Three weeks in. What stuck.
We installed the Meta AI Mac beta on day one and used it for ordinary work rather than demos. One feature earned a permanent place. One is cleverer than it is useful. And the whole thing is free, 16 MB, and better behaved than most desktop AI apps.
Independent review not affiliated with Meta, not sponsored, and nobody gave us anything. One Mac, three weeks, ordinary work.
Worth installing, not worth rearranging your life for
It is free and it is 16 MB, so the bar for “worth it” is low and it clears it comfortably. The more interesting question is which parts survive contact with actual work.
System-wide dictation
Types into whatever has focus, not just the chat box. Within a week we were dictating replies, commit messages and notes without thinking about it. This alone justifies the install.
Attaching a window
Handing it a spreadsheet and asking what changed is genuinely impressive the first ten times. Then the snapshot model starts to grate, and we reached for copy-and-paste instead.
Launch at login
A menu-bar resident we did not want resident. Nothing wrong with it just an assumption about how much of your day this app deserves that we did not share.
One Mac, three weeks, ordinary work
This is a hands-on account rather than a benchmark. We installed the beta on release day on a single Apple silicon Mac running macOS 15, granted all three permissions, and then used it for the work that was already in front of us: email, spreadsheets, notes, a fair amount of writing, and the sort of small lookups that would otherwise have gone to a search engine.
The tasks in the table below are the ones we did repeatedly, scored on how well the app handled them in the course of getting something done not on whether a carefully constructed prompt could produce an impressive answer. There is a difference, and it is the whole point of a three-week review rather than an afternoon one.
What we are not claiming: that one machine is a sample, that three weeks is a long-term test, or that our workload resembles yours. Read the task table and weight it by what you actually do.
Nobody at Meta was involved, we paid nothing because the app is free, and the site carries advertising but no relationship with Meta. During the test the model behind the app moved to Muse Spark 1.3, which changed a few things mid-flight noted where it matters.
Ten things we actually asked it to do
Scored out of five for usefulness in context. A low score is not always the model's fault sometimes the friction is in getting the material to it.
| Task | Rating | What we found |
|---|---|---|
| Tightening a rambling paragraph | 5 | The thing it is best at, and the thing we used most. Paste, ask for shorter, take the version you prefer. |
| Dictating a reply in Mail | 5 | Fast and accurate on the built-in microphone. The one feature that changed how we work. How it works. |
| Asking what changed in a spreadsheet | 4 | Reads an attached window well. Loses a mark because every edit means attaching it again. |
| Explaining an error message on screen | 4 | Share the window, ask what it means. Genuinely faster than typing the error into a search box. |
| Summarising a long PDF | 3 | Fine for the shape of a document, shakier on specifics. Independent testing found long-document accuracy slipped in 1.3, and we believe it. |
| Generating an image for a post | 4 | Quick and good enough to publish. Free-tier limits arrive sooner than you expect. What meters. |
| Drafting from a bullet list | 4 | Reliable, slightly over-formal by default. Tell it who it is writing to and it improves markedly. |
| Quick factual lookups | 3 | Convenient, and we still checked anything that mattered. Same caveat as every assistant. |
| Working across two windows at once | 2 | You cannot. One window per attachment, so comparing two documents means pasting one in. |
| Anything it should do on your behalf | 1 | Not this app. It reads and answers; it does not act. That is Muse, and it is a separate download. |
Ratings reflect our experience on one machine over three weeks against the 1.0 beta. A later build could move any of them.
Dictation is the reason to install it
We did not expect this. Window sharing is the feature that gets demonstrated, and it is the more novel of the two but dictation is the one we used every day without deciding to.
The difference is that it types into any app. Not into a chat box you then copy from, but directly into the Mail reply, the Slack message, the commit summary, the note. Once that is true, the friction of speaking instead of typing drops below the friction of typing, and the habit forms on its own. By the end of week one we were dictating first drafts of anything longer than a sentence.
Our rough tally: in the third week we used dictation dozens of times and attached a window twice. The feature that changed our behaviour was not the impressive one.
It is not flawless. Accuracy depends heavily on the microphone the Mac's built-in array was consistently better than Bluetooth earbuds, which macOS drops to a low-quality call profile when used as an input. Proper nouns and technical terms are where errors concentrate, so dictating prose and typing acronyms became the routine. And the permission it needs is the broad one, which is a real cost worth weighing. What Accessibility grants.
Six rough edges, in the order they annoyed us
Most are beta problems rather than design mistakes, which matters a design mistake is permanent and a beta problem is a to-do list.
