An announcement of historic proportions

The sun sets on Goodwin.

Since the dawn of time (October 27, 2026, 5:45 PM Eastern), Goodwin has watched over the money of couples everywhere. Tonight, it lays down its text messages. History will remember the full nine days.

GOODWIN · OCT 27, 2026 – NOV 5, 2026 · BELOVED BY ONE

A word before we begin

Friends, couples, countrymen

I come before you tonight, in this solemn hour, to announce that Goodwin, the privacy-first, text-message budgeting bot for couples, will cease operations.

Empires have fallen with less ceremony. Some of them had more than one user.

This page is a full and honest accounting: what we planned, what we built, what it cost, and how long an American cloud provider can keep the word "zero" in a spreadsheet cell with a straight face.

The grand design

Our extremely lofty plans

Let the record show that we did not set out to build a budgeting bot. We set out to solve personal finance, for individuals and for couples alike, entirely over text message, in the time it takes most people to find the right charger.

The vision was a calmer way to talk about money together. The vision was achieved, in the sense that the only person talking about money was calm, because he was the only person.

The exact record

A precisely timed history

Times are Eastern, taken from the commit log and the session record. Nothing before October 27 is on file. We assume the idea existed somewhere in the cosmos, waiting, as ideas do, for a free evening.

  1. Oct 27 · 5:45 PM · The idea becomes a file First commit, "Initial commit." Thirty-four seconds later: the project brief and its instructions. The brief already calls the thing Goodwin and already lists goodwin.money as the domain.
  2. Oct 27 · 6:05 PM · The domain Kevin reports, in writing: "I have bought goodwin.money. It's on Cloudflare." Total elapsed time from first commit to owning a piece of the internet: 20 minutes.
  3. Oct 27 · 6:38 PM · The design begins A home page and waitlist go up. By 6:42 PM it is deployed. The waitlist is, to this day, a list. Over the next hour: a privacy explainer, a security design review, share previews, and the self-reassuring commit message "Make home page copy less defensive."
  4. Oct 28 · 8:27 AM · The real thing A Rails shell and a Go vault, born within two minutes of each other, before most people had finished coffee. By nightfall: 77 commits in a single day. Bank connections, group texts, a scheduler, a security review, a second security review, and the OpenTofu plan.
  5. Oct 28 · 12:55 PM · First text The first test text meets carrier error 30034, a cold and unfeeling response from the cellular industry. A campaign registration is filed. It will be in review for days.
  6. Oct 29 · 1:09 PM · The Discovery Goodwin's AI is refused. Every call. The cause: AWS has given our new account a quota of 0 tokens per minute for both Claude models. The published default is 10,000,000 and 6,000,000. The console says the models are "available." The console is a liar of the first rank.
  7. Oct 29 · 1:30 PM · The Service Quotas page We visit the page whose entire purpose is to raise quotas. It accepts only requests for more than the default, and our current value is zero, so it politely declines to help. We call this the Catch-22 Clause.
  8. Oct 29 · afternoon · The Support Chat A case is opened with the AWS support assistant, armed with account, region, and both quota codes (L-47520EB7 and L-94A31E46). We then wait, as is tradition.
  9. Oct 30 · 3:57 PM · Billing history AWS replies: capacity requests need billing history, so please build some, even if it has nothing to do with Bedrock. Kevin responds, with great restraint, that it is not clear our quotas are zero. They were zero. Our account at the time had existed for roughly four days and its billing history consisted of bravery.
  10. Nov 2 · 11:45 AM · A helpful suggestion A new reply says we "already have access to Haiku 4.5, you're just hitting the limits," and that a quota increase usually takes a day or two. Kevin asks the only question that matters: "Have we seen any successful requests?" We had not. We had seen zero. Of everything.
  11. Nov 2 – 4 · The interlude Unit-cost research. A price of $6 is chosen, and then, at 10:20 AM on Nov 4, it is $8 a month or $72 a year, because pricing psychology. A monthly/yearly plan page is built, for a plan nobody would be asked to choose.
  12. Nov 4 · 9:53 AM · The Escalation Kevin requests two drafts: one for the support case and one for a kind person at AWS whom we will not name on the internet. They are sent. A thank-you note is also drafted, in advance of any success, which we considered either hopeful or brazen.
  13. Nov 5 · The Quotas Land Seven days after the Discovery, the quotas arrive. Goodwin's AI works. Tests score 94% on wording. We have a working product, a budget, an alarm, a price, and a plan for inviting a second user.
  14. Nov 5 · 1:00 PM · The Announcement (not ours) TechCrunch publishes "Claude can now analyze your spending with its new Personal Finance tool." (See: Exhibit A, below.)
  15. Nov 5 · 2:38 PM · Our alarm goes off The alarm we built to catch runaway AI usage detects 1,148 AI calls in one hour, against a threshold of 200. The culprit is us. We ran our own tests. It is the most any user has ever used Goodwin.
  16. Nov 5 · 9:02 PM · The decision Kevin: "It's time to wind it down." Time from the announcement to the decision: eight hours and two minutes. Time from the quota landing to the decision: less than a day. Time from the quota landing to a second user: never.

