Meta adds computer use to Muse for Mac

• Computer Use is rolling out to Muse desktop today, according to the rollout post.
• Muse Mail gives each agent an email address, so users can forward work or include it in a thread, the Muse Mail announcement says.
• Voice and real-time video now let a conversation continue while Muse performs background work, according to the Connect announcement.
• Muse Realtime Avatar is designed to answer in under a second during long conversations, the Realtime Avatar announcement says.

Meta’s official Connect recap lists the Muse app, WhatsApp, web, and Mac desktop as ways to reach the agent. The research post says reference media can turn portraits, animals, and ordinary objects into live conversational avatars.

Meta’s Connect recap says Computer Use can drive any app on a Mac after the user grants permission. The queue-and-walk-away interaction is central: the Mac computer-use demo says users can line up jobs and leave Muse working at the laptop.

The same recap says the earlier Mac client could manage messages and find old files. Computer Use extends that client from app-specific assistance to operating the desktop’s applications.

Muse Mail makes the agent an addressable participant in email. Users can add it to an existing thread or forward material for it to handle, the Muse Mail announcement says.

The Connect panel groups connectors into four inventories:

• Shopping: American Eagle, Best Buy, Dick’s Sporting Goods, Fanatics, Gap, Michael Kors, Sephora, Ulta Beauty, Walmart, and Wayfair.
• Productivity: Box, GitHub, Obsidian, and Notion.
• Payments: Shop and PayPal.
• Travel: FlightAware, Virgin, Turo, and Fever.

Muse’s broader security design separates the agent runtime from credentials and outbound-policy services, according to Ground Truth’s report. The Muse security architecture depicts the components behind that separation.

The system described in the diagram has three relevant boundaries:

• Runtime cell: tool execution occurs in isolated Linux containers.
• Credential service: the model never receives real credentials.
• Sentinel: a separate permission authority checks outbound calls, applies network approvals, and performs credential insertion.

Muse Realtime Voice emits speech tokens carrying both content and prosody; Realtime Avatar consumes that same stream to produce synchronized video. A fixed-length motion history supplies context for each next chunk, which the shared-stream architecture says bounds computation over arbitrarily long conversations.

Meta’s inference compression path is unusually concrete:

1. A 40-step diffusion teacher with three-way classifier-free guidance requires 120 evaluations per video chunk.
2. An unguided two-step causal student uses a fixed-length KV cache.
3. Self-forcing is intended to limit drift, with the distillation note claiming 60 times fewer evaluations at near-teacher quality.

Meta evaluated matched avatars in two-to-three-minute conversations inside Runway Characters and HeyGen LiveAvatar, and the live-call evaluation reports higher overall preference for Muse. Meta also says all Realtime Avatar output is watermarked as AI without added latency in the watermarking claim.