Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.
Realtime Voice
The browser Realtime path is enabled by default and preserves the existing HTTP
dictation queue as its fallback. The default model is
qwen3-asr-flash-realtime.
The browser sends start, 250 ms PCM16 audio, and finish frames over the
avibe-asr-v1.<cloud-token> subprotocol. The cloud route authenticates the
existing asr capability, forwards audio to the configured DashScope realtime
endpoint, and returns non-authoritative previews followed by one final text.
The preview is transient editor state; only the server-cleaned final result is
committed to the draft. Provider credentials remain server-side. Handshake,
transport, protocol, timeout, and upstream failures activate the existing HTTP
recording fallback without changing legacy or dictation endpoint contracts.
Cloud token
The browser opens the socket with the token from GET /api/cloud/token. A
remote-access browser receives a token for its signed-in avibe.bot user,
minted through POST /api/v1/instances/{id}/user-token. The Workbench opened
on this machine's own loopback origin (a local browser or the desktop shell)
has no avibe.bot session, so a direct, same-origin loopback request receives a
token whose principal is the paired device, minted through
POST /api/v1/instances/{id}/device-token. That token grants nothing the
device secret does not already grant. Proxied requests, LAN setup hosts,
requests without same-origin evidence, an unpaired instance, and a refused or
unavailable mint get 503 cloud_unavailable, and the browser keeps the local
HTTP relay.
Reply context
When dictation starts in an existing chat, start may also carry reply: the
full body of the chat's latest Agent result, without Avibe's generated
footer. The cloud extracts recognition hotwords from it; the browser neither
extracts nor truncates it, and never logs it. The reply is read once, when
dictation starts and before the microphone opens. A result row arrives whole, so during a running
Turn the previous reply is used. The field is omitted when the reply is empty
or longer than 200,000 characters. It is also omitted when the transcript
window does not reach the live tail. The Workbench home, the new-session sheet,
and Show Page dictation have no latest reply and never send it.
The start schema is strict, so an unknown field closes the socket. The
browser therefore sends reply only when the cloud token that opened the
socket declares realtime_reply_context. avibe.bot returns capabilities with
user-token and device-token. GET /api/cloud/token relays it as a
bounded list of strings. A backend that predates the field declares none.