Meta Introduces Its AI Agent, Muse

Access is available via web, iOS, Android, and WhatsApp with Meta AI glasses support coming soon.

Summary

Meta launches Muse, an autonomous AI agent capable of executing everyday tasks and making purchases

The service connects directly to user apps to deploy subagents, build tools and operate in the background

Muse debuts in the U.S. with a free tier and two premium subscriptions while running on an isolated secure virtual machine

Meta introduced Muse on Tuesday, September 8. This personal AI agent is engineered to autonomously execute everyday tasks for users in the U.S The rollout shifts the landscape from passive chatbots to proactive agents built to act on a user’s behalf. This evolution demands deep integration into daily digital routines. Extensive integration requires users to weigh convenience against granting access to personal data.

Muse Spark powers the system by connecting to everyday apps and services like email, calendars, payments, health and fitness, smart home devices, dining, shopping and music. Users maintain control by opting into these connections one at a time. Once integrated, the agent handles diverse tasks including sending emails, booking travel, lowering bills, filling out forms, creating plans and making purchases. Checkout protections are provided through Link by Stripe, with Shopify and 1Password integrations coming soon. Personalized settings allow users to select an avatar, give the agent a name and adjust communication preferences.

Technical infrastructure enables the system to push the boundaries of AI autonomy. If a linked service has a public API, the agent sets up a connection using user-provided credentials. Where no API is available, Muse accesses the service via a browser. Tarek Sheasha, Meta Superintelligence Labs VP, stated the model was trained for zero-shot tool calling using CLIs and skills. Background operation is possible even after the user leaves the app. The system can autonomously launch swarms of subagents, build its own tools and edit its own code to optimize performance.

Rigorous digital safety measures require the agent to run in an isolated, persistent Linux virtual machine called Muse Secure VM. This environment houses a full browser, storage, CPU and memory. A separate Sentinel agent monitors the environment from within the same virtual machine while remaining strictly isolated at the system level. Meta claims the agent lacks visibility into actual passwords or payment methods. The company also states it does not share conversation data with its advertising systems.

Access is available via the web, through iOS and Android apps and within WhatsApp chats. Meta also plans to bring the agent to its AI glasses soon. Getting started requires a payment card on file to unlock a free tier with a usage meter. Two paid subscription plans are available for heavy users. The Power tier costs $20 per month while Maximum is priced at $100 per month.