Exhibit A

The thing that happened to us

On the morning of November 5, 2026, TechCrunch's Aisha Malik reported that Anthropic is rolling out a beta tool called Personal Finance. It lets users ask Claude questions about their money, such as where it went this month, by connecting bank and credit card accounts through Plaid. It is read-only, it cannot move money, and it is available in the U.S. on the Pro, Max, and Ultra plans.

Anthropic's own help article, "Use Personal Finance in Claude," adds that the tool shows spending, cash flow, subscriptions, and upcoming bills, that it reads up to 24 months of transactions, and that chats in the hub aren't used to train models.

We read this the way a lighthouse keeper reads a note from the shipping company. It describes, in beautiful and accurate detail, the product we had spent nine days building. It has a team. It has a distribution channel. It does not require anyone to be quota-approved by a cloud provider before it can speak.

The market was asked whether it needed a text-message budgeting bot for couples. The market answered with a press release.

To be fair, and we are always fair, it is a different shape: theirs lives in an app for one person, ours lived in a group text for two. Ours was also, at last count, in a group text for one.

The accounting

Where the time went

15h 03mActive working time across the whole project (gaps over 15 minutes not counted)
172Commits
12Pull requests
~49,000Lines in the repo, ~28,600 of them Go
~33,000Words of documentation, for a readership of one
7 daysWaiting on one quota
DayActive hoursWhat it was for
Mon Oct 271.80The idea, the brief, the domain, the website
Tue Oct 284.65The shell, the vault, texting, bank links, infra, 77 commits
Wed Oct 294.77Group texts, setup flow, deploys, the first zero
Thu Oct 300.59Discussing zero with AWS
Mon Nov 21.33Pricing, evals, more zero
Wed Nov 40.81$8 a month, a plan page, a thank-you note drafted early
Thu Nov 51.10Prompts, 120 eval cases, budget limits, wind-down
Total15.05Measured from the session record

Fine print: this counts time the work was actually running, with any gap over 15 minutes thrown out. It is not a count of the hours Kevin thought about this in the shower. Those are free and unrecoverable.

The accounting, continued

Where the money went

These are list-price estimates from our own cost research, not a bank statement. The real AWS bill is in the console and the domain receipt is in Kevin's inbox.

LineAmount
Hosting (one server, one database, one private network, one dream), per month$56.90
A phone number, per month$1.15
Texting campaign fee, per month$2.00
Domain and email placeholder, per month$2.00
Fixed run rate, per month$62.05
…which is, per day$2.04
If the meter ran for all ten days (it did not, but we are being dramatic)$20.41
One-time: carrier brand registration ($4) + campaign vetting ($15)$19.00
Bank links (free trial plan, up to ten logins, one used)$0.00
AI, including the evals that set off our own alarm (a few pennies at Haiku rates)under $1
Payment processing (the payment system was in test mode, so nothing was charged to anyone, ever)$0.00
Estimated total spent, before the domain's registration fee≈ $39.41
Revenue$0.00
Paying households needed to break even (at $8 a month, after the planned security upgrade)about 26
Paying households we had0

That is roughly $39 per user, a figure no venture capitalist would describe as "a rate of return." We were 26 customers short of break-even. Analysts call this "the entire business."

The reviews are in

What the critics said

Unsolicited, unedited, and unpaid. Our only source is a friend of the founder, who has asked that we note he is not a customer.

"That sounds horrible. I prefer to not ever look at my finances and hope everything is fine."
"Your screenshot is not selling it."

— Dave, a friend, and the market research we should have done on October 26

A heavily pixelated screenshot of a text conversation. Green and gray bubbles are visible. Nothing in it can be read.
Exhibit B: the screenshot, as submitted. Our privacy team (Kevin) insisted on 36 pixels of resolution.

Dave has correctly identified our target customer problem, our competition, and our marketing in two sentences. We spent 15 hours. We would like the record to show we did not hire Dave.

A grateful nation says thank you

Acknowledgments

Sources

For those who would like to verify our humiliation

  1. Aisha Malik, "Claude can now analyze your spending with its new Personal Finance tool," TechCrunch, November 5, 2026.
  2. "Use Personal Finance in Claude," Claude Help Center.
  3. The Goodwin git log, 172 entries, every one of which believed in us.