The error said the key was invalid. We made a fresh key and entered it again. Same error. That was the moment our test of Meta Muse stopped being about what it could do and became about who fixes it when it can't.
We'd gone in with an assumption plenty of owners share: if a business lives on Facebook, Instagram and WhatsApp, the agent built by the company that owns those platforms is the obvious one to wire the whole business to. For a good part of the work, Muse was the natural choice. Then we asked it to reach two ordinary business systems.
This is a field note, and we mean it as one. Muse is a very good personal agent. What we say about the product beyond our own test comes from Meta's help page and independent write-ups, as of September 2026.
What Muse does well
Start with speed. In our use, you ask and the work comes back quickly enough that you keep asking.
The Meta side is as easy as advertised. Facebook, Instagram and Threads connect automatically when the accounts share an Accounts Center. Nothing to configure, no keys to copy. Beyond Meta's own platforms, connectors include Gmail and Google Workspace, along with personal sources such as Apple Health and Android SMS.
The defaults are sensible too. Meta says that by default Muse won't take many important actions, like sending an email, without your approval, and that many connectors can be set so Muse retrieves data but takes no actions.
An independent review that ran ten business tasks, published September 27, 2026, scored meeting prep and background tasks highest, and said business briefs, content plans and social copy got sharply better once Muse could see the reviewer's Facebook, Instagram and Threads context. It gave 8.0 of 10 across the scored tasks and 6.5 of 10 as an overall impression. In our own use, the Meta side and the personal-assistant work were as good as advertised. There's a plan that costs nothing, with paid tiers above it. We haven't seen per-task pricing published.
What we asked it to do next
A services business doesn't run on social accounts alone. It runs on systems of record: the place the customer history lives, the place the job data lives. So we asked Muse to connect two ordinary business systems, a memory service and a database, through its custom connectors.
You ask Muse to build a connector, it retrieves the service's API information and guides you through it, and your credentials go into what Meta calls a Secure Credentials Store. We created fresh API keys for both systems and entered them.
Both connections returned an authentication error saying the key was invalid. We created new keys and entered those. Same error, repeatedly. Meanwhile the same connection method worked for a third service, so the method itself wasn't broken.
To its credit, the agent reasoned about the failure instead of repeating itself. It concluded that either the key format didn't match what the service expected, or the stored credential wasn't being picked up on its side. It suggested we test the same key from our own machine, and it offered to send a private note to the platform team. A sensible response. The road ended there anyway.
The likely cause, in plain language
We can't see inside Meta's credential store, so treat this as the likely explanation and not a confirmed one.
A key is only half of a login. The other half is how the key is presented. One service wants it under one label. Another wants a particular word in front of it. A third wants it in a different place in the request altogether.
A credential flow that presents every key the same way will work for the services that happen to match and fail for the ones that differ. That fits what we saw: one service connected, two returned an invalid key error, and new keys changed nothing. Our conclusion is that nothing about the keys was wrong.
The part that failed is the part a business can't afford to have fail, and there was nobody to call.
How this lines up with what others found
Our result is one test. It sits alongside documented findings, all as of September 2026.
- The independent test stopped at a login too, by choice. When that reviewer asked Muse to publish to WordPress, it reached the login screen and offered to keep the credentials in its secure store. The reviewer stopped there, on a rule of never handing a password to an agent during a test. The same review reported that Muse once said an email hadn't been sent when it had. Its advice: a good fit for businesses centered on Facebook and Instagram, for research and for background work, and one to avoid where the job needs credentials or broad file access.
- Custom connectors are yours to vouch for. Meta's help page states: "Meta doesn't review custom connectors or how they use your information, so grant access with caution." When we read that page, it had no troubleshooting steps for a failed connection.
- It runs in Meta's cloud. A developer write-up from mid-September noted that Muse runs its integration code on a virtual machine in Meta's cloud, so a server on your own laptop or office network isn't reachable.
- The connector help page doesn't mention WhatsApp. When we read it in September 2026, Meta's connector help page described no WhatsApp connection, for business use or otherwise.
- Trust is still being worked out. On September 20, 2026, Amazon blocked Muse from shopping on its site, saying the agent doesn't identify itself while browsing and raising privacy and credential concerns. Meta's position is that Muse has no visibility into people's passwords or payment methods.
None of this makes Muse a bad product. It's a consumer product, and a good one.
Our verdict, and what a services business needs on top
Muse is a very good personal agent. On the evidence of our test, we wouldn't yet wire a company's systems of record to it.
What our test exposed was less a bug than a missing role. When a connection fails, someone has to read the service's instructions, fix it and tell you when it's working again. A services business needs three things on top of any agent.
- Someone accountable for the connections. A name and a phone number, where we had a help page.
- Approved sources. A decision, made by your people, about what the agent may read and where its answers come from.
- Review points. A person at the decision point before anything reaches a customer or touches money.
We call that an AI department. Nexus is the one we install: a digital employee in the channel your team already uses, working from approved sources, with the work landing in your own folders. We compared the wider field in the best AI employee alternatives for small services businesses.
We'd still set Muse up for a client. If the two-day audit shows that the costly problem is marketing and research on Meta's platforms, Muse is a fine answer, and we'll say so and help set it up.
Frequently asked questions
Is Meta Muse good for a small business?
For marketing, research, meeting prep and content planning, yes, especially if the business runs on Facebook or Instagram. As of September 2026, an independent ten-task review advised against it where a job needs credentials, and in our own test two custom connections to business systems failed at sign-in.
Why did the custom connectors fail?
We can't confirm the cause. The likely explanation is that different services expect a key to be presented in different ways, and a credential flow that presents every key one way will fail on the services that differ.
Can Muse handle WhatsApp for a business?
As of September 2026, Meta's connector help page doesn't describe a WhatsApp connection. We wouldn't plan a customer workflow around it until it does.
Would Advisy set up Muse for a client?
Yes, when the two-day audit points there. If the work that matters most is marketing and research on Meta's platforms, Muse fits.
Who picks up when the connection breaks?
An agent can be fast and still leave you alone at the moment a connection breaks. In a company that owns its AI department, a named person picks up, and the owner knows the name.
Make a list of the systems your business can't run without for a day. Those are the ones that need a person behind the connection. Tell us how your business runs today, and the two-day audit will sort which workflows a personal agent can carry and which need more around them.
Sources, checked September 2026: Meta's help page on Muse connectors, the AI Agents Library ten-task review of Muse, the Parallel write-up on Muse custom integrations and Implicator's report on Amazon blocking Muse. The connector test described here is our own.