The snapshot model is the real one. Attaching a window captures it once, so scrolling the document or editing a cell means attaching it again. In a demo you attach once and ask one question. In actual work you iterate, and after the fourth re-attachment you paste the content in instead which is what we ended up doing. Whether this is fixable depends on whether Meta considers a live view acceptable, and there are good privacy arguments for the current behaviour, so we would not bet on it changing.
The permission behaviour is the most confusing. macOS reads these grants when an app launches, so ticking a box while the app is running does nothing until you quit and reopen it. We lost twenty minutes to this on day one and have since watched everyone else lose the same twenty minutes. It is macOS being careful rather than a bug, but the app could say so. It comes up again after updates, which reset the grants.
Nothing crashed in three weeks. For a 1.0 beta of a desktop app from a company with no recent Mac track record, that is worth saying plainly.
The rest are smaller: Option-Space is claimed by Raycast, Alfred and half the launcher market, and there is no published shortcut list to consult; the long-document accuracy regression in 1.3 is real and documented; and image limits on the free tier arrive faster than the marketing implies. None of them made us uninstall it.
Scored where it counts
Assistant quality
Strong at rewriting and drafting, ordinary at long documents, honest about what it does not know more often than not.
4Desktop integration
Dictation is excellent. Window sharing is clever and awkward. No menu bar clutter, no background chatter.
4Performance and footprint
16 MB, native, instant to open, invisible when idle. Better behaved than most of its competitors.
5Setup and permissions
Three separate grants, a relaunch requirement nobody tells you about, and a sign-in that hands off to the browser.
2Reliability
No crashes in three weeks. The failures we saw were permission confusion rather than instability.
4Value
Free, with limits that only bite on image and video generation. Nothing to weigh up unless you generate in volume.
5A capable, well-made first release with one genuinely habit-forming feature, held back by a snapshot model that fights the way people work and a setup experience that loses twenty minutes of everyone's time. Free, so the only real cost is the permissions which is precisely the thing to think about before installing.
Scored against version 1.0 beta in September 2026. We will revise this page as builds land rather than leaving a stale verdict up.
Who it is for, and who should not bother
Install it if you talk more easily than you type
Dictation into any app is the whole case, and it is a strong one. If you already use Meta AI in a browser, this is a free upgrade to how you reach it.
Think twice if the permissions bother you
Screen Recording and Accessibility are broad, and without them the app is a browser tab in a nicer window. That is a reasonable thing to decline. What they grant.
Skip it entirely on an Intel Mac
It will not run, and no workaround exists. The browser version is the whole story there. Why, and what to do instead.
Review questions
Is the Meta AI Mac app worth installing?
If you already use Meta AI and have an Apple silicon Mac, yes it is free, it is 16 MB, and dictation alone justifies it. If you would be granting Screen Recording and Accessibility to an app you are unsure about, the browser version does the chat part without asking for anything.
Which feature turned out most useful?
Dictation, by a wide margin. It types into any app rather than just the chat box, and by the end of week one we were using it without deciding to. Window sharing was more impressive in demos and much less used in practice. Why.
What is the biggest weakness?
The snapshot model. Because the app captures a window once rather than watching it, every change means attaching it again and that friction was enough to stop us using a genuinely clever feature. The full list.
Is it better than the browser version?
For the same tasks, no the assistant is identical. The app adds reach: a panel over other apps, a window it can read, and dictation into any application. Those three things are the entire case for installing it.
Does it feel like a beta?
In places. Nothing crashed in three weeks, but permissions that appear granted and do not work until you relaunch, and updates that quietly reset them, are the kind of rough edges a 1.0 has and a 1.2 usually does not.
Is it fast?
Yes, and this is where it beats its competitors outright. A 16 MB native build rather than an Electron wrapper, instant to open, and genuinely invisible when idle. It is the best-behaved desktop AI app we have had installed.
Did the model change during your test?
Yes. Muse Spark 1.3 landed partway through, bringing better tool use and a documented slip in long-document accuracy. We noticed the second more than the first, which is reflected in the PDF summarising score.
Who should skip it?
Anyone on an Intel Mac, since it will not run; anyone unwilling to grant the permissions, who loses the only reasons to install it; and anyone who wants an assistant that acts rather than answers, which is Muse and a different product entirely.
Were you given anything by Meta?
No. The app is free, nobody at Meta was involved, and we have no relationship with them. The site carries advertising, which is disclosed, and it does not influence what we write here.
Will you update this review?
Yes, as builds land. A verdict on a public beta has a short shelf life, and we would rather revise the page than leave a stale score sitting at the top of it.
Free, light, and one habit-forming feature.
Three and a half out of five. Install it for the dictation, keep expectations low for the rest, and think about the permissions first.
Independent review not affiliated with Meta, not sponsored, nothing provided. Scored against version 1.0 beta on one Apple silicon Mac over three weeks. Check it runs on yours.