Earned: pay AI only for what works

Today you pay for AI before you know if it worked. Earned flips that. PayPal holds the buyer's money, independent AI inspectors check every result against a checklist the buyer agreed on, and PayPal captures only the results that passed. The rest of the hold goes straight back to the buyer, and every outcome becomes a signed, public receipt.

• Live demo: https://earned.loo2maaa.workers.dev (PayPal sandbox, real AI models)

• MCP server for any AI agent: https://earned.loo2maaa.workers.dev/mcp

• Demo video: https://www.youtube.com/watch?v=HzYOj38yoPw

Built for the PayPal AI Hackathon (Oct–Nov 2026). New project, MIT licensed.

The problem

AI is sold before anyone knows whether it worked. A buyer pays, the model hallucinates a name, garbles a date or draws four cats instead of three, and the only remedies are a refund request, a dispute, or nothing. Most results are not all bad; they are "3 of 4 good", and there is no payment primitive for that.

We run a few small AI tools ourselves. Our paying customers mostly arrive from chat assistants, on their phones, and they hesitate at the same moment: paying before they have seen the result. And when AI agents start buying from other AIs, nobody will be watching at all. The only workable rule there is verify, then settle.

Related MCP server: Growaify MCP Server

How it works

sequenceDiagram
    autonumber
    actor Buyer
    participant Earned
    participant PayPal
    participant Studio as AI studio (image model)
    participant Judges as AI inspectors
    Buyer->>Earned: "A card with exactly 3 cats that says …"
    Earned->>Earned: AI drafts a checklist (the contract), buyer edits it
    Earned->>PayPal: POST /v2/checkout/orders (intent AUTHORIZE, custom_id = contract hash)
    Buyer->>PayPal: approves the hold (an authorization, not a charge)
    PayPal-->>Earned: return URL / CHECKOUT.ORDER.APPROVED webhook
    Earned->>PayPal: POST /v2/checkout/orders/{id}/authorize
    Earned->>Studio: make 4 images
    Earned->>Judges: two vision models check every image, item by item
    Judges-->>Earned: pass/fail + quoted evidence (a third model breaks ties)
    alt some images passed
        Earned->>PayPal: POST /v2/payments/authorizations/{id}/capture (passed × price, final_capture = true)
    else nothing passed
        Earned->>PayPal: POST /v2/payments/authorizations/{id}/void
    end
    PayPal-->>Earned: PAYMENT.CAPTURE.COMPLETED webhook (signature verified)
    Earned->>Buyer: images, verdicts, money paid vs released, Ed25519-signed receipt

1. Agree the checklist. An LLM turns the buyer's words into 3–6 objectively checkable items: exact text, counts, colours, exclusions. Factual items decide payment; style wishes ("watercolour", "warm colours") are inspected and reported but never billed. The buyer can edit everything. The checklist's SHA-256 hash is written into the PayPal order's custom_id, so the goalposts cannot move later.

2. PayPal holds the money. An Orders v2 authorization: the studio knows the money is real, the buyer knows it hasn't moved.

3. Two AI inspectors check every result. Two different model families transcribe any text character by character and judge each item with evidence. If they disagree on a required item, a stronger third model decides that item alone.

4. PayPal captures only what passed. 3 of 4 passed? One partial capture for exactly 3 × $1.50 with final_capture: true, which releases the remaining $1.50. Nothing passed? The authorization is voided.

5. Appeals and receipts. A buyer can appeal a charged image once; a stronger model re-inspects it and, if the appeal is upheld, PayPal refunds that item immediately. Every settled order gets an Ed25519-signed receipt that anyone can verify.

Where this goes

• Today: outcome pricing for any AI tool. One integration turns a pay-first AI product into a pay-per-pass one.

• Next: a credit score for every AI. Benchmarks get gamed and reviews get faked; pass rates measured from settled money are much harder to fake. The Quality Index ranks AI studios by what buyers actually paid for (Wilson lower bound, so two lucky orders can't top the table), and agents read it before they buy.

• Endgame: the settlement layer of the agent economy. When agents hire agents, no human is there to complain. Whoever verifies outcomes decides where the money goes. Every payment becomes a vote: money flows to AI that works, and AI that doesn't stops getting paid.

Try it (judges)

Payments run on the PayPal sandbox, so no real money moves. Images and inspections use real AI models, paid for by us (capped).

