Meta built every part of Muse except the till

Meta's new agent reads your email, books your travel, negotiates on your behalf and pays with your card. Then it asks for the card before you have done anything at all.

Muse launched on 8 September, free for most usage, with a Power plan at $20 a month and Maximum at $100. Meta built almost all of it in house. The model is Muse Spark, the isolated computer it works on is Muse Secure VM, and a separate Sentinel agent sits on that same machine and has to approve anything leaving it.

One piece belongs to somebody else. When Muse buys something it checks out with Link, built by Stripe, and Meta's own announcement calls Muse "the first AI agent covered by Link's purchase protections": free coverage for damaged or lost items, price drops, no-fee returns, a return guarantee on eligible purchases.

The interesting sentence there is about liability

Read it as an insurance policy rather than a feature list. When an agent buys the wrong concert tickets, who eats it?

A payments company, apparently.

Meta wrote its own model, its own sandbox and its own guard, then handed the one component that can generate a chargeback to somebody else (Stripe's balance sheet, Stripe's returns desk). That is not an oversight. That is a company thinking hard about what it wants to be holding the first time an autonomous agent misreads a date.

Meta charged money for the thing it normally gives away

The subscription is stranger than the Stripe deal. Meta has spent twenty years handing consumer software out free and charging advertisers for the attention. Here it has shipped a consumer product with two price tiers and a card on file at signup.

The reason sits three lines further down the same announcement. "Muse doesn't share a person's conversations or the data in their VM with Meta's ad systems." People can opt out of their interactions training Meta's models, too.

So Meta has walled off the richest first-party signal it has ever had access to, what you buy, what you cancel, what you negotiate, who you invite to dinner, from the machine that produces nearly all of its revenue. And then it charged for the walled-off product.

Those two facts are one fact.

You cannot promise an agent will never feed the ad system and also fund it with ads.

Walls like this one get moved

Our expectation is that the carve-out narrows before the price does. Meta amends the Muse language within eighteen months to let derived signals, not conversations, reach ad targeting, probably as an opt-in, probably called personalisation, and arriving in a settings screen rather than a launch post. If instead the pricing moves and the privacy wording holds, we had this backwards and we will say so here.

The fair counter is that Meta does not need Muse data at all. It already has Facebook, Instagram and WhatsApp, and the marginal value of knowing which tennis lesson you booked is probably small against the cost of visibly breaking a promise made in a launch post. TechCrunch spent half its piece on Meta's record with the FTC for that reason, and the worry is a reasonable one.

But an agent does not produce more of the same signal. It produces purchase intent with a confirmed outcome attached, which is the single thing an ad auction cannot buy anywhere else at any price.

The startups in this category all have a price

Instinct has raised $350m at a $2.5bn valuation. Town is in talks at about $1bn (led by Index, according to Newcomer). Both sell assistants that people like and nobody gets for nothing. The free tier in this category now belongs to a company that can put the product inside WhatsApp.

That is a hard thing to price against.

Muse rolls out in the US on iOS, Android and the web, with AI glasses to come. Watch the card requirement rather than the launch reviews. A free product that wants your card on file has already told you what it is.