Open-source Grok Bot alternative for named AI teammates. Local-first desktop, web, and mobile clients with per-bot tools, instructions, and memory. Independent fork of T3 Code.
Stored-reply playback
LEO-402 owns reading an existing assistant message on the current client. LEO-401 owns speech-provider credentials, synthesis operations, and live-call audio. This client must not add a second key store or generate a new bot turn to produce speech.
Playback identity
One playback owner exists per client. Identity is environment, thread,
message, content version (updatedAt), provider, and voice. Replacing the
identity, changing content, navigating, disconnecting, logging out, or
starting a live call cancels in-flight generation and releases the media
handle. At most one in-memory audio resource is retained.
Spoken text
replyMarkdownToSpokenText converts stored Markdown to speech text. It keeps
prose, link labels, labeled table rows, and inline code. Fenced or indented
code and images are skipped with a disclosure. Empty, code-only, and oversized
replies are not speakable. Limits are 100,000 source characters and 20,000
spoken UTF-16 code units, with no silent truncation.
Automatic readout
The preference is client-local, off by default, and stored as a boolean only. The tracker hydrates the current chat's already-loaded assistant messages, then accepts later completions in that chat. History loads, reconnects, and message edits do not replay.
Synthesis boundary
The client uses the authenticated voice.synthesize operation with the
provider and voice resolved from the environment's voice settings. ChatGPT
realtime remains a live call and must not be used as a TTS fallback. Platform adapters
(createBrowserReplyAudio, createExpoReplyAudio) accept already-decoded
audio bytes and own temporary object URLs or cache files.
Privacy
Do not log spoken text, audio bytes, or provider error payloads. Errors shown in the UI are generic recovery copy.