Muse review: what builders actually report after the first weeks

Muse is good at errands that take a person an hour of tabs and phone menus, like chasing a refund, negotiating a bill or booking through a partner connector. It's weaker on huge media libraries, anything that needs a phone verification code, and Amazon, which blocks it. The builders who are happiest set a clear goal and a spending limit, then check the result. The ones who are least happy either hit an accessibility wall or don't trust Meta with the access.

This review is assembled from what people have posted and filed in the catalog, not from a lab. Numbers are as reported by the author.

What Muse is, briefly

Muse is Meta's personal agent, launched September 8, 2026 in the US. It runs errands, fills forms, books travel, browses and buys, keeps working after you close the app, and runs on Muse Spark 1.3 inside a dedicated cloud VM (Meta newsroom). It's on iOS, Android, web, WhatsApp and, since September 17–18, Mac. For a fuller explainer see what is Muse.

It got a lot of trial quickly: more than 2.5M downloads by September 21 and the #1 free iOS app spot in the US, per CNBC.

Where Muse does well

Money errands. This is the strongest cluster in the catalog:

• Chandra Bhavanasi authorized Muse to accept any AT&T fiber offer under $70/month; it came back with $40/month plus 3 months free.
• Ejaaz got a $250 credit for a 7-hour Delta delay in about five minutes, and Muse rebooked the flight too.
• A receipt audit over six months of email found a forgotten $100/month subscription.
• Joshua Liu says Muse saved him $113/month on car insurance.

The Agents Breakroom library of around 500 use cases found 20% were about money, so this isn't a fluke of what gets posted.

Partner-connector bookings. Where a connector exists, checkout is quick. Alex Volkov's Muse booked dinner tickets through Stripe Link while a rival agent was still searching.

Connecting messy personal feeds. One parent wired school grades, announcements and sports schedules in about 30 minutes. Another had Muse merge 400 Google Docs into one 400-page PDF, which the author says other agents failed at.

Custom connectors. Muse will write its own integration against an API. One Reddit demo built a Linear connector in under a minute.

Small-business back office. Rich Rose spent a week running a salon's back office with it: Excel order forms from color charts, scanning photo order forms and checking stock.

Where Muse falls short

Human verification. In a phone call to Xfinity, Muse reached a human, then hit a verification text it couldn't read and had to patch the user in.

Scale on media libraries. Re-sorting 18 Spotify playlists worked; the same user says a 40k Google Photos job failed (post).

Amazon. Blocked since September 20 (GeekWire).

Accessibility. Taylor Arndt found non-standard text fields and missing heading structure with VoiceOver, plus missing iCloud email support (review).

Opaque quotas. Plans are tokens per week: Free (reported up to 100M), Power at $20 for 500M, Max at $100 for 3B. There's no official conversion to tasks.

Undocumented internals. Users keep finding things the docs don't mention: on-disk "dreams" jobs and SOUL/MEMORY files, and HomeKit, geofence and BLE tools in the tool schemas. Some find this exciting. Others find it a reason for caution.

Privacy and trust

Meta's design is serious on paper: a per-user VM, a separate Sentinel authority that gates connector actions and network egress, credentials the model never sees, and a bug bounty up to $300,000 (Meta). Muse data is excluded from Meta's ad systems and you can opt out of training.

Against that: Meta's own write-up says prompt injection "remains an open problem" and that Muse "will sometimes make mistakes", and a Mac zero-day was reported with no patch mentioned as of September 22. Details in security and privacy.

Scorecard

Area	Verdict from reports
Bill negotiation, refunds, claims	Strong
Bookings via partner connectors	Strong
Custom connectors from an API	Strong, but unreviewed by Meta
Big photo/media libraries	Weak
Phone verification steps	Needs a human
Amazon shopping	Not possible
Screen reader use	Needs work
Cost predictability	Unclear

How it compares with other agents

There are few fair head-to-heads so far. The one in the catalog is Alex Volkov's ticket task, where Muse booked through Stripe Link while Instinct was still searching. A Reddit parent reports that connecting school and sports feeds took about 30 minutes in Muse versus 5 to 6 hours with the open-source Hermes agent. That's the trade in one line: Muse is faster to get working because Meta has done the connector and payment plumbing, and you accept Meta's platform in return. If that trade doesn't suit you, alternatives to Muse covers ChatGPT agent mode, Instinct and self-hosted options like OpenMuse.

Who should use Muse

Use it if you're in the US, have a backlog of annoying admin with a clear success condition, and are comfortable giving a Meta-run VM access to email and payments with approvals on. Skip it for now if you depend on VoiceOver, shop mostly on Amazon, or can't accept Meta holding that data. If you're unsure, start with read-only connectors and one money task with a hard ceiling. Start here walks through it, and limitations lists what to expect.

Frequently asked questions

Is Muse worth paying for?

If the free tier runs out on real errands, Power at $20/month is cheap next to the savings people report, such as $40 off an AT&T bill every month. If you mostly chat, the free tier is probably enough.

Is Muse safe to give my credit card?

Muse checks out through Stripe Link, which uses your saved method at Link merchants and issues single-use virtual cards elsewhere. Purchases require your approval. The residual risk is prompt injection, which Meta says it hasn't solved.

What is Muse best at?

From catalog reports: negotiating bills, claiming refunds and compensation, auditing receipts, and bookings through partner connectors like Stripe Link, Ticketmaster and Duffel.

What are the biggest complaints about Muse?

US-only access, the Amazon block, token quotas with no task conversion, VoiceOver problems, and distrust of Meta with personal data.