Claude Personal Finance went live on November 5 as a beta for US subscribers on the Pro, Max and Ultra plans, and its most notable design decision is what it refuses to do.
Bank accounts attach through Plaid, which means credentials never pass to Anthropic and the model never holds a banking password.
The permission set runs one way. Moving, managing or investing money is out of reach entirely, and on investment, loan and card accounts the only data returned is a balance and a credit limit.
Questions about spending across two years of transactions are in scope, along with duplicate charge and fee detection and dashboards built from balances. Sixteen US banks is the connection ceiling.
Restraint sold as the feature
Anthropic is marketing the absence of hands. A assistant that can transact inherits failure modes no benchmark covers, so the company picked the narrower and more defensible product.
Data handling follows suit. Conversations in the hub are excluded from training and kept out of the memory Claude shares across sessions.
The part worth reading twice
Anything a figure has been written into, whether chat, artifact or shared link, remains there until someone deletes it by hand. Unlinking a bank cuts off future reads. It does not pull back numbers already committed to a conversation.
That gap is easy to overlook in a release that otherwise signals caution at each turn, and a shared dashboard is exactly the surface that invites it.
This is the first Claude feature to handle regulated, high-consequence personal data at consumer scale. Anthropic has spent the year arguing that agents need receipts and bounded authority, and the release turns that argument into a product, with one loose thread deliberately left in view.
