Meta Muse: The Complete Guide to Meta's AI Agent App (2026)

Meta has used the "Muse" name for so many things this year that searching for it is confusing. Muse Spark is a model API. Muse Glimmer is an open-weights model. Muse Code is a terminal coding agent. The product that took over the App Store charts in September is simply called Muse. It is a personal AI agent app built for consumers, not developers, and it is the most direct attempt yet by a big platform company to ship an assistant that acts for you instead of only answering questions.

This guide covers what Muse does, how Meta says it keeps your data and money safe, what it costs, the Mac app and developer connector program that followed in the next ten days, and how it compares with ChatGPT, Claude Cowork and Gemini Spark. Where a number comes from a news report rather than Meta, we say so. Several details Meta simply hasn't published.

What is Meta Muse?

Muse is a personal AI agent that Meta announced on 8 September 2026. Meta's pitch is an assistant that "doesn't just answer questions" but does the work. In practice you describe an outcome, such as "find a cheaper car insurance quote", "book dinner for four on Friday" or "turn this recipe into a grocery order". Muse then plans the steps, opens a browser, fills out forms, sends messages and comes back when it needs you.

Three things set it apart from a chatbot:

• It works in the background. Tasks keep running after you close the app. That is the difference between an agent and a chat window.
• It remembers. Meta says Muse "remembers what matters to a person" and makes suggestions without being asked, based on what you have saved. You can tell it to forget specific things.
• It keeps a record. Every user gets "a complete audit trail of everything it has done and plans to do".

Muse runs on Muse Spark 1.3. Meta's launch post only says "Muse Spark, Meta's most capable model to date". TechCrunch names version 1.3.

Muse vs Muse Spark vs Muse Glimmer vs Muse Code: what's the difference?

Meta Superintelligence Labs now brands almost everything "Muse". Here is how the four products you are most likely to confuse fit together:

If you want to build on Meta's models, read our Muse Spark 1.3 guide for the API, the Muse Glimmer guide for self-hosting, and the Muse Code guide for the CLI. For the bigger picture of Meta returning to open weights, see Meta's open-weights return. The rest of this article is about the app.

What can Meta Muse do?

Meta and launch coverage describe these task types:

• Email and messages. Draft and send email, send invitations, and use WhatsApp and Instagram through Meta's own connectors.
• Travel and bookings. Book travel and restaurant tables. OpenTable is among the listed connectors.
• Forms and admin. Open a browser and complete multi-step forms for you.
• Bills and negotiation. Meta says Muse can "negotiate on their behalf", and TechCrunch lists lowering bills as a launch use case. Business Insider, via Pulse2, reported one user saving $3,500 a year on car insurance. That is one anecdote, not a benchmark.
• Shopping. Turn recipes into grocery lists and buy things. Checkout goes through Link by Stripe.
• Planning and research. Build plans, research services (TechCrunch mentions finding dog sitters) and pull context from connected apps.

Early reports also note limits. Business Insider found Muse could only see the first frame of Instagram videos. Meta has not published a list of tasks Muse refuses or cannot do, so expect it to fail on sites that block automation or need a login you have not connected.

How does Muse work? The Secure VM and Sentinel

The architecture matters more than the feature list, because Muse acts with your accounts and your money. Meta describes three layers:

1. Muse Secure VM. Each user gets "a dedicated, virtual machine (VM) that houses both the agent and a person's data". It includes its own browser, which is where Muse fills forms and clicks through sites. Independent write-ups describe it as a Linux VM with a filesystem and terminal. Meta has not documented that level of detail itself.
2. Sentinel. A separate agent controls what leaves the VM. Meta's wording: "Nothing Muse does reaches the internet unless the Sentinel approves it." In effect it is a second model acting as a gatekeeper on the first, the same pattern security teams use for agent egress controls.
3. Credential isolation. Meta says Muse has "no visibility into people's passwords or payment methods". Payments go through Link by Stripe, which comes with purchase protections. Meta says Shop Pay and 1Password support are "coming soon".

One caveat is important. These protections are operational: Meta runs the VM and the Sentinel. Meta has announced a future Muse Confidential VM in which "the whole VM, including a person's data and conversations with Muse, is encrypted with a key only they hold, so not even Meta can access it." It has not shipped and has no date. Until it does, you are trusting Meta's policies, not cryptography.

Where is Meta Muse available?

Muse is US only. Meta has not given a timeline for other countries. That is the biggest practical difference from Google's Gemini Spark, which already reaches Pro subscribers in more than 160 countries.

How much does Meta Muse cost?

Meta's launch post says only that Muse is "free for most of what people need, with subscription plans". The tier names and prices come from TechCrunch's launch report:

Neither Meta nor any outlet we could find has published per-tier quotas: how many tasks, how long they can run, or how many run at once. Treat the tiers as usage buckets, not feature gates, until Meta documents them. The structure matches the rest of the market: $20 is the standard consumer AI price, and $100 matches Claude Max 5x and the entry Google AI Ultra plan.

