A Reddit thread about Claude's new money feature pulled 264 upvotes and 76 comments in three hours. The top reply, sitting at 166 points, wasn't about privacy or convenience.
i tell myself i won't use it for security and privacy reasons, but i already know that i won't be able to help myself.
That's the product in one line.
Anthropic announced Personal Finance on November 5. You link a bank or card account through Plaid, then ask Claude where your money went.
Access is read-only. It can't move, manage, or invest a dollar.
The example in the announcement is oddly specific. Someone asks where their money went, Claude totals $4,309.37 across connected accounts, then flags $1,024.51 of food and drink and a membership renewal nobody remembered. Ask what's left before payday and you get $1,393.22.
I read the announcement, then the Hacker News thread, then the Reddit comments. The press covered it as a big launch. The actual reaction was messier than that.
Because the people most excited about this are the ones who already know it's a bad trade.
What you actually get
here's a question people always ask about these launches. is it just a chat window on top of my bank app?
Kind of. But the chat window is doing real work.
Anthropic says Claude can see the bank name, account name and type, last four digits, balances, credit limits, and up to two years of transactions. That's the whole list. No routing numbers, no passwords, no ability to move money.
And the questions it answers are the ones you'd ask a friend who's good with numbers.
Where did it go. Is this month weird. What's coming out before Friday.
The announcement leans on small stuff that's easy to miss. A charge that hit twice. A fee you didn't expect.
A bill that's been creeping up all year.
Ask what your APR cost you and it explains the term, then adds up the interest.
That's more than my bank app does. Mine shows a pie chart and calls it a day.
| Claude plus Plaid | Your bank app | DIY with Claude Code | |
|---|---|---|---|
| Setup | one Plaid login | already there | a weekend |
| Data leaves your machine | yes | no | no |
| Follow-up questions | yes | no | only if you built them |
The pipe has a history
The first time i tried to explain Plaid to someone, i called it the plumbing behind every app that asks for your bank login. That's still the best version.
Plaid is how Venmo, Robinhood, and a hundred budgeting apps read your accounts. It's also how Claude does it now. You sign in on Plaid's page, not Anthropic's, which is why Claude never sees your bank password.
Sounds clean. Here's the part the launch posts left out.
In 2021, Plaid settled a class action for $58 million. The suit claimed it over-collected banking data and wasn't clear that it was touching login credentials.
Around 98 million people were in the class. Payouts landed near $36 each.
Plaid changed its disclosures after that. The company says it never sold data.
But the point stands. Anthropic's privacy promise covers Claude. It does not cover the pipe.
Anthropic says chats inside the Personal Finance hub aren't used for training. That's a real commitment and i believe they mean it. It just doesn't answer what Plaid keeps.
The banks get a vote
what actually happens is the feature only exists while banks allow it. And banks are in a mood.
In late October, RBC, TD, CIBC, and BMO all said they'd block personal AI agents from direct access to customer accounts. Scotiabank said it would limit them. That was a week before Anthropic shipped this.
The distinction matters and people keep missing it. Those blocks target agents logging in as you. Plaid is a different door, one banks already accept.
But the mood is the story. Axios reported that Morgan Stanley, Bank of America, JPMorgan, and Wells Fargo had found third-party AI agents inside customer accounts, sometimes doing things users didn't ask for.
So the product isn't really the chat. The product is the connector. And Anthropic doesn't own the connector.
One commenter put it better than i can:
the interesting bet isn't the chat UI, it's whether banks keep the connector open.
Half the thread already built this
most tutorials tell you to pay for the thing. The ClaudeAI thread went the other way.
The second most upvoted comment offered to share a homegrown version that loads PDF statements and never leaves the laptop. A dozen people asked for the repo. Then someone raised the obvious problem.
"If it's not a local model," one reply said, "it kinda has to leave your computer."
That's the real tension. You can have local data or you can have a model that reads messy transaction codes. Getting both is the hard part.
Another user said he already does this with Claude Code. Drop statements in a folder once a month, let the script file them, get a dashboard. Build once, reuse forever, free.
I've done versions of this. It works for about four months, then you skip a month, then the folder is a mess.
Where it's actually better
ok but let me be fair to the thing. There's one job it does better than my spreadsheet ever did.
Catching the slow leaks. The $12 subscription that became $19. The annual renewal you forgot.
The quarterly fee that never gets questioned.
A person scanning a statement skips those. A model reading two years of rows does not.
And the follow-up question is the part that matters. My bank app tells me i spent $600 on food. It can't tell me whether that's normal for me, or which week broke the pattern.
Claude can, because it has the history and you can keep asking.
Anthropic is upfront that it can be wrong. The announcement says to ask which transactions are behind a number and to check important figures with your bank. That's a strange thing to put in a launch post, and i respect it.
A small detour
Anthropic's launch images have this hand-drawn look. Sketchy hands holding cards, arrows, little coins. It's meant to feel warm and human.
One commenter said the drawn hand icon creeps them out and they don't know why.
I looked at it again after reading that. Now it creeps me out too.
There's something about a sketched hand reaching toward a bank card that reads less like "friendly design system" and more like "someone is about to take something." Which is, if you sit with it, exactly the anxiety this whole product is fighting.
Design teams spend months on this stuff. Then one guy on Reddit wrecks it with eleven words.
And now i can't unsee it. Thanks, fuzzypetiolesguy.
Real talk
If you're outside the US, none of this exists for you yet. Beta, US only, Pro, Max, Ultra. That's the whole map.
If you have one checking account and one credit card, skip it. Your bank app already shows this month's spending, and you don't need a language model to read a number back to you.
If you keep a local-only setup, don't touch this. You already know why.
And if you're going to try it, connect one account. Not four. See what it catches in a month before you hand over two years of history from every institution you use.
The honest critique is that the trade is lopsided. You give up a lot of signal about your life, your habits, your relationships, your health if you pay medical bills from that account. What comes back is a better answer to a question your bank already answers badly.
For some people that's worth it. For most, it's a convenience upgrade with a permanent privacy cost.
The ending
The top comment in that thread is still the one i keep thinking about.
A guy says he knows he shouldn't. He's going to do it anyway. And 166 people agreed with him.
That's not a technology problem. That's a negotiation people have already lost, quietly, in public, before the feature finished rolling out.
I'll probably try it too. One card, one account, see what it finds.
But i keep coming back to another commenter. Five years ago companies paid billions for half this data. Now we hand it over and pay $20 a month for the privilege.
He was joking. Mostly.
Top comments (1)
The “what’s left before payday” example needs one extra check: does it subtract bills that have not posted yet, pending card charges, and money already reserved for annual costs? A current balance is not the same as spendable cash. I’d ask for the transaction list AND the assumptions behind that number before using it to make a spending decision. Also worth correcting the table: keeping statement files locally does not keep their contents local if Claude Code sends them to a remote model.