Meta's Muse Keeps Working After You Close the App, With a Keychain Device Due in December

Meta unveiled Muse, a personal AI, on September 8 local time. The service is built to handle everyday chores on a user’s behalf and to break longer-running goals into multiple steps. According to several reports, Meta then introduced a dedicated device and new conversational features at Meta Connect on September 23. The arc is clear: a personal assistant that started as an app is spreading across the devices people use in daily life.

What happens to your task when you close the app

At the heart of Muse is an agent, an AI that carries out work on its own. It runs on the Muse Spark model inside a virtual computer in the cloud that is provisioned separately for each user. It gets things done using a browser and connected apps, and it keeps working on the job you assigned even after you close the app.

The change here is that you no longer have to sit and watch the chat window. That does not mean it decides everything by itself. At sensitive steps, such as sending an email or making a purchase, it asks the user for approval. In effect, the authority to carry out work and the authority to make the final call are kept separate.

At the time of the announcement, a staged rollout was under way in the United States across iOS and Android apps and the web. Meta said users can also chat with it on WhatsApp. The company explained that ordinary levels of use would mostly be free, with a subscription plan for heavy users. It did not disclose specific subscription pricing.

The keychain device and the glasses: when can you use them?

Coverage from Connect focused on the growing number of ways to reach Muse. The Verge reported a dedicated email address for Muse along with a video chat feature. The Decoder also described a capability for controlling a Mac. The direction is a move beyond typing instructions into a chat window toward touchpoints such as email and video.

A dedicated device, Muse Charm, was also revealed. Ars Technica reported that the keychain-sized device is slated for release in December. Support for AI glasses is in the works as well. In the official September 8 announcement, however, glasses support was listed as something coming later. The Verge’s Connect coverage also described it as a feature still under development. The date a device is announced and the date you can actually use it are two different things.

What do 1.8 million downloads actually tell us?

The evidence of early interest is the download count. According to Apptopia figures cited by AI Times, Muse logged 1.8 million iOS downloads in the United States and Canada over the 13 days after launch. That is more than the 1.3 million ChatGPT recorded in the early days of its mobile release, compared within the same regions and operating system. Still, that number is only a count of how many times the app was downloaded.

The Decoder reported that Muse had more than 500,000 users in its first week. User counts and download counts are different metrics. Both figures indicate the level of interest, but neither shows how accurately the AI completes the tasks it is given. Meta’s official announcement included no performance figures such as a task success rate.

There has also been debate about whether Meta created the design on its own. TechCrunch reported that while Meta said it developed Muse from scratch, it acknowledged drawing heavy influence from the open source project OpenClaw. The outlet also relayed observations that file names and contents in some workspaces resemble each other. That alone is not enough to judge whether code was copied outright.

So can you trust it with your work?

In Muse, external connections and app actions are screened by a separate oversight AI called Sentinel. Meta stated that even with such safeguards, Muse can make mistakes or be influenced by malicious content. Ars Technica also reported a vulnerability that could have let an attacker take over another person’s Muse. Meta has fixed that vulnerability, Ars Technica said. Because the range of work Muse handles on a user’s behalf is so broad, approval safeguards and defenses against external content become key criteria for evaluation.

How far the company can see into your data deserves scrutiny too. Under the current setup, Meta can access data when it is needed for purposes such as operating the service. Muse Confidential VM, designed to prevent that, has not yet shipped. In Korea, AI Times reported that the domestic launch timing is undecided. The significance of Muse lies in an AI that does not stop at answering but goes on to carry out actual work. For users in Korea, the next criteria to check are Korean language and local service support, subscription terms, real-world task accuracy, and the level of data protection.