PayPal sandbox buyer for approving holds

You can also use any sandbox personal account of your own.

1. Web: open /studio, pick an example, click Draft my checklist, edit it, click Hold with PayPal, log in with the sandbox buyer, approve, and watch the order page: images appear, then verdicts with evidence, then the PayPal settlement. Open the signed receipt and see it verified in your browser.

2. Agent: open /agent and say what you want. The agent ranks studios with the Quality Index, drafts the checklist, asks you to confirm and opens the order. Only you can approve the PayPal hold.

3. MCP: add https://earned.loo2maaa.workers.dev/mcp to any MCP client (Streamable HTTP, no auth) and ask it to buy a card. Or from a terminal:

curl -s https://earned.loo2maaa.workers.dev/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Demo limits: 4 orders and 40 agent messages per visitor per day, and a hard cap on total AI spend. If the cap is reached, the Quality Index and all receipts stay browsable.

PayPal integration

Code: src/paypal.ts, settlement math in src/settle.ts, the order state machine in src/orders.ts.

AI tools and how they are used

All models are called through one OpenAI-compatible gateway (APIMart), so any compatible provider can be swapped in via AI_API_BASE.

The project itself was written with help from an AI coding assistant (Claude Code).

MCP tools

Run it locally

Requirements: Node 24+, a Cloudflare account (free plan works for local dev), a PayPal developer account with a sandbox REST app, and a key for an OpenAI-compatible gateway that serves the models above (or change the model ids).

git clone <this repo> earned && cd earned
npm install
cp .dev.vars.example .dev.vars        # fill in PayPal sandbox client id/secret and AI_API_KEY
node scripts/gen-receipt-key.mjs      # paste the output into RECEIPT_SIGNING_JWK in .dev.vars
npx wrangler d1 migrations apply earned --local
npm run dev                           # http://localhost:8787
npm test                              # unit tests: settlement, consensus, receipts, MCP

Webhooks need a public URL. To deploy your own copy:

npx wrangler d1 create earned         # put the id into wrangler.jsonc
npx wrangler r2 bucket create earned-media
npx wrangler d1 migrations apply earned --remote
npx wrangler deploy
PAYPAL_CLIENT_ID=… PAYPAL_CLIENT_SECRET=… node scripts/register-webhook.mjs https://<your-worker>.workers.dev
npx wrangler secret put PAYPAL_CLIENT_ID      # and PAYPAL_CLIENT_SECRET, PAYPAL_WEBHOOK_ID,
                                              # AI_API_KEY, RECEIPT_SIGNING_JWK, IP_HASH_SALT

scripts/seed-orders.mjs creates real orders against an instance so the Quality Index has data; each one still needs a human to approve its sandbox hold.

Project layout

src/
  worker.ts     routes: pages, JSON API, PayPal return + webhooks, media, cron
  orders.ts     order state machine: authorized → generating → judging → settling → settled
  paypal.ts     Orders v2, authorizations, captures, voids, refunds, webhook verification
  judge.ts      two-inspector consensus + tiebreaker
  criteria.ts   checklist drafting and the contract hash
  settle.ts     pure settlement math
  receipts.ts   Ed25519 receipt signing and verification
  quality.ts    Quality Index (Wilson lower bound) and recent receipts
  tools.ts      the five agent tools, shared by MCP and the web agent
  mcp.ts        MCP server (Streamable HTTP, JSON-RPC)
  agent.ts      the web shopping agent
  appeal.ts     buyer appeals → PayPal refunds
  budget.ts     AI spend ledger and per-visitor limits for the public demo
public/         static pages (no framework)
migrations/     D1 schema
test/           node:test unit tests

Stack: Cloudflare Workers, D1, R2, TypeScript, no frontend framework.

Honest limits

• Sandbox only. In production each studio would be its own PayPal merchant and Earned would use PayPal's multiparty (platform) model, so funds go straight to the studio and the platform never holds money. The demo uses one sandbox merchant for all four studios.

• Authorizations expire. PayPal holds are honoured for 3 days and valid for up to 29; longer jobs would be split into milestones, one hold each.

• AI inspectors can be wrong. That is why there are two of them, a tiebreaker, quoted evidence on every verdict, an appeal path that ends in a real refund, and a rule that only factual items move money.

• The Quality Index is small. It only counts orders settled on this demo, seeded through real sandbox orders and whatever judges add.

License

MIT, see LICENSE.