Meta's standalone AI app makes voice and cross-device continuity the frontend

One app, four surfaces, and a deliberate handoff gap

Meta already runs its assistant inside WhatsApp, Instagram, Facebook, and Messenger. The new standalone app is a fifth surface, and it also absorbs the Meta View companion app for Ray-Ban Meta glasses — merging device management into a new Devices tab where existing users' paired glasses, settings, and media transfer automatically.

The interesting detail is where continuity stops. You can start a conversation on your glasses and pick it up in the app or on the web. You can move between app and web in both directions. But you cannot start in the app or on the web and resume on the glasses. That asymmetry is spelled out plainly in the announcement, and it tells you the sync architecture flows outward from the glasses rather than treating all three surfaces as equal peers.

For a company positioning glasses as "the most exciting new hardware category of the AI era," building the continuity model around them is a coherent bet — but it means the frontend is not yet a symmetric mesh of devices.

Voice as the default, with the microphone made visible

Meta calls voice "the most intuitive way to interact" and builds the app around starting a conversation with one button. There's a "Ready to talk" setting to leave voice on by default, and — notably — a visible icon that indicates when the microphone is in use. On a device that lives in your pocket and connects to camera glasses, surfacing mic state is a frontend decision about trust, not just aesthetics.

The app ships two distinct voice paths. The standard experience uses Llama 4 for more conversational responses. Separately, there's a toggleable full-duplex demo that generates voice directly rather than reading written text aloud.

It doesn't have access to the web or real-time information, but we wanted to provide a glimpse into the future by letting people experiment with this. You may encounter technical issues or inconsistencies so we'll continue to gather feedback to help us improve the experience over time.Montana Labs

Shipping the full-duplex feature as an explicitly labeled, opt-in demo — walled off from web access and flagged as inconsistent — is an honest way to ship an unfinished capability. It lets Meta gather real usage data without presenting experimental speech generation as production quality.

Personalization pulled from existing Meta data, gated by geography

The app's personalization story leans on Meta's existing account graph. Meta AI can remember things you tell it, pick up context, and — the substantive part — draw on your profile and the content you like or engage with across Meta products. Link Facebook and Instagram in the same Accounts Center and it pulls from both.

This is where the announcement's rollout map matters. Personalized responses are live only in the US and Canada. Voice conversations, including the full-duplex demo, cover the US, Canada, Australia, and New Zealand. The core personalization pitch and the core interaction mode are both regionally constrained at launch, which means the app most people can download is not the app being described.

The Discover feed turns prompting into a social object

The app includes a Discover feed for sharing and exploring how others use AI — you can see prompts people share and remix them. Meta stresses that nothing posts to your feed unless you choose. This is the most Meta-specific move in the release: it applies the social-feed pattern the company knows better than anyone to the act of prompting an assistant.

The same feed appears in the web interface, which also gains voice, an improved image generation experience with style and lighting presets, and — in select countries being tested — a rich document editor that exports to PDF plus document import for analysis.

What the handoff rules signal for anyone building multi-surface assistants

The specific implication of this launch is that Meta is building its assistant frontend as a set of unevenly connected surfaces, not a uniform layer. The one-directional glasses handoff, the region-by-region gating of voice and personalization, and the web-only document editor test all indicate a company shipping capabilities where the underlying plumbing is ready rather than waiting for parity.

For teams designing their own cross-device AI experiences, the useful lesson is in the constraints Meta chose to state out loud: label the experimental voice mode, show the microphone indicator, and be explicit about which resume paths work. A multi-surface assistant is defined less by its model than by the honesty of its seams — and this first version documents its seams rather than hiding them.