Meta has officially launched the personal AI assistant Muse (muse.ai). It no longer just answers questions; it can open a browser, connect apps, fill out forms, handle emails and purchase processes, and continue tasks after users leave the app. Muse is powered by the Muse Spark series models and was among the first to launch on iOS, Android, and web for US users.
Muse is more than just a chatbot
Muse's core is "doing things for the user." It has a standalone cloud computer, file system, terminal, and Chromium browser, can search web pages, create documents, run code, and break long-term goals into plans for continuous follow-up. Tasks can run in parallel and proactively notify users once completed.
More importantly, context persistence. Muse remembers conversations and goals and can proactively offer suggestions. For example, extracting schedules from emails, preparing shopping lists, booking restaurants, and even continuously adjusting execution plans around a long-term goal. Personal AI Agents are shifting from "Q&A" to long-term hosting.
Security is placed at the core layer of the Agent
When AI can send emails, log into websites, and make payments, security questions are more direct than whether the model answers are accurate. Meta has configured a standalone Muse Secure VM for Muse, with each user having an isolated environment; Passwords and payment credentials are stored in secure areas that agents cannot directly read.
Web access also requires a dedicated Sentinel Agent. Sensitive operations such as sending emails and purchasing goods require manual confirmation by default, and browsers have added prompt injection, data leaks, and high-risk form detection. Meta is also offering a Muse vulnerability reward program with rewards of up to $300,000.
Meta is pushing personal AI toward long-term hosting
Muse's Spark 1.3 is dedicated to enhancing long tasks, tool calls, multitasking, and agent workflows. Rather than simply pursuing chat experiences, Meta wants to prove that AI can be a continuously running digital executor, rather than a dialog box waiting for prompts.
Muse is still in its early stages, with the first batch only available in the US, most basic usage free of charge, and higher usage offered through subscriptions. The real test is clear: when AI starts to click "send" and "pay" for users, beyond capability limits, permission boundaries and accountability will be the key to whether personal agents can become widespread.
