Muse got the agent right
I think Muse is the best consumer AI product Meta has ever shipped, and one of the best anyone has. This page is about what it got right, and the one place it stopped short.
What it got right
Muse is an agent, not a chatbot. You give it a goal and it handles the rest: it sends email, books travel, negotiates bills, fills forms in a real browser, builds grocery lists from your recipes, and checks out with a single-use card. Every send and every purchase waits for your approval. There is a full audit trail. People name it, give it a face, and Meta's own design notes say that is the feature people light up about most.
The architecture underneath is the part I want to dwell on, because it is the reason this proposal is possible. Every Muse user gets their own persistent Linux computer, a Muse Secure VM, with its own browser, its own files, and a credential store the agent cannot read. A separate Sentinel agent is the sole permission authority: Muse proposes, Sentinel approves. Memory lives in editable files you can open, edit, and ask it to forget.
Meta already promised the next step. A Muse Confidential VM, encrypted with a key only you hold "so not even Meta can access it," is due later this year. That sentence concedes the entire argument of this site: the right trust bar for a personal agent is that the company that made it cannot see inside.
And the market answered. 2.8 million installs and 642,000 US daily users in twelve days, ahead of ChatGPT's first twelve days on iOS. The number one free app. Muse on Mac followed nine days after launch. Connectors for developers opened on day ten.
Where it stopped
Your Muse computer is in Meta's cloud. At launch Meta can technically read it. It runs a closed model, Muse Spark 1.3, that only Meta can serve. Your memory file is yours to edit and download, but it does not port anywhere. The interface is a text thread with tabs. The developer platform is a directory Meta curates, with no published protocol, no pricing, and no terms. There is no on-device inference, and no home hardware.
None of these are mistakes. They are the fastest way to ship a great agent in 2026. But each one is a place where Muse is halfway to the version of itself that wins the decade.
The tell
Thirteen days after launch, Amazon blocked Muse from shopping on its site. The reasons given were that it does not identify itself as an agent and reads purchase history without disclosure. Those are fixable. The deeper problem is structural: a cloud agent acts from Meta's IP space, under Meta's identity, and every platform on the internet will treat it as Meta knocking on the door.
An agent that runs from your house, on your network, with your identity and your credentials in a box you own, is a very different visitor. It is you.
One more thing Meta already built
On August 10 Meta released Muse Glimmer: a 30 billion parameter model distilled from Spark, Apache 2.0, post-trained for agentic tasks, that runs in under 20GB of memory on a single consumer GPU or an M-series Mac, and ships with its own drafter model for fast decoding. Meta built the local model for the box. It just has not built the box.
Sources
- Meta, Introducing Muse (2026-09-08)
- Meta AI Research, how we built safety into Muse (2026-09)
- Meta, How we designed Muse (2026-09)
- Meta AI Research, Muse Spark 1.3 (2026-09-02)
- Meta AI Research, Muse Glimmer (2026-08-10)
- Meta, Muse Platform connectors (2026-09-19)
- TechCrunch, Muse is outpacing ChatGPT's early mobile launch (2026-09-21)
- TechCrunch, Muse hits Mac (2026-09-18)
- SiliconANGLE, Amazon blocks Muse (2026-09-21)
- TechCrunch, will consumers trust it (2026-09-08)
- Meta Help Center, how Muse handles privacy (2026-09)