Supervised autonomy
Safety, supervision, and limits
The self-hosted beta is designed for control under live-stream pressure. Kira should make the show sharper, not less responsible, and Agenda Mode keeps important topics human-approved. Report reproducible problems through OpenCohost GitHub Issues without exposing secrets.
- Use OpenCohost GitHub Issues for reproducible bugs and setup feedback; never include API keys, private logs, or sensitive information.
- Agenda Mode is semi-automatic: topics and direction still need human approval.
- OpenCohost adds resilience so failures degrade the show instead of collapsing it.
- Kira does not browse the web, manage files, control the PC, or automatically ban users.
- Viewers should not be able to redirect Kira into private setup disclosure.
- Performance depends on selected models, VRAM, hardware, and what else is running.
FAQ
FAQs
Is OpenCohost just a wrapper?
No. The chassis is the product value: orchestration, context management, repetition mitigation, agenda control, streamer guardrails, resilience, and live-ready coordination across models.
Can one PC run all of this?
Sometimes, yes, but the honest answer is hardware-dependent. LLM, TTS, and Whisper/VAD can compete for VRAM and compute, especially during gaming or high-load streams.
What belongs to OpenCohost and what belongs to LiveAudio?
OpenCohost contains the co-host chassis: LLM use, voice output, profiles, reactions, Agenda Mode, context, resilience, avatar, and music. LiveAudio is a separate connected app for VAD, transcription, subtitles, transcripts, and voice context.
Is Kira free?
Yes. Local-first mode avoids default cloud/API billing, but you still pay through hardware, VRAM, electricity, and setup time.