Muse subscription pricing is separate from the developer API. If you call Muse Spark 1.3 directly, you pay per token ($1.25 input / $4.25 output per million on the standard tier). The Muse Spark 1.3 guide covers that, including the cheaper Contributor tier.

What is Muse for Mac?

On 17 September 2026 Meta released Muse for Mac, a free macOS app, US only, that lets the agent work on your computer and not only in its cloud VM. According to MarkTechPost, it works across the built-in Files, Mail, Messages, Calendar and Notes apps. Meta's examples include:

• Organising a cluttered Downloads folder
• Finding a file you have lost track of
• Filling out forms using data from local files
• Summarising your messages and notes, for example an end-of-day recap
• Continuing a task across phone, laptop and WhatsApp while it runs in the background

Muse for Mac permissions and privacy

Local access is opt-in and adjustable. Full Disk Access is optional. Destructive or outbound actions, such as deleting files or sending messages, need your approval before they run. Meta frames this as "all with your explicit permission."

The launch still caused some concern. Social Media Today reported users saying Muse had read their private DMs on the Mac and suggested actions based on them. Meta's reply: "Muse can only do this if the user has granted Muse explicit permission to do so." Both statements can be true. If you give an agent broad access to Messages, it will use it. Grant access app by app, and skip Full Disk Access unless a specific task needs it.

Meta has not published a minimum macOS version or said whether Intel Macs are supported.

What are Muse Connectors?

Connectors are what give Muse abilities inside other services, such as reading a calendar, sending email or booking a table. At launch Meta highlighted WhatsApp, Instagram and Stripe Link. CellCog's review of the connector directory also lists Gmail, Google Calendar, Outlook, Plaid, OpenTable, Google Docs, Spotify, Function Health, Withings, Tailscale and Peloton, plus Muse-only connectors for Instagram, Threads, Facebook and Messenger.

On 18 September 2026, Meta opened the Muse Connectors developer program. Here is what is known:

• How to apply: a "Submit a connector" form at muse.ai/platform. You describe your product and how users would use it with Muse.
• Review: Meta checks submissions against functional, security and legal requirements and runs end-to-end testing. Approved connectors appear in the Muse directory, and some may get editorial placement.
• Two connector types: reviewed directory connectors, and custom connectors that Muse builds for an individual user without review.
• Payments: Meta has partnered with Stripe, so connectors can take payments through Link.

What has not been published: an SDK, an API spec, any mention of MCP (Model Context Protocol), fees or revenue share. Some coverage names Notion and Granola as early partners, but we could not confirm that on a Meta page. If you are a developer deciding whether to build, the honest status is: submit interest now and expect the technical contract later. If you already expose your service through an MCP server, keep it. Nothing Meta has said suggests Muse will use MCP, but nothing rules it out either. Our agent benchmarks overview covers how today's agents handle tool calls.

Is Meta Muse safe to use?

It depends on what you give it access to and how much you trust Meta's operating policies. The case for and against:

In Muse's favour: Muse cannot see passwords or card numbers. Payments go through Stripe Link, not a card stored with the agent. Outbound actions pass through a separate Sentinel. Every action is logged in an audit trail. Meta also says Muse "doesn't share a person's conversations or the data in their VM with Meta's ad systems."

The reasons for caution: TechCrunch's launch coverage points to Meta's record: FTC actions in 2011 and 2019 (the latter a $5 billion penalty), 2023 charges over violating its privacy order, the 2019 discovery of passwords stored in readable form, Cambridge Analytica, and a recent $18 billion child-safety settlement with 29 states. The end-to-end encrypted Confidential VM has not shipped yet. The ads promise covers ad systems, and Meta has not said the same about model training on consumer Muse data.

Practical rules we would follow:

• Connect the fewest accounts that make the task work. Add more only when you need them.
• Read the audit trail for the first few tasks of each new type before you trust it with that type unattended.
• Keep approvals on for anything that sends, deletes or buys.
• Don't connect work accounts (company email, internal docs) unless your employer has approved it. A consumer agent with broad inbox access is a data-handling question for your security team, not just for you.

How does Meta Muse compare to ChatGPT, Claude Cowork and Gemini Spark?

TechCrunch names Google's Gemini Spark and Anthropic's Claude Cowork as Muse's direct competitors. OpenAI has changed course: ChatGPT's standalone agent mode was removed in early August 2026, and OpenAI's help centre now points users to ChatGPT Work for multi-step tasks and a cloud browser for browsing workflows (see the OpenAI community thread). The table below only includes facts we could source. No independent benchmark compares these consumer agents directly.

Sources: Anthropic's Claude pricing page (Cowork on Pro and Max), Google's AI subscriptions post and Chrome Unboxed for Spark, and our Gemini Spark guide.

