Skip to content
IntentReceipt

Receipts for what the AI was asked to do.

When an AI agent buys something for you, nobody can prove what you asked for, what the agent understood or what you confirmed. IntentReceipt records all three, signs them into a tamper-evident ledger linked to the PayPal order, and uses AI to resolve disputes with evidence that is checked against the record, fairly for buyer and seller.

PayPal sandbox:
connected
AI (groq):
openai/gpt-oss-120b
Intent receiptChain intact
seq 1 → 2 → 3 · Ed25519 · ir-key-1Open

How it works

  1. 1

    Your request is pinned down before you pay

    The agent turns your words into requirements, and code, not the model, checks each one on the confirmation card. The exact card you confirm is hashed, signed (Ed25519) and chained, so nobody can quietly change it later.

  2. 2

    Tied to the PayPal order

    PayPal Orders v2 creates the order only after a signed confirmation exists, for exactly that item and amount. The order carries the receipt id (in custom_id), and the capture joins the same chain.

  3. 3

    Disputes get evidence you can check

    When a dispute arrives, an AI analyst reads the receipt, order and shipping history. Each citation points at a field in the record, and code rejects any that do not match. If the agent was at fault, it recommends a refund. A human approves every action.

Who it is for

  • Merchants accepting agent-initiated orders: contest the disputes you should win, refund the ones you should not fight, and stop over-blocking agent traffic.
  • Dispute and risk teams: structured, verifiable evidence about intent instead of screenshots.
  • Agent developers: a receipt format and a signal when the agent, not the buyer, got it wrong.

Honest scope

  • PayPal sandbox only. No real money, no production credentials.
  • Seeded demo orders and demo disputes are labelled SIMULATED; precomputed analyses are labelled CACHED.
  • The ledger is tamper-evident (hash chain + signature), not a blockchain. Inspired by agent payment mandates; not an AP2 implementation.
  • The analyst recommends; it does not adjudicate liability. A human approves every action.