QWE AI Academy Can Muse finish work after you lock the phone?
Meta shipped Muse on September 8, 2026 – US only – on iOS, Android, muse.ai, and WhatsApp (core chat). Chatbots wait. Muse keeps going on its own box after you leave.
Demos look clean. Real use is messier: connector scope, approval fatigue, and how fast multi-step browser runs chew the free pool. Below is setup, one complete goal, and the sharp edges early testers hit.
Giving software a live goal still feels odd the first week. It’s closer to lending your keys to a careful stranger than opening another chat tab.
Chat assistants vs an agent on its own machine
Normal assistants outline steps. You still copy data, open tabs, and re-explain context when the thread dies. Muse takes a goal, plans, works in the background, and only pings you for irreversible moves.
It runs inside a dedicated Muse Secure VM – isolated Linux with its own browser, storage, and CPU. A separate Sentinel process gates egress and sensitive actions (Meta’s security write-up). That split is why approvals exist at all: the worker agent proposes; Sentinel enforces your rules.
Setup in about ten minutes
Log in with your Meta account. Name the agent, pick or build an avatar, set tone once – it sticks across sessions. Model stack is Muse Spark per Meta’s launch materials.
- Settings → Data Controls: training on interactions is on by default. Opt out now if you don’t want chats used to improve models. Conversations aren’t fed into ad systems, according to the Meta newsroom post. Full Confidential VM mode with a user-held key is slated later in 2026 – not day-one.
- Check the free meter. Zuckerberg put free capacity at roughly 100 million tokens per week (weekly reset) as of launch week. Power is $20/mo; Maximum is $100/mo when you want heavier continuous hand-off.
- Connectors: Facebook, Instagram, and Threads auto-link if they already share Accounts Center. Add Google Workspace, Spotify, Apple Health, Ticketmaster, OpenTable, or Plaid one at a time. Prefer read-only until you trust the loop. Details live in Meta’s connectors help page.
Missing a service? Tell Muse to build a custom connector. It walks API setup and drops credentials in a vault it can’t freely read. Meta does not review those custom builds – you own the third-party risk.
Set Web access and connector defaults to ask before purchases and outbound sends. Loosen later. Approval banners blur once they feel routine.
First workflow that isn’t a toy prompt
Skip “make me productive.” Give Muse data it can already reach and a hard stop.
Try: “From my Google Drive notes on last quarter’s client calls, pull open action items, build a simple tracker artifact in Library, draft follow-up emails in Gmail, and hold every send for my approval. Flag anything older than 30 days.”
Muse reads Drive via the Workspace connector, writes the tracker as a Library artifact, drafts mail, and surfaces approval cards before send. Watch live with Open browser or the activity log under the avatar. Close the app – background work continues. Purchases (when you get there) go through Link by Stripe one-time cards; your real number stays hidden, and eligible first-agent buys pick up Link protections for returns, price drops, and damage.
Afterward open Goals. Progress sticks there; Muse may push unprompted next steps if it spots a pattern.
How much autonomy feels okay to you – read-only connectors forever, or send-email after one successful dry run? That preference matters more than any feature list.
Daily controls
Avatar tap → full activity log (past, live, queued). Take over the browser mid-run or stop the task cold. Library holds artifacts outside the chat thread so long goals don’t turn into one endless scroll.
The catch is reliability under load. Pre-launch dogfooding covered by Reuters reporting saw stalls (UI stops updating around the 15-minute mark), silent disconnects that needed re-login, and rare unauthorized data-exposure bugs before hardening. Long browser jobs freeze? Stop, re-prompt with the last known state. Free tokens look huge until a multi-step browser swarm with sub-tasks runs; early public tests (Gizmodo and others citing the same 100M figure) still showed headroom after heavy sessions – often single-digit to low-teens percent burned – but the meter moves faster than plain chat.
FAQ
Is Muse free forever?
Free tier with the weekly token pool covers everyday use as of September 2026. Paid tiers are for heavier hand-off. Meta says most people should stay on free; power users pay for compute.
What if Gmail is connected and Muse drafts a send I hate?
Sentinel blocks sensitive actions until you clear an approval card – once, this task, this site, always, or deny. Example: it drafts a client nudge, you open the card, edit the tone, approve once. You can also lock the connector to read-only. Everything lands in the audit log.
Outside the US? On glasses?
US launch only as of September 2026. WhatsApp carries core messaging but drops some app surfaces (full Ideas/feed-style tabs). AI glasses are listed as coming soon with no firm public date. No official non-US timeline in the launch materials – watch Meta’s channels if you’re abroad.
Open muse.ai or the app. Name the agent. Opt out of training if you want. Connect one low-stakes read-only source. One concrete goal with a finish line. Sit through the first approval. That’s enough signal to know whether Muse earns a daily slot.