Meta Muse (Personal Agent App)

by Meta — billed as “The World’s First Personal AI Agent Built for Everyone”

Meta’s autonomous personal agent — takes tasks and projects off your plate rather than just answering questions: sends emails, books travel, fills out forms, and completes shopping checkouts, running persistently in the background and returning when it needs approval or has an update. Launched 2026-09-08, one week after Instinct’s Series B made headlines — the two are now the two most-discussed entrants in the 2026 personal-agent wave.

See https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/

Biggest competitor to Instinct

Where Instinct is an invite-only startup product (Spear Street Technology,
$2.5B valuation, still scaling compute), Meta Muse launches with Meta’s full consumer
distribution behind it — native iOS/Android apps, a web app at muse.ai, and direct
integration into WhatsApp, which alone puts it in front of billions of existing
users. On pure reach/scale, this is Instinct’s most consequential competitor, even
though the two take different bets on how a personal agent should behave (see below).

vs. Instinct — reactive/goal-oriented vs. ambient/proactive

Meta MuseInstinct
CompanyMetaSpear Street Technology (startup)
DistributionNative apps, muse.ai, WhatsApp (billions of existing users)Invite-only private beta
BehaviorReactive/goal-oriented — converts stated goals into plans, works in the background, returns for approvalAmbient/proactive — follows up on dropped threads, arranges things without being asked
Underlying modelMuse Spark (Meta Muse model family)Not disclosed
Security architectureMuse Secure VM (isolated agent + data), independent “Sentinel” agent approves all internet requests, upcoming Confidential VM with user-held encryption keysPersistent cloud computer with cached credentials
PricingFree tier + two paid tiersFree (private beta)
Launch2026-09-082026-08 (invite-only)

Key capabilities

  • Sends emails, books travel, completes form submissions
  • Shopping checkouts via Link (Stripe’s wallet for agents), covered by purchase
    protections
  • Persistent/background work — keeps running after the app is closed, surfaces when it
    needs approval or has an update
  • Contextual memory — recalls casually mentioned preferences and proposes unprompted
    actions (e.g. turning a saved recipe into a grocery list)
  • Price negotiation on the user’s behalf

Architecture & trust framing

Built on Muse Spark, part of the Meta Muse model family (the same family
as Muse Spark 1.2 / Muse Spark 1.3, though the
consumer agent app is a distinct product from those coding-focused model releases). Runs on
Muse Secure VM, isolating the agent and user data; a separate Sentinel agent
independently approves outbound internet requests; Muse cannot view stored passwords or
payment details; sensitive actions (sending an email, making a purchase) require explicit
approval; full audit trail of activity; opt-out available for using data in training. An
upcoming Muse Confidential VM will encrypt everything with user-held keys.

Availability

US-only, 18+, rolling out as of launch. Free tier covers most needs; two paid tiers for heavier use.

  • Instinct — the closest direct competitor; opposite behavioral bet
    (proactive/ambient vs. reactive/goal-oriented)
  • Meta Muse — the model-family series hub (Spark, Glimmer, Code, Image, Video)
    this agent app is built on top of
  • Gemini Spark — Google’s equivalent always-on personal agent, announced
    2026-05-19, built on Google Antigravity + Gemini 3.5 Flash — the third major entrant in
    this category alongside Meta Muse and Instinct

Sources