What Muse can't do: limitations to read before you build
Meta Muse is US and Canada only, 18+, token-limited, blocked by Amazon and still prone to mistakes. The limits builders keep hitting, with sources.
Meta Muse can't be used outside the US and Canada or by anyone under 18, can't buy on Amazon, and can't tell you how many tasks your token quota will cover. It also still makes mistakes. Meta's own Mac help page warns that "your agent may be inaccurate or take unexpected actions." Most other limits builders report come from the connector platform being new, the agent's VM being small, or benchmark scores running ahead of hands-on results.
None of this makes Muse useless. The catalog has plenty of real wins, collected in what people build with Muse. But if you know where the edges are before you start, you'll plan around them instead of finding them halfway through an errand.
Who can use Muse right now
Muse launched on September 8, 2026 in the US and added Canada on September 18 (iPhone in Canada). It is for people 18 or older (or the local age of majority). No UK, EU or India date has been announced, and region is checked at the account level when you sign in, so a VPN doesn't get around it. Meta describes Muse and its subscriptions as being in "limited testing" on its subscriptions help page. Muse for Mac launched in the US on September 17. If you're elsewhere, Meta Muse outside the US covers what you can use today.
Surfaces are iOS, Android, the web at muse.ai, WhatsApp and Mac. AI glasses were announced at Connect 2026 but are listed as coming "in the coming months", and the Muse Charm pocket device has details promised "later this year" (Meta Connect roundup). If you're planning a build around glasses or Charm, you can't test it yet.
Token quotas with no task conversion
Muse plans are measured in Muse tokens per week: Free is usage-limited (The Batch reports up to 100M a week), Power is $20/month for 500M, and Max is $100/month for 3B. Meta doesn't publish how many tokens a typical errand costs, a gap Layer3 Labs points out. So you can't budget a workflow in "tasks" before you run it. The full tier comparison is in Muse pricing.
In practice, long jobs are where this bites. The census of a 77k Instagram story archive uses a resumable prompt that saves cursor state every 10 pages, which is the right design when you don't know whether the job will outlast your quota. See reducing agent costs for more patterns.
Shopping limits: Amazon and checkout
Amazon blocked Muse from buying on its store effective September 20. GeekWire reports that Amazon said the agent accessed its store without permission, didn't identify itself and appeared to store credentials. One builder's workaround is a self-hosted Amazon cart API Muse calls over SSH, which only adds to a cart; it doesn't route around the ban at checkout.
Elsewhere, purchases go through human approval. Meta's security write-up says all purchases requiring payment, and checkout pages with saved payment credentials, need your sign-off. That's a feature, but it means "buy this while I sleep" won't finish unattended.
Verification codes and phone steps
Agents stall where a human has to prove identity. In one Xfinity phone negotiation, Muse got through the phone tree to a human, hit a verification text it couldn't read, and patched the user into the live call. Plan for a handoff whenever a site sends a code to your phone.
Scale limits on big personal archives
The same Reddit user who sorted 18 Spotify playlists by decade (about 200 songs each) reports that a similar attempt on 40k Google Photos failed. Structured data behind a good connector works. Huge media libraries reached through a browser are much less reliable.
The agent's VM is small and closed
Each user gets a dedicated Muse Secure VM. Greg Ceccarelli, who got Claude Code running inside his Muse VM, says it has 2 vCPUs, about 8GB of RAM and 100GB of disk, and no inbound ports. He had to build a bridge to reach it. Heavy builds, local model inference or anything that needs to accept incoming connections belongs on your own machine. The homelab-via-Tailscale setup is one way to give Muse access to hardware you own.
Connector platform gaps
Custom connectors are built by Muse from any API and aren't reviewed by Meta. Directory connectors are reviewed. As of September 18, CellCog noted that no SDK, API spec or developer terms had been published, and fees and revenue share were undisclosed. Meta got more than 1,500 applications in under a week, so expect a queue.
Some connectors people expect are missing. Taylor Arndt's accessibility review names iCloud email as a gap that kept Muse out of her work.
