Running the whole thingwith the wifi off

The two engines

HuddleOwl splits the work in two, and they are switched independently because they have very different requirements.

Transcription

A Whisper-family model, running through whisper.cpp. It turns audio into text, live, for both tracks at once. It handles 23 languages, and it detects the language once per track and then locks it, so a call cannot drift into being transcribed as a different language halfway through.

Coaching

A local language model, running through llama.cpp, which writes the live cues and the post-meeting brief. This is the one with real hardware requirements, because it has to answer while you are still talking.

What your machine needs

• About 3GB of disk for both models together.
• 8GB of RAM is the floor. 16GB is comfortable.
• Apple Silicon is where this is fastest, by a wide margin.
• macOS only, today. The Windows local engines are being built.

The app checks your hardware before it offers a model and will tell you if something is too big for the machine, rather than downloading three gigabytes and then reporting itself unavailable.

Downloading them

Settings, then the models panel. Each model shows its size, its licence and what it is for, and downloads with a progress bar you can leave running. Once a model is installed, switch the matching engine to this machine and that is the whole setup.

You can verify the result the honest way: turn the wifi off and run a meeting. Live transcription, live cues, the brief and the follow-up email all still work.

Already running Ollama or LM Studio

HuddleOwl will use it. If you have a local model server running, it can serve the coaching engine and there is nothing to download at all. You choose which model answers, rather than the app picking whichever one the server happened to list first.

One thing worth declaring in settings: how much context your own server actually serves. Ollama takes its window from its own process environment, which no API exposes, so the app cannot measure it and assumes a documented default unless you say otherwise.

Or use a cloud model instead

Point HuddleOwl at over 100 providers, or at your own endpoint, with your own key. You pay that provider directly and there is no per-seat markup from us.

Be clear about what changes: in that mode your transcript is sent to the provider you selected, from your machine, under their terms. It does not pass through us, because there is no server of ours for it to pass through. On Windows this is currently the only mode, which the Windows guide says plainly.

The two keys are separate

Transcription and analysis take different keys on purpose. Running the coach on a cloud model does not force your audio into the cloud too, and running transcription on-device does not stop you using a large cloud model for the brief. Mix them however the particular call deserves.