Inspiration
There's an old folk tale. A traveller walks into a hungry village with nothing but a pot and a stone. He sets the stone boiling, and one by one the villagers add what they have: a carrot, an onion, a handful of barley. Soon there's soup for everyone. Nobody gave money. Everyone gave something, and everyone could see the soup.
Today we give money, and it disappears. Roughly 4 in 5 first-time donors never give again (Fundraising Effectiveness Project, 2025). The top reason lapsed donors give is not knowing what their money did. 68% of people say trust in a charity is essential, but only 18% highly trust charities (BBB Give.org, 2026). Small charities, most of them with budgets under \$500K, have no staff to prove their spending. Meanwhile, AI makes fake charities, fake photos and fake reports cheap to produce.
Every giving platform answers this the same way: send cash, then try to prove where it went. I wanted to try the opposite: don't send cash, send the blanket.
What it does
Stone Soup turns donations into verified purchases. A donor doesn't fund a cause. They fund a specific item a small charity needs, and PayPal pays the supplier directly.
- Charities post exactly what they need. They write it in plain language ("rice for 30 people for two weeks") or snap a photo of a bare shelf. Gemini turns that into priced needs and matches each one to a supplier's catalog.
- Every price is checked against the market. Channel3 finds about 10 real retail products for each item, and Gemini filters out the irrelevant ones. The comparison is saved so donors can see it. A price above the market needs a written reason.
- The donor gives with one PayPal approval. A single PayPal Orders v2 checkout has one purchase unit per supplier, and each unit's
payeeis that supplier's own account. Stone Soup never holds the money. - Or the donor asks the Giving Agent. "I have \$20, help people stay warm." Gemini searches the open needs with function calls and proposes a basket. The server prices and checks every proposal, and the donor approves once in PayPal.
- Delivery is proven with a photo. Charity staff upload a picture of what arrived. Gemini vision checks it's the right item in roughly the right quantity. Every photo is fingerprinted, so a reused picture is caught. Doubtful deliveries go to a person and are never rejected automatically.
- The donor hears back with a personal note, the delivery photo, and the whole journey of their gift ("Follow the dollar"), including the PayPal capture IDs that prove it.
Around every gift:
- Verified charities only. Each charity's EIN is checked against IRS records through ProPublica's Nonprofit Explorer API. The charity also posts a code on its own website, and Gemini with Google Search confirms the site is really theirs.
- Scam Check. Paste any fundraiser link, text or screenshot, and get the red flags along with the evidence for each.
- Give from any AI. An MCP server lets any AI assistant find needs and set up a gift. Only the person, in PayPal, can move money.
- Group pots and a live wall. Friends pool gifts toward one need, and a shelter's big screen drops a stone in the pot for every gift.
- Open ledger and funder reports. These are built only from real payments and verified deliveries (AG Grid).
How we built it
- App: Next.js 16 (App Router) and React 19 in TypeScript, with Postgres 16 and the Drizzle ORM. It deploys to Render from a blueprint file.
- PayPal: Orders v2 with
payeeand multiple purchase units, the JS SDK v6 buttons, andPayPal-Request-Ididempotency so retries are safe. Signature-verified webhooks act as a backup, andpayer-actionapproval links serve gifts created by AI assistants. - AI: Google Gemini through
@google/genai. It uses structured JSON output, vision, function calling, and Google Search grounding. Calls fall back through a chain of models, with timeouts and a cooldown for overloaded ones. - Market data: Channel3 product search, re-priced by product ID so price checks can be reproduced.
- Charity records: IRS data from ProPublica's Nonprofit Explorer API.
Paying the supplier the exact price. A donor can choose to cover PayPal's fee. PayPal charges a percentage plus a fixed fee on the gross amount, so simply adding the fee on top isn't enough. To make the supplier net exactly $N$:
$$ G = \left\lceil \frac{N + f}{1 - r} \right\rceil, \qquad r = 3.49\%,\ f = \$0.49 $$
For two \$18 blankets, $N = \$36.00$, the donor pays $G = \$37.81$, and the supplier receives exactly \$36.00.
Catching reused photos. Each delivery photo gets an exact SHA-256 hash and a 64-bit perceptual difference hash. Two photos count as the same picture when their perceptual hashes are close:
$$ d_H(h_1, h_2) = \sum_{i=1}^{64} \left[\, h_1^{(i)} \neq h_2^{(i)} \,\right] \le 10 $$
That rule still catches a photo after it has been re-saved, resized or lightly cropped.
Challenges we ran into
- The AI must never decide amounts. A language model choosing how much money moves is a risk. The Giving Agent only picks which needs and how many units. The server reads prices from the database, enforces the budget, and builds the order itself. Text written by charities is passed to the model as data, never as instructions.
- Paying suppliers directly. PayPal's Agent Toolkit can't set a
payee, so I built that part on Orders v2 directly. On day one I proved in sandbox that one checkout can pay several different business accounts. - Fees on small gifts. PayPal's fixed fee applies to each payee, so a \$1 share to a charity would mostly go to fees. I set a minimum share and made fee cover exact.
- Market data is uneven. Channel3 is strong for retail goods and weak for groceries. Search results needed a relevance filter, and items with no reliable data are labelled honestly instead of getting a made-up benchmark.
- Fake charity pages. In 2026, US states complained about donation pages created for charities without their consent. So Stone Soup never creates a charity page by itself. It needs an IRS record, a code on the charity's own website and a web check, and a human reviewer steps in when those aren't conclusive.
- Overloaded models during a demo. Model fallbacks, timeouts and per-IP rate limits keep the app working when one model is busy.
Accomplishments that we're proud of
- The core idea works end to end in the PayPal sandbox. Money goes straight from the donor to the supplier, with no middleman account.
- One \$18 blanket can be followed all the way: Giving Agent, PayPal approval, supplier paid, delivery photo checked by Gemini, donor thanked, and every PayPal ID on record.
- Trust checks run at every step: verified charities, market-checked prices, fingerprinted photos, and a human review queue for anything doubtful.
- It works from any AI assistant without the assistant ever being able to move money.
- I built all of it solo, including the demo film.
What we learned
- Trust is a product feature, not a marketing line. The strongest proof is structural. If the platform never holds the cash, there's no cash to misspend.
- Put the AI where judgement helps and keep it out of where judgement is dangerous. Gemini is excellent at turning messy input (a shelf photo, a paragraph, a fundraiser screenshot) into structure. Arithmetic, prices and money movement belong in plain code.
- PayPal's Orders API is more flexible than it looks. With
payeeand multiple purchase units, it can serve as a marketplace payment layer. - Honest limits build credibility. "No market data" and "sent to a person for review" are better answers than a confident guess.
What's next for Stone-Soup
- Real suppliers and charities: onboard suppliers through PayPal Partner Referrals and pilot with a few local shelters and food banks.
- Tax-deductible gifts: route gifts through a fiscal sponsor or donor-advised fund so they're deductible.
- Trust scores for charities and suppliers, built from their delivery history.
- Real sign-in for charities, with account recovery, instead of the demo's browser-bound access.
- Soup Club monthly giving with PayPal Vault, always with the donor's explicit consent.
- More catalogs and countries. Every place has people who need something specific, and someone nearby who can supply it.
Built With
- channel3
- nextjs
- paypal