Muse: Meta's agent, on a computer of its own

Muse is an agent rather than a chatbot: you give it an errand or a goal and it does the work, in steps, without being walked through each one. Every user gets a dedicated virtual machine in Meta's cloud with a browser, a file system and a terminal, so it can search and fill in forms, write the small programs a task needs and build its own connectors to services that have an API. It keeps one long conversation rather than separate threads, tracks goals in a tab of their own, sends you messages when something changes, and asks before anything consequential — sending an email, making a purchase. What it produces comes back as files: documents, spreadsheets, PDFs, images, podcasts, single-page web tools. It is reached from the iPhone and Android apps, from muse.ai, from a Mac app that can be given access to your files and local apps, and from inside WhatsApp; Meta says AI glasses are next.

Underneath it is Meta's own model, Muse Spark — version 1.3, according to Meta's security write-up — and around it is more machinery than most agents publish. A separate process Meta calls the Sentinel approves every network request and every connector action, logins sit in a credential store the agent cannot read, and purchases go through Link by Stripe with a single-use card number tied to one merchant and one amount. Its personality and memory are plain files you can read and edit: Memory.md, Soul.md and Identity.md, the last described as the agent's name, creature and vibe. Those are the file names and fields of OpenClaw's workspace templates; Meta has not said where the convention came from. Muse is not a coding agent. It writes code for its own tasks, but Meta's coding agent is a separate product, Muse Code, sold separately.

The data terms are the part to read slowly, because the defaults point one way. Training on your conversations is on from the first message, covers what connectors and the Mac app bring in, and when switched off also stops the use of past interactions; Meta strips names, email addresses, phone numbers and Social Security numbers first. Conversations are not shared with Meta's ad systems, though the sites the agent visits see it as you. And Meta's own security write-up says today's design restricts staff access by policy without preventing it; a version designed to lock Meta out, Muse Confidential VM, is promised for later this year. Sign in from an account that shares an Accounts Center with Facebook, Instagram or Threads and all three connect to Muse automatically.

We have not run it. Every way in, the free one included, needs a Meta account, and opening one to review a product is not something this desk does. Everything here is documented evidence read on the date at the top: Meta's Help Center — the subscription articles and ten of its thirteen Muse articles — the Muse terms and privacy policy, Meta's subscription terms, the two Newsroom announcements, and the design and security write-ups Meta published at launch. Where those sources disagree — on where you can pay, and on whether an action can be undone — we print both.