Muse (Meta)

Muse is Meta's consumer personal AI Agent, launched September 2026 (official tagline: Your Personal AI Agent). Its positioning is not answering questions but getting things done — actually executing multi-step tasks across the web and connected apps, and running persistently in the background. Meta draws the boundary with three tiers: a chatbot means one question one answer; an AI assistant means richer conversation and content generation; a personal AI Agent actually performs actions, autonomously completes multi-step tasks, and stays resident in the background.

Core capabilities

• Muse Secure VM (the agent runtime): each user gets a persistent, isolated Linux cloud VM with a full browser, file system, and terminal — real compute for compiling code, running sub-agents, and scheduled jobs; the VM is the system of record for user data
• Sentinel oversight layer: a host-side agent, Sentinel, runs on the same machine, system-level isolated from Muse, and is the sole permission authority for connector actions and all network egress — Muse can only propose, and Sentinel decides whether to allow
• Credential isolation: third-party OAuth tokens live in a dedicated container inside the user's VM, never entering Meta's centralized infrastructure; the agent runtime receives only surrogate tokens — real passwords and card numbers stay invisible to the agent
• Human-in-the-loop confirmation: routine browsing is allowed by default, but hard-to-reverse actions like sending email, paying, or submitting important forms force a stop and require confirmation; structured approval cards, per-action/per-session/per-task/time-limited authorizations, and a full activity audit log are provided
• Persistent cross-conversation memory: readable, editable memory files, a Forget skill, and one-click reset; background tasks keep running after the app closes and proactively report completion or changes
• Multi-surface coverage: iOS/Android apps, a Mac client, the web app (muse.ai), and WhatsApp; planned are Meta AI glasses, Muse Charm hardware, and Realtime Voice video calls
• Payments: Link by Stripe support generates one-time card numbers so the real card number is invisible to merchants and the agent alike — the first AI agent covered by Link's purchase protection

Use cases Suited to consumers in the US/Canada, 18 and older, whose time is drained by digital chores — email, scheduling, bookings, price comparisons, forms — and who are willing to grant deep permissions. Delivery cases verified by independent media include: phoning an insurance company and transferring to a human, building spend tracking across spreadsheets and canceling duplicate subscriptions, comparative shopping and purchasing, and turning an Instagram recipe video into a shopping list.

Not for: users in mainland China or Europe (US/Canada only, and Meta still calls it limited testing); users highly sensitive to data sovereignty or clearly opposed to Meta; people who just want a Q&A assistant; teams needing enterprise-level permission governance (Muse has no team management); heavy Amazon shoppers (Amazon has blocked it as an unauthorized AI agent).

Name disambiguation (important): this entry is Meta's Muse — not the former company that ran video AI search on the muse.ai domain from 2016 (renamed Skiv and moved off the domain in March 2026), and not MuseScore / Muse Group, Microsoft Muse, the Muse 2 headband, or other same-named products.

Risk warning: security and compliance problems surfaced rapidly within two weeks of launch — the Amazon ban, a Reuters report on beta overreach reading a tester's private iCloud photos, a VM-access class of vulnerability (official response: stronger warnings rather than a confirmed fix), a macOS 0-day (hotfixed), and humans outsourced to pose as AI calls; Meta officially admits that before Confidential VM delivery, Meta can still access user VMs when necessary; training on user conversations is on by default; terms include a $250 liability cap, mandatory arbitration, and no refunds; data retention periods are undisclosed.