Skip to main content
VoiceOS is a voice assistant that lives in the Mac notch. Users talk, the agent acts, and answers arrive as cards in the notch. A custom integration puts your thing on that surface. You declare tools; the agent calls them when the user asks; your results render as native cards — or as fully custom UI when you want every pixel. If it has an API, you can build it: internal dashboards, home automation, your CRM, a coffee tracker.

First-party integrations answering in the notch. A custom integration renders on this exact surface.

Think of it like the App Store: VoiceOS is iOS, your integration is an app. One manifest declares what it does (tools), where it runs (runtime), what it needs (auth and permissions), and how it confirms actions (declarative UI). VoiceOS wires up the rest. You never touch VoiceOS internals — which is also why your code can never reach VoiceOS’s backend or secrets.

​
Pick a path

Build with AI — no code

Describe your integration in a sentence. The Integration Studio plans it, builds it, designs its cards, and lets you test-drive it live. ~3 minutes.

Build with code

A standalone folder with a manifest and a small MCP server. Full control over logic, cards, and widgets. ~10 minutes.
Both produce the same thing: an integration folder VoiceOS can install. Studio-built integrations are real folders too — open one in your editor, or hand it to a coding agent, and keep building.

​
What it feels like

Say you built an ElevenLabs integration.
1

The user speaks, and the agent confirms

“Read this in Rachel’s voice” — the agent picks your generate_speech tool. Because it acts (it spends API credits and writes a file), VoiceOS first renders the confirmation card you declared in the manifest, before any of your code runs.

The user edits the arguments; the edited values are what your handler receives.

2

Your server runs, your card appears

The handler calls the ElevenLabs API and returns JSON. The agent narrates the data; the notch shows your card.

A glance card — rendered by VoiceOS's own components from three lines of JSON.

Read-only tools skip the confirmation entirely — “what voices do I have?” just answers, instantly, with a card.

​
What you can build

​
The path through these docs

Read top to bottom and you’ll go from nothing to a polished, branded integration.
  1. Quickstart — build one with AI, in the app, in three minutes.
  2. How it works — the architecture in one page.
  3. Build with code and defining tools — schemas, handlers, background execution.
  4. Result cards, confirmations, setup fields — native UI, the easy way.
  5. Custom widgets and the Widget Kit — pixel-level control.
  6. Branding, testing, sharing — ship it.
  7. Reference — every field, block, cap, and bridge message.
The Integration SDK is in developer preview. The folder contract — manifest, MCP server, cards — is stable and versioned (schemaVersion: 1). See Limits & stability.