Local setup expectations
The self-hosted beta ships two install paths: a CI-generated Windows alpha installer published on GitHub Releases, and a source build documented in the OpenCohost core and UI repositories. Either way you bring Ollama and the models you select and download. There is no hosted service, bundled model, or universal hardware profile.
- Quick path (Windows): download the alpha installer from the OpenCohost GitHub Releases page, run the setup executable, and accept any SmartScreen prompt for unsigned alpha builds.
- Developer path: follow the OpenCohost core and UI repository documentation for a source build (Python, Node, pnpm, Rust); this site intentionally does not invent exact source commands.
- Install Ollama, then choose and download the models you want to run locally; OpenCohost does not bundle models. Kira cannot start without Ollama running.
- Use a modern Windows machine with a dedicated NVIDIA GPU and enough VRAM for the selected local modes.
- Start with lighter LLM or voice modes if you are also running a heavy game on the same machine.
- Treat Ollama, the selected TTS path, and Whisper/VAD as workloads that compete for resources when enabled together.
- Keep OBS and streaming tools updated before connecting avatar, music, or audio states.
- Expect hardware, electricity, VRAM, and setup time in local mode. Free does not mean compute-free.
Core modules and connected apps
OpenCohost is the technical chassis of the self-hosted stack: it coordinates local components rather than hosting or bundling every model. Ollama runs models you choose and download; speed, quality, GPU/VRAM, and RAM requirements depend on that choice.
OpenCohost orchestrator
Coordinates LLM output, voice generation, agenda state, profiles, chat reactions, context, and resilience.
Ollama LLM path
Provides local model responses for the co-host, with speed and quality tied to model choice and hardware.
Voice output path
Provides Piper local synthesis, optional Edge-TTS light synthesis when privacy allows it, and an active path for custom voice work.
LiveAudio bridge
Separate app for Silero VAD, OpenAI Whisper open-source transcription, subtitles, transcripts, and voice context.
OpenCohost-designed behavior
Profiles, personalization, chat reaction, repetition mitigation, and context handling are designed by us.
Optional visual ambience
Avatar and music modules improve the show feel, but they are not required for the core AI loop.