Anthropic launched Personal Finance in Claude on Thursday — a beta for Pro, Max and Ultra subscribers in the US, on web, desktop, iOS and Android. Users link bank, credit card, loan and investment accounts through Plaid, the aggregator most banking apps already use, then ask questions about their own money. The constraint that matters is stated plainly in the documentation: Personal Finance can only read. It cannot move, manage or invest anything.
What Claude actually sees
For each account you choose to share, Claude reads the bank name, account name, type and last four digits, balances and credit limits, and up to 24 months of transactions — amount, date, merchant, category and the bank's own description. It does not get full account numbers, unless the bank writes one into an account name or a transaction description. For investments and loans it reads balances, not holdings, trades, rates or terms. You can link up to 16 banks; sign-in happens inside Plaid, so Claude never sees your bank password.
The dashboard shows spending, cash flow, accounts, subscriptions and upcoming bills as reorderable widgets. Questions like "where did my money go last month?" or "which subscriptions am I paying for?" work on the page or in an ordinary chat.
The privacy details that matter
Conversations inside the Personal Finance hub are not used to train models. In that hub, Claude is also designed not to save anything to memory or make the chat searchable later, and to leave it out of the monthly recap — including the title. Say the same things in a normal chat and your ordinary privacy settings apply.
Two caveats are uncomfortable and unusually well documented. First, a shared chat link shows the messages and the figures from your accounts, and an artifact can read the linked accounts of whoever opens it, provided that person has Personal Finance enabled. Second, balances can be a day or two stale, and an account linked through two logins can be counted twice, which can make a budget look like it has more money left than it does. Anthropic states outright that Claude is AI and can make mistakes, and that important figures should be checked against the bank's own app.
Limits
US banks only. Not usable in Claude Code, scheduled tasks, Cowork tasks or projects; not on Free, Team or Enterprise plans; not in Claude Duet. Data exports do not yet include the dashboard. Unlink your last bank and the dashboard, widgets, budgets and setup answers are deleted permanently.
Why this is more than a feature
The assistant has moved from "text I typed" to "data that exists about me", and the engineering answer is a permission class: read-only, chosen accounts, explicit linking, revocable at any time. Plaid does the credential work so the model never holds a password — the same pattern that made fintech integrations tolerable in the first place.
What is genuinely new is the surface area. A chat window that can quote your transactions is also a chat window that can leak them into a shared link or an artifact, and Anthropic documenting that in the help centre, rather than a footnote, is the right call. Whether people trust a model with their ledger depends less on benchmark scores than on how visible the line between reading and acting stays — and today that line is drawn by product design, not by the model's judgement.