Accessibility with VoiceOver
The same review found non-standard text fields and missing heading structure in chats when using VoiceOver. If your users rely on a screen reader, test with one before recommending Muse.
Mac security and permissions
Muse for Mac needs Full Disk Access, Automation and Notifications. On September 21 Patrick Wardle reported a zero-day where a local app changes an undocumented setting to redirect dictation traffic, exposing voice prompts and auth tokens (Malwarebytes). Meta hot-fixed it on September 22 by stripping the debug setting from production builds (Unite.AI). The lesson outlasts the bug: an agent with Full Disk Access is only as safe as the other software on your Mac. More in Muse security and privacy.
Model limits for developers
For Muse Spark 1.3 and Muse Glimmer, benchmark claims and hands-on results diverge:
- DataCamp measured a net 12% cost increase across three coding tasks, against Meta's claim of about 25% fewer tokens (test write-up).
- MindStudio's open-ended game test produced "a cube shooting at other cubes" (write-up).
- Glimmer has a 4,096 visual-token image limit, which is why Glimmer HD Vision tiles 4K images.
- The DFlash drafter gave only a 0.90%–1.16% speedup on an M5 Pro MacBook via oMLX, far below Meta's headline 1.8x on an M5 Max.
The benchmark gap gets its own post: Muse Spark benchmarks.
What the failures in the catalog have in common
We counted 11 agent builds in the catalog that describe something going wrong (as of September 28, 2026). Three were sites or payment processors blocking the agent, like Ryanair rejecting a US card; in two of those, Muse or the builder found a workaround. Three were the agent being confidently wrong: a closed restaurant, a wrong bus departure it then fixed itself, and Namecheap redirects it reported as set that weren't. The other five were human checkpoints or infrastructure: a verification text, a Tailscale tunnel that returned empty replies, Reddit's new-account filter, a 40,000-photo job and multi-agent coordination. The practical lesson: check the result, not the agent's report of it.
Quick reference
| Limit | Status (Sep 2026) | Workaround |
|---|---|---|
| Region / age | US and Canada, 18+ | None official; VPNs don't work |
| Amazon purchases | Blocked since Sep 20 | Other retailers, cart-only tools |
| Quota planning | Tokens, no task conversion | Resumable prompts, small pilots |
| Phone verification | Needs a human | Plan a handoff |
| VM resources | ~2 vCPU / 8GB, no inbound ports (as reported) | Tailscale or tunnel to your own box |
| Connector SDK | Not published as of Sep 18 | Custom connectors from an OpenAPI spec |
Start with small, reversible tasks. Start here has a sensible first week, and you can submit your own build if you find a way around one of these.
Frequently asked questions
Can I use Muse outside the United States?
Only in Canada, which got Muse on September 18, 2026. Everywhere else is still waiting: Meta hasn't published dates for the UK, EU or India, and region is checked on your account, so a VPN won't work.
Why can't Muse buy things on Amazon?
Amazon blocked Muse purchases effective September 20, 2026. GeekWire reports Amazon said the agent accessed its store without permission and didn't identify itself. Other retailers, including Shop Pay merchants, Walmart and Best Buy, have announced support.
How many tasks do I get on the free Muse plan?
Meta doesn't say. Plans are measured in Muse tokens per week, and The Batch reports the free tier as up to 100M tokens a week. There is no published conversion from tokens to tasks.
Does Muse work with screen readers?
Partly. Taylor Arndt's VoiceOver review found non-standard text fields and missing heading structure in chats. Test it with your own assistive setup before relying on it.
Can Muse run tasks completely unattended?
Many, yes, including background work after you close the app. But purchases, checkouts with saved cards, sending data out of its VM and phone verification codes all stop for a human.
Is Muse safe to use on a Mac?
It's as safe as your Mac is. Muse for Mac needs Full Disk Access and Automation, and a dictation zero-day reported on September 21 let local malware redirect its traffic. Meta hot-fixed that on September 22, but keep the app updated and limit which apps it can act in.
Numbers throughout are as reported by the build authors or by Meta, not verified by shipwithmuse. Official documentation lives at muse.ai/platform.