What runs locally
How much of Zeph works with no network at all — and where the cloud comes in.
The device runs its own operating system, and both halves of the voice pipeline have a local option. Here is where the line falls, and what you choose to put on each side of it.
The device runs its own operating system
The screen, the buttons, the launcher, Settings and the built-in glances are the device's own software — the clock and the battery on a face come from the device itself. An app you install runs there too. All of it is standalone: it keeps working with your laptop closed and in a bag.
Most apps declare required instead — it is the default. Their work happens in the companion on your Mac, which computes their screens and pushes them to the device. With no companion reachable, the device dims a required app's content and says Not connected. A third kind, enhanced, works offline and gets richer when the companion is there.
The app format, for the curious
Apps are compiled to WebAssembly — WASM for short — the same format whether they run on the device itself, as a standalone app does, or in the companion on your Mac, which is where a required app runs.
A glance that arrived with an installed package draws its last cached snapshot when the companion is away, and a placeholder before the first one lands.
The link between the device and your Mac is a direct one, over Bluetooth or USB. It does not route through us.
Speech and language can run on your Mac
Speech-to-text can run entirely on your machine. A model downloads once and then works offline, with no key and no account.
Language models can too — through a local runtime you point the companion at, like Ollama or LM Studio, or through Apple's on-device foundation models, which are already part of macOS. Your prompts go to a process on your Mac.
That covers the model. An agent step reaches outward the moment it calls a connector into one of your accounts, or an MCP tool that lives on someone else's server.
Every provider the companion ships, and which side of the line it sits on:
- Providers
- 13
- Run locally
- 6
- Free
- 6
- Metered
- 1
| Provider | What it does | Runs | Cost |
|---|---|---|---|
| Anthropicanthropic | modelsClaude models for agent and summarization features. | Cloudapi_key | Not stated |
| Apple Intelligence (local)apple-fm | modelsOn-device language model via Apple Foundation Models. No key, no download. | Localapple-fm | Free |
| Deepgramdeepgram | sttttsStreaming speech-to-text with live partial transcripts, plus Aura text-to-speech on the same key. | Cloudapi_key | Metered |
| Google Geminigemini | modelsGoogle's Gemini models (1.5 Pro, Flash, and newer) power the assistant, agent workflows, and tool use over the Gemini API. Connect with an API key from Google AI Studio. | Cloudapi_key | Not stated |
| Groqgroq | modelsGroq serves open models (Llama, Qwen, and more) on its LPU inference hardware behind an OpenAI-compatible chat-completions API for very low latency. | Cloudapi_key | Not stated |
| LM Studio (local)lm-studio | modelsembeddingsLocal language models + embeddings via LM Studio's OpenAI-compatible server. | Locallm-studio | Free |
| MLX Whisper (local)mlx-whisper | sttOn-device speech-to-text via Apple MLX (Metal). No key, runs offline. | Localmlx-whisper | Free |
| Ollama (local)ollama | modelsembeddingsLocal language models + embeddings via a running Ollama daemon. | Localollama | Free |
| OpenAIopenai | modelssttttsOpenAI's GPT models (GPT-4o, o-series, and newer) power the assistant, agent workflows, and tool use over the Chat Completions API. The same key also drives Whisper speech-to-text (/v1/audio/transcriptions) and text-to-speech (/v1/audio/speech). Connect with an API key from platform.openai.com. | Cloudapi_key | Not stated |
| OpenRouteropenrouter | modelsOpenRouter is an OpenAI-compatible model gateway: a single API key routes chat completions to 400+ models (Anthropic, OpenAI, Meta, Mistral, and more) behind one endpoint. | Cloudapi_key | Not stated |
| Parakeet (local)parakeet | sttOn-device speech-to-text via NVIDIA Parakeet (ONNX). No key, runs offline. | Localparakeet | Free |
| Vercel AI Gatewayvercel-ai-gateway | modelssttembeddingsVercel AI Gateway is an OpenAI-compatible model gateway: a single API key routes chat completions to hundreds of models across providers (Anthropic, OpenAI, Google, xAI, and more) with unified billing, fallbacks, and observability. | Cloudapi_key | Not stated |
| Whisper (local)whisper | sttOn-device speech-to-text via whisper.cpp. No key, runs offline. | Localwhisper-cpp | Free |
Cost is what the provider’s manifest states. Most cloud vendors state nothing, and an unstated cost is not a free one — those rows say so rather than guess.
What does need the network
- Cloud providers, if you choose one. You bring your own key; it is verified once and stored in your system keychain. Your audio and text go to that provider, with your key.
- Connectors and remote MCP tools, when an agent step calls one. A connector acts on an account you hold elsewhere — Gmail, Calendar, Slack. A tool can be an MCP server reached over HTTP.
- The catalog, to discover and install apps and skills.
- Updates — the companion pulls its own from a public release feed, and firmware images from a public release manifest. The providers, model runtimes and audio driver it bundles ride that same companion download.
- Diagnostics, unless you switch them off — anonymous usage and crash reports that help us make Zeph better. What they carry, and the one switch that stops them, is on Privacy.
Accounts
You do not need an account to use Zeph today. Install the companion, pair a device, add a provider key if you want one, and go.