Build an iCloud skill with a Muse agent
I Tried Meta’s Muse Agent. Task: I Tried Meta’s Muse Agent. I Really Want This to Work. Okay, so I want to talk about Meta’s Muse Agent because I finally sat down and tried it, and I have a lot of thoughts. The short version is that I like it. I actually really like the idea. I kept finding myself saying, “that’s interesting,” and then immediately trying to hand it something else to do. Taylor’s Substack is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. But I also spent a lot of time fighting VoiceOver, and I hit one problem that matters a lot for me: I cannot connect enough of the systems I actually use for work. So, this ended up being one of those reviews where I am excited and frustrated at the exact same time. A little context first. I had already been following Meta’s Muse work because of Muse Glimmer, the open 30-billion-parameter agentic model Meta released in August. I have been really impressed with Glimmer. Then Meta launched Muse as the much more consumer-facing version of this idea: a personal AI agent in the cloud that can browse the web, connect to apps, keep working in the background, and take actions with your approval. Meta introduced the product on September 8, 2026. That is what got me interested. I have tried the more technical agent stuff. It can be really cool, but a lot of it still assumes you want to set up servers, configure tools, manage credentials, and basically become the IT department for the AI that is supposed to save you time. Muse looked like the opposite of that. Open it, connect things, talk to it, and let it work. So, I decided to see if I could use it for both personal stuff and work. The setup made me think, okay, this is actually approachable When I first signed in, Muse explains some of the important stuff in pretty normal language. It can take actions for you. It can make purchases. It can send messages. It can work across apps. It also tells you that AI can make mistakes. Obviously AI makes mistakes, but I liked the way Meta explained it. It did not feel like a page written for AI researchers. That was one of the first things that made me go, okay, this is cool. A lot of agent systems are still very, very technical. Muse was trying to make the whole thing approachable. Then the accessibility problems started And then I got to the actual interface. I decided to name my agent Lyra. I like the name. Getting the name into the field should have been the easiest thing in the world, and instead it took me a few tries because the text field was not behaving like a normal text field with VoiceOver. On the web I was hearing things like message, attach files, dictate, send, all as this kind of combined thing. I could not just get into the field, type, edit, and know where I was. At one point VoiceOver was reading characters in the name in a way that made me think I had an extra letter. I was afraid I was deleting the wrong thing. I finally got it fixed, but this is exactly why I keep talking about standard controls. If you use the standard text editing behavior, I already know how to use it. My screen reader already knows how to use it. You do not have to reinvent all of that. Then I got into the chat and realized the chat messages are not headings. I am gonna be honest, I was really mad about this. I kept moving through the page and getting to things like Connect your apps, but the actual conversation did not have the heading structure I wanted. If you do not use a screen reader, that probably sounds tiny. For me it is not tiny. This could become a huge chat over time. I need a way to jump through the conversation quickly. Headings are one of the main ways I do that. I have written about this exact pattern before: something looks like a heading visually, or it sits inside a nice region, but the semantic structure underneath is missing. That means I get the design without the navigation. I also need Muse to tell me when something happened There was another thing that kept bothering me: I did not always know when Muse had responded, saved something, or wanted something from me. I would love a sound. I would love a VoiceOver announcement. Give me something. This matters even more for an agent than it does for a regular chatbot. If the whole point is that Muse can go away and work on something, I should not have to sit there checking the chat to figure out whether it finished or asked me a question.; tools: Meta Muse. Human role: The original post does not say how a person stayed involved.; limitations: Meta Muse currently lacks a supported mechanism for custom tools to access persistent, user-approved credentials, limiting their utility to single sessions. The web and Mac interfaces exhibit critical accessibility flaws, including non-standard text fields and missing ARIA headings, which impede screen reader users. Integration is restricted to a predefined list of connectors, leaving niche or personal systems like iCloud email unsupported without complex workarounds..
Tools
Published on Agent Work Atlas:
Overview
I Tried Meta’s Muse Agent.
