Meta's Muse Wants the Keys to Your Mac's Messages
Meta's Muse assistant has arrived on the Mac with permission to read Messages, Calendar and Notes — a level of access that shifts the privacy conversation around the company's AI push.
Meta's Muse AI assistant has launched a Mac app that can access a user's Messages, Calendar and Notes, and Meta shares last closed at 665.75, down 2.43% on the day.
Meta's AI assistant, Muse, has landed on the Mac with an app that can reach into a user's Messages, Calendar and Notes. That is a meaningful escalation. An assistant living in a browser tab or a chat thread only knows what you type into it. An assistant wired into the operating system's personal data stores knows who you talk to, when you are free, and what you wrote down when you thought nobody was reading.
Writing for The Verge, the assessment is that Muse is genuinely effective — and also "a little creepy." Both things can be true at once, and in consumer AI they usually are. The features that make an assistant useful are the same features that make it unsettling, because usefulness here is a direct function of how much of your life the software has been handed.
Desktop access changes the privacy question
Messages, Calendar and Notes are not incidental files. Messages contains conversations with people who never agreed to have their words processed by an AI system. Calendar exposes a map of a person's professional and personal network, including meeting titles that often carry confidential context. Notes is where people keep drafts, passwords they should not be keeping there, medical details and half-formed ideas.
Granting an assistant read access to all three at once is qualitatively different from granting access to any one of them. Cross-referencing is where the value comes from — and where the exposure comes from. The relevant questions for any user are the same ones that apply to every assistant of this type: what is processed on the device, what leaves it, how long anything is retained, and whether the data touches model training.
Those are not questions the lead answers, and they are not questions users can resolve by reading a permissions dialog. They are the questions that will determine how this product is received over the next several months.
An assistant that cannot explain itself
One detail stands out as more than a quirk: Muse does not know how to describe itself. Jason Aten, a contributing editor at Inc Magazine, posted on Threads about the assistant, and the gap between capability and self-description is worth pausing on.
An AI product that cannot articulate what it is, what it can see, or what it does with what it sees is a product that cannot obtain meaningful informed consent from its user. If the assistant itself is the primary interface — and for a chat-based tool, it is — then the assistant's inability to state its own data practices is a design failure, not a cosmetic one. Users increasingly ask the model rather than read the documentation. When the model shrugs, the trust chain breaks at the first link.
This is also a brand problem specific to Meta. The company carries more accumulated scepticism about personal data handling than most of its rivals in AI assistants. A product that is vague about itself while sitting on top of a user's private messages is the worst possible combination for a company in that position.
Where Muse sits against rival assistants
The desktop is now contested ground. Assistants that only answer questions are commoditised; assistants that act on a user's own context are the differentiator every major platform is chasing. Meta's advantage in that race has never been the operating system — it does not own one on the desktop — so a Mac app that plugs into Apple's native data stores is a way of buying context it cannot get from hardware ownership.
Assistants that only answer questions are commoditised; assistants that act on a user's own context are the differentiator every major platform is chasing.
That is the strategic logic. It also explains the discomfort. When a platform company owns the OS, the data access at least stays inside one trust boundary. When a third party asks for the same access, the user is being asked to extend trust across a boundary, to a company whose core business has historically been built on behavioural data.
Things worth watching from here:
- Whether Meta publishes clear, specific retention and training policies for Muse's Mac-level data access, rather than folding it into a general AI privacy page.
- Whether processing happens locally on the device or is sent to Meta's servers — the single most consequential technical detail.
- How European and US regulators treat an assistant that ingests third-party message content without those third parties' consent.
- Whether the assistant is updated to describe its own capabilities and data handling accurately when asked.
How the market is treating Meta right now
Meta Platforms (META) last closed at 665.75, down 2.43% on the day from a previous close of 682.31, with a session range of 660.80 to 690.15, as of the last trade at 20:00 GMT on 18 September 2026. That is a wide intraday swing and a close near the low end of it — the stock gave up ground steadily through the session rather than gapping down and stabilising.
The move ran against the broader technology tape. The Nasdaq 100 tracker (NASDAQ: QQQ) closed at $721.45, up 0.63%, while the S&P 500 tracker (NYSE: SPY) finished at $761.69, down 0.12%, and the Dow tracker (NYSE: DIA) at $515.88, down 0.48%. A decline of that size in Meta on a day when the tech-heavy benchmark rose is a company-specific story rather than a sector one, though a single session's price action should not be read as a verdict on any one product launch.
The cost of context
The deeper issue Muse surfaces is not unique to Meta. Every assistant vendor has concluded that the path to a product people use daily runs through personal context — email, calendar, messages, files. There is no version of a genuinely helpful assistant that operates on nothing. The industry is effectively asking consumers to trade the contents of their private communications for convenience, and doing so with interfaces that make the trade feel like a routine permissions prompt.
What makes Muse's case instructive is the combination: real capability, broad access, and a product that cannot explain itself. For Meta, the reputational maths on AI products is tighter than it is for competitors, because the company starts from a lower baseline of public trust on data. Shipping an assistant with Messages access and no clear self-description narrows the margin for error further.
The practical advice for anyone installing it is mundane and unglamorous: grant the narrowest permission set that still makes the tool useful, and treat Notes and Messages access as the highest-stakes toggles in the list. If the assistant cannot tell you what it does with what it reads, that is information in itself.
Key facts
- META last close: 665.75, -2.43% (as of 20:00 GMT, 18 Sep 2026)
- Mac app data access: Messages, Calendar and Notes
- Nasdaq 100 (QQQ) close: $721.45, +0.63% same session
- Named commentator: Jason Aten, contributing editor, Inc Magazine, posting on Threads
Frequently asked questions
What is Meta's Muse?
Muse is Meta's AI assistant. It has been released as a Mac app that can access a user's Messages, Calendar and Notes. Early assessments describe it as an effective assistant, though one that feels somewhat unsettling given the breadth of personal data it is permitted to read on the desktop.
Why is Muse's Mac access considered a privacy concern?
Messages, Calendar and Notes together give an assistant a detailed picture of who a user communicates with, their schedule and their private written material. Message content also includes words from third parties who never consented. The combination of all three data sources, rather than any single one, is what raises the exposure.
What does it mean that Muse cannot describe itself?
Reporting notes that the assistant does not know how to explain what it is. Because users increasingly ask the AI directly rather than reading documentation, an assistant that cannot state its own capabilities or data practices makes informed consent harder to obtain. It is a design gap with consent implications, not merely a cosmetic flaw.
How did Meta stock perform in the latest session?
Meta Platforms last closed at 665.75, a fall of 2.43% from a previous close of 682.31, with the session ranging between 660.80 and 690.15, as of the last trade at 20:00 GMT on 18 September 2026. The stock closed nearer the bottom of that range than the top.
Did the broader market move the same way?
No. On the same session the Nasdaq 100 tracker QQQ closed up 0.63% at $721.45, while the S&P 500 tracker SPY slipped 0.12% to $761.69 and the Dow tracker DIA fell 0.48% to $515.88. Meta's decline ran against the direction of the tech-heavy benchmark that day.
What should users check before granting Muse desktop permissions?
The key unknowns are whether data is processed on the device or sent to Meta's servers, how long anything is retained, and whether content is used for model training. Users can also limit exposure by granting only the permissions needed, treating Notes and Messages access as the highest-stakes toggles.