How to read the comparison:

• Muse is the only one with a real free tier. That, plus Meta's distribution through WhatsApp and Instagram, is why it reached #1 so fast.
• Muse is the most consumer-focused. Stripe checkout, bill negotiation and grocery carts are errands. Cowork is aimed at knowledge work: files, documents, research.
• Muse is the most geographically limited. Anyone outside the US can't use it yet.
• Model quality is roughly comparable on paper. VentureBeat reports that Muse Spark 1.3 at xhigh effort ties Claude Opus 5 on the Artificial Analysis Intelligence Index. There is no published head-to-head against the newest releases, so see our Claude Opus 5.5 guide and Grok 4.7 guide for what those models add.

How popular is Meta Muse?

Very popular, very quickly. Business Insider reported that Muse reached #1 among free apps on the US App Store about a week after launch, ahead of ChatGPT. On 21 September TechCrunch reported Apptopia estimates for Muse's first 12 days:

These are third-party estimates, not figures Meta has released. The 95% Facebook overlap explains the growth: Muse is spreading through Meta's existing audience. That also means these numbers say little yet about whether people outside Meta's ecosystem will adopt an agent from Meta.

What to watch at Meta Connect 2026 (23–24 September)

Meta Connect runs 23–24 September in Menlo Park, with Mark Zuckerberg's keynote at 4pm PT on the 23rd. Meta's agenda lists Muse Spark and Muse Glimmer sessions. Previews from Engadget and others expect:

• Muse on AI glasses. Meta already calls this "coming soon", so it is the most likely announcement.
• New glasses hardware and a lightweight mixed-reality headset, reported under the codename "Phoenix".
• Possibly news on the promised Muse Spark open-weights release. Zuckerberg said it is coming "soon" but has given no version, size, licence or date.

These are expectations, not announcements. We will update this guide after the keynote, and in particular we will add any glasses integration, new regions or published tier limits.

Should you use Meta Muse?

If you are in the US and have errands that eat your week, such as comparing quotes, rebooking, filling out forms or chasing refunds, the free tier costs nothing to try. Start with low-stakes tasks and read the audit trail. If your main need is document and file work, Claude Cowork is built more for that. If you are outside the US, Gemini Spark is the realistic option today.

For developers, the more useful questions are about the model and the platform. You can build on Muse Spark 1.3 today through the API, and you can join the queue for Muse Connectors. But without a published SDK or spec, it is too early to plan a roadmap around the connector program.

If your team is building agent features on these models and needs more engineers, Codersera can help you hire vetted remote developers who have shipped LLM integrations.

FAQ

What is Meta Muse?

Muse is Meta's personal AI agent app, launched on 8 September 2026. It performs tasks for you in the background: sending email, booking travel, filling out forms, negotiating bills and buying things through Link by Stripe. It runs on Meta's Muse Spark 1.3 model inside a dedicated cloud VM for each user. It is available in the US on iOS, Android, muse.ai, WhatsApp and Mac.

How much does Meta Muse cost?

Muse has a free tier with a usage meter. TechCrunch reports two paid plans: Power at $20 a month and Maximum at $100 a month. Meta itself says only that Muse is "free for most of what people need, with subscription plans", and it has not published the usage limits for each tier.

Is Muse the same as Muse Spark?

No. Muse is the consumer agent app. Muse Spark is the closed AI model behind it, which developers can call through Meta Model API as muse-spark-1.3 for $1.25 input and $4.25 output per million tokens. Muse Glimmer is Meta's separate open-weights model, and Muse Code is its terminal coding agent.

Is Meta Muse available outside the US?

Not yet. Muse launched in the US only, on every platform, including Muse for Mac. Meta has not announced a timeline for other countries. Google's Gemini Spark and Anthropic's Claude Cowork are currently the widely available alternatives outside the US.

What can Muse for Mac do?

Muse for Mac, released on 17 September 2026, lets the agent work inside Files, Mail, Messages, Calendar and Notes. It can organise Downloads, find lost files, fill forms from local data and summarise messages. Access is opt-in, Full Disk Access is optional, and deleting or sending requires your approval. It is free and US only.

Can Meta see my data in Muse?

Meta says Muse can't see your passwords or payment methods and doesn't share conversations or VM data with its ad systems. But the current Secure VM protections are policies Meta enforces, not encryption only you control. Meta has announced a Muse Confidential VM, encrypted with a key only you hold, but it has not shipped yet.

Sources

• Meta, Introducing Muse, a personal AI agent (8 Sept 2026)
• TechCrunch, Meta debuts its Muse AI agent (8 Sept 2026) and Muse is outpacing ChatGPT's early mobile launch (21 Sept 2026)
• MarkTechPost, Meta launches Muse for Mac
• Social Media Today, Meta expands Muse agent connections, launches Muse for Mac
• CellCog, Muse Connector Platform: what Meta opened to developers
• Meta, Connect 2026 agenda