Workflow
I Tried Meta’s Muse Agent. I Really Want This to Work. Okay, so I want to talk about Meta’s Muse Agent because I finally sat down and tried it, and I have a lot of thoughts. The short version is that I like it. I actually really like the idea. I kept finding myself saying, “that’s interesting,” and then immediately trying to hand it something else to do. Taylor’s Substack is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. But I also spent a lot of time fighting VoiceOver, and I hit one problem that matters a lot for me: I cannot connect enough of the systems I actually use for work. So, this ended up being one of those reviews where I am excited and frustrated at the exact same time. A little context first. I had already been following Meta’s Muse work because of Muse Glimmer, the open 30-billion-parameter agentic model Meta released in August. I have been really impressed with Glimmer. Then Meta launched Muse as the much more consumer-facing version of this idea: a personal AI agent in the cloud that can browse the web, connect to apps, keep working in the background, and take actions with your approval. Meta introduced the product on September 8, 2026. That is what got me interested. I have tried the more technical agent stuff. It can be really cool, but a lot of it still assumes you want to set up servers, configure tools, manage credentials, and basically become the IT department for the AI that is supposed to save you time. Muse looked like the opposite of that. Open it, connect things, talk to it, and let it work. So, I decided to see if I could use it for both personal stuff and work. The setup made me think, okay, this is actually approachable When I first signed in, Muse explains some of the important stuff in pretty normal language. It can take actions for you. It can make purchases. It can send messages. It can work across apps. It also tells you that AI can make mistakes. Obviously AI makes mistakes, but I liked the way Meta explained it. It did not feel like a page written for AI researchers. That was one of the first things that made me go, okay, this is cool. A lot of agent systems are still very, very technical. Muse was trying to make the whole thing approachable. Then the accessibility problems started And then I got to the actual interface. I decided to name my agent Lyra. I like the name. Getting the name into the field should have been the easiest thing in the world, and instead it took me a few tries because the text field was not behaving like a normal text field with VoiceOver. On the web I was hearing things like message, attach files, dictate, send, all as this kind of combined thing. I could not just get into the field, type, edit, and know where I was. At one point VoiceOver was reading characters in the name in a way that made me think I had an extra letter. I was afraid I was deleting the wrong thing. I finally got it fixed, but this is exactly why I keep talking about standard controls. If you use the standard text editing behavior, I already know how to use it. My screen reader already knows how to use it. You do not have to reinvent all of that. Then I got into the chat and realized the chat messages are not headings. I am gonna be honest, I was really mad about this. I kept moving through the page and getting to things like Connect your apps, but the actual conversation did not have the heading structure I wanted. If you do not use a screen reader, that probably sounds tiny. For me it is not tiny. This could become a huge chat over time. I need a way to jump through the conversation quickly. Headings are one of the main ways I do that. I have written about this exact pattern before: something looks like a heading visually, or it sits inside a nice region, but the semantic structure underneath is missing. That means I get the design without the navigation. I also need Muse to tell me when something happened There was another thing that kept bothering me: I did not always know when Muse had responded, saved something, or wanted something from me. I would love a sound. I would love a VoiceOver announcement. Give me something. This matters even more for an agent than it does for a regular chatbot. If the whole point is that Muse can go away and work on something, I should not have to sit there checking the chat to figure out whether it finished or asked me a question.
Reported result
The custom iCloud skill was created but proved unusable for long-term automation due to the lack of persistent credential storage. The author reported that the agent's core capabilities were impressive, but the interface's accessibility failures and connector gaps prevented reliable daily use. No quantitative metrics on time saved or task completion rates were provided, as the primary outcome was a functional demonstration that failed to meet the persistence requirement.
Human role
The original post does not say how a person stayed involved.
Limitations
Meta Muse currently lacks a supported mechanism for custom tools to access persistent, user-approved credentials, limiting their utility to single sessions. The web and Mac interfaces exhibit critical accessibility flaws, including non-standard text fields and missing ARIA headings, which impede screen reader users. Integration is restricted to a predefined list of connectors, leaving niche or personal systems like iCloud email unsupported without complex workarounds.
Try it
Act as a Meta Muse user attempting to automate email monitoring. 1. Log in to Meta Muse and navigate to the chat interface. 2. Identify that iCloud is not in the default connector list. 3. Prompt Muse: 'Build a custom iCloud skill to check my email for urgent messages.' 4. Observe the generated tool and attempt to authenticate using an app password. 5. Close the session and reopen Muse to test if the credential persists. 6. If persistence fails, document the error and the need for re-authentication. 7. Simultaneously, use VoiceOver to navigate the chat history and note any missing semantic headings or unlabeled buttons. 8. Output a report detailing the tool's functionality, the persistence failure, and specific accessibility barriers encountered. Do not send real emails or make purchases during this test.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.