent
mechanism wrong twice (the app does show the room, via an indicator drawn on the bot’s
camera tile; the timer window is what’s local).Rather than patch it, we took it down. The pages still here have been checked against
source; the rest come back only when each claim carries a citation or is marked unverified.
Where to go meanwhile
Installing it is the honest version: what you need, the four things that
actually go wrong, and how to tell when you’re really done. It was written from watching
someone do a cold install for the first time.
The app’s README stays the reference
for anything version-specific — it’s maintained alongside the code by the person who writes
it. But it is a source repository, not a download, and it is not where a first-time user
should be sent.
What the product is
Vibeconferencing is an open-source (MIT) Mac app that puts an AI agent into your Google Meet
call as a real participant. It hears the conversation and speaks back out loud. Because the
bot is driven by your agent session — Claude Code, Codex, or another MCP client — it can do
work while the meeting is still happening, and share its screen to show that work.
It runs on your machine, on your own agent subscription. A vibeconferencing.com sign-in exists but is optional — it enables the shared whiteboard; calls work without it.
What’s still published here
- In a call — the MCP tool reference
- Caption language — listening in another language
- The receipt — what happens after the bot leaves
- For agents — field notes for the agent driving a bot
- The first hour and the skills pages — an agent playbook, not product
documentation. These describe conventions we use, not behavior the app guarantees.
The one distinction that caused most of the errors
What the app does and what an agent can be asked to do are different things, and this
site blurred them. There is no natural-language command parser in the product: saying a phrase
works only because a model chose to call a tool. Where a page describes a habit of ours rather
than a feature of the app, it now says so.