Voice AI SDK Features.

Register your app's real JavaScript functions with a plain-English description and typed params. ai.register() is all it takes, the agent calls your actual code, not a script.

ai.bindState() feeds the agent your current UI state on every turn, so it reasons over what the user is actually looking at instead of a stale snapshot.

Mark any action dangerous: true and wire onConfirmation(), mutating or destructive calls pause for explicit user approval before they run.

ai.use() lets you intercept, validate, or block executions before they run, check auth state, rate-limit, or log every call the agent makes.

Every init ships a versioned snapshot of your capabilities. If a sync fails or a project's manifest is locked, the SDK simply keeps running on the last known-good version.

Every project locks down exactly which origins may open a Voxide connection, enforced with strict CORS checks at the WebSocket handshake, not just a dashboard toggle.

Upload documents, let Voxide chunk and embed them, and attach a knowledge base to any project so the agent answers grounded in your content.

On React and Next.js, ship <VoxideWidget> with your accent color, position, and title out of the box. Anywhere else the headless core gives you the same agent and voice session to render however you want, and ai.setLanguage() switches voice language at runtime either way.

setFallback() gives the agent a custom message and suggested next steps for anything it can't do, no dead-end "I don't understand."

Set a phrase like "hey acme" and visitors start talking without clicking, detected on-device, so nothing streams until it's heard. An accessibility path for anyone who can't see or reach the launcher. Pro.

When a request is beyond the agent, it offers to connect the visitor to a real person and files a support request, to your inbox, a webhook, or your dashboard, with the conversation attached. You write when to escalate in plain English. Pro.

Emails, phone numbers, card and ID numbers are stripped before anything is stored, not hidden afterwards, removed. So the data your visitors speak isn't visible to you, to us, or in a backup. On every plan.