Muse has only been live for a few weeks, and Meta is already acting like it has something to prove. At the company’s annual Connect event in Menlo Park on Wednesday, CEO Mark Zuckerberg took the stage in a T-shirt that read “Building is My Love Language” and spent most of his keynote talking about one thing: the Muse personal AI agent. According to TechCrunch, the company rolled out a significant batch of new capabilities, integrations, and partnerships that suggest Meta is treating Muse as its most important consumer product in years.
For context, Muse launched in early September. It’s an AI agent powered by Muse Spark, Meta’s multimodal model built specifically for agentic work. The pitch from day one has been accessibility: give everyday users an AI that can actually handle their busy work, connecting to email, calendars, and third-party apps without much friction. The competitor set here is obvious. OpenAI has Operator, Anthropic has its own computer-use features, and Google is pushing Gemini deeper into Workspace. Muse is Meta’s answer to all of them.
A face for your agent
The most visually striking announcement was a digital avatar feature. Meta is building a new AI model called Muse Realtime Avatar that gives each user’s agent a distinct face, body, and voice. Users will be able to video chat with this avatar in real time, assign it tasks, and ask it questions as if it were a person on a call. Zuckerberg named the physical avatar “Jolly” during the keynote. There’s an obvious echo of Meta’s earlier Metaverse ambitions here, but framing it as a practical productivity interface rather than a social space feels like a smarter positioning move.
Smart glasses, Mac control, and its own inbox
Three other announcements stood out for developers and power users evaluating whether Muse is worth building around.
- Smart glasses integration: Muse will work inside Meta’s smart glasses via a wake word. The agent can guide workouts, log meals, book appointments, and help users buy things they see in the world. Meta says this is coming “in the coming months,” which is not a precise timeline.
- Mac computer use: Muse is coming to Mac and will be able to operate any app on a user’s desktop, similar to what OpenAI’s Operator and Anthropic’s computer-use features already do. Meta’s chief AI officer Alexandr Wang described it simply: “You can walk away from your computer and it keeps working for you on all the jobs you lined up.”
- Its own email address: Because Muse runs in the background and handles tasks independently, Meta is giving it a dedicated email address. Users can forward emails directly to their agent or add it to existing threads. Wang said this is coming “soon,” again without specifics.
The shopping angle is where Meta’s business model shows
Zuckerberg was unusually transparent about how Meta plans to make money from Muse. The agent is free up to a large number of tokens, and over time Meta intends to take a small fee from transactions. That explains the aggressive push into commerce partnerships.
Meta has locked in deals with Stripe and Shopify, and Muse can now browse the entire Shopify product catalog and complete purchases using Lync and Shop Pay. PayPal support has been expanded. On the retail side, new partners include Best Buy, Gap, Sephora, Walmart, Wayfair, Expedia, and soon Instacart. GitHub, Granola, and Notion are also listed as new connectors.
For small businesses specifically, Meta opened its platform to third-party connector development and received more than 1,500 applications in under a week. That’s real traction. If Meta can get developers building integrations at that pace, the connector ecosystem becomes a meaningful moat.
Why this matters right now
The agent space is moving fast and getting crowded. But most competitors are targeting enterprise or technical users first. Meta’s distribution, hundreds of millions of people already on WhatsApp, Instagram, and Facebook, gives Muse a user base that OpenAI and Anthropic simply don’t have access to out of the gate. The question is whether Muse is reliable enough to earn the kind of trust that agentic tasks require. Letting an AI buy things, manage your email, and run your Mac is a different category of risk than asking it to summarize a document. The features announced this week are directionally right. Execution is everything from here.
