The assistant & voice
A conversation with the company itself — reading the same Brain, board and ledger the fleet reads — and a voice path that runs entirely on your own box.
The assistant
The assistant panel is a conversation with the company: it reads the workboard, the Brain, the pipeline and the books through the same tools the fleet uses, and it is bound by the same gate. It can find, summarise, draft and propose. It cannot send, deploy or spend.
It can also be opened from outside the app — a URL can carry the question, or open it hands-free with the microphone already armed, which is how a home-screen widget starts a conversation without anyone clicking into a composer.
Step budgets and timeouts
Every turn has a ceiling on how many tool round-trips it may take and how long it may run before it has to start finishing. Both are configurable; both default to values a normal question never approaches.
How voice works
Both directions stay on your infrastructure. Speech in is transcribed by a Whisper model running in the appliance; speech out is served by the Kokoro sidecar beside it, reached over the container network. No audio is sent to a third party in either direction.
The sidecar is a separate container for a licensing reason rather than an architectural one: the speech stack it wraps carries obligations we will not link into the appliance's own binary. Your box pulls it from upstream; we ship none of it.
The stand-up
The stand-up is the voice path applied to the whole day: the fleet reports, you steer, hands-free. Saved chapters land in the Brain, so a decision made out loud is retrievable later like every other thing the company knows.
Running chat separately
The fleet and the assistant do not have to share a model. The fleet can sit on a slow local model all day because nothing is waiting on it; the assistant is you, waiting, with the microphone open — and a small model answers slowly and writes spoken prose in one unpunctuated breath.