Local models
Summaries, descriptions and digests run on a model on your Mac. Cloud models exist only behind a switch that is off by default.
The text features — summaries, action item descriptions, to-do descriptions, follow-up emails and the day digest — are written by a language model running on this Mac. That is the default and the ordinary case. Cloud models exist, but only behind a setting that is off by default, and the app says when text would leave. You can also run summaries on a server of your own, on this Mac or another computer: see Summary servers.
What runs locally
| Feature | Model |
|---|---|
| Summary and action items | The local summary model |
| Action item and to-do descriptions | The local summary model |
| Follow-up email | The local summary model |
| Day digest | The local summary model |
| Agent memory search | Embeddings from the local model (see Agent memory) |
| Transcription | Whisper, always local; dictation may use Apple's on-device speech instead |
By default, inference runs on 127.0.0.1:11434. Nothing about a meeting is sent to write a summary unless
you have turned cloud models on and picked one, or chosen a summary server
off this Mac. Embeddings always use the local Ollama.
Settings
| Setting | Where | Default |
|---|---|---|
| Where summaries run, Status, Refresh, Model | Settings → AI → Where summaries run | Ollama on this Mac; the model chosen at setup |
| Allow Ollama cloud models | Settings → AI → Cloud models (there is no second switch in recording settings) | Off |
| Summary model | Settings → Meetings → Recording & models | The model chosen at setup |
| Summary Setup (Ollama's install and models, other servers) | Settings → AI → Change…, or PrivateGoat → Ollama Setup… | Offered when no model is available |
| Embedding model for memory | Settings → Knowledge → Meaning search (embeddings) | Automatic |
The memory check
Before loading a summary model in Ollama on this Mac, PrivateGoat checks whether the model fits in the Mac's free memory. A server on another computer is not checked. It judges by what the model took the last time it was loaded, as Ollama reported it, and keeps the latest of those measurements, so one reading that came out too small cannot let a model load on a Mac that is short on memory. Both passes of a summary ask for one context size, so the model is not loaded a second time.
- It fits: the job runs.
- It does not fit: nothing is loaded. The notice names the apps using the most memory — for example "Google Chrome (12 GB)" — and the summary starts by itself once enough memory is free.
- Other ways out: Summarize with a smaller installed model when one is suggested, Another model for this one summary, or Not now.
- Summarize anyway…: runs the job after a warning that the Mac may stop responding. It is never offered while recording.
What is stored
| Item | Where |
|---|---|
| The chosen model and the cloud switch | App preferences |
| Downloaded models | Ollama's own store — PrivateGoat does not copy them |
| Speech models | ~/Library/Application Support/Sonavia/Models |
What leaves your Mac
- Local inference: nothing. It stays on the loopback interface.
- A model download or pull:
ollama.com(see What leaves your Mac). - The cloud switch on: the cloud model list is fetched from
ollama.comat most once a day and cached for 24 hours, and any text you send to a cloud model goes toollama.com. No API key or account credential is stored by PrivateGoat for this. - The cloud switch off: no listing request is made and cloud models are not shown.
When it is off or fails
- No local model yet: recording and transcription still work; the summary panel says summaries need a local model and offers Set up Ollama…. See Summaries and action items.
- The model does not fit: the memory check above applies: wait for memory, or choose the smaller alternative.
- The cloud listing fails: the last listing is kept.
- A cloud model is selected: the UI says that the text will leave the Mac; transcription still never does.
- A recording starts: AI work pauses and an interrupted summary is queued again when the recording ends.
Related
- Summaries and action items — the main consumer of the model.
- Agent memory — embeddings, and the fail-closed cloud rule.
- What leaves your Mac — every request in one table.
Keys and sign-ins
Your Jira sign-in lives in the macOS Keychain, guarded by Touch ID where the system allows it, like a password manager.
Summary servers
Summaries run on this Mac with Ollama by default. Choose Ollama or an OpenAI-compatible server on another computer, or a cloud service, and see exactly where the text goes.