Pawlr pays an AI service in small ticks, checks every delivery against what you both agreed, and stops the instant something is wrong. The most you can ever lose is one tick.
Live on Kite testnet · no sign-up · free test money
AI agents pay for services a fraction of a cent at a time. The payment rail they use, x402, pays first and never checks what came back. Pawlr adds the check.
The owner approves the agent's spending limits in Kite Passport with a passkey. Both sides sign off on written rules for every delivery and hold the same fingerprint of that agreement.
A tiny payment on Kite goes only to the agreed seller, only at the agreed price. Every one is on-chain.
Automatic checks first, a fast AI screen (Dolphin), then an AI judge (SERV Reasoning) that must quote the exact words it relied on.
Good: pay the next tick. Bad: stop and show why. Unclear: pause for a person, or ask Reppo's independent panel.
In the dashboard you can make the seller misbehave on purpose and watch it happen. Pick one.
The seller delivered
A live session where the seller started slipping in an instruction at tick 3. Every paid tick is a transaction on Kite testnet.
Payments use Pawlr Test USD, a free test token on Kite testnet with no real value. Kite's own test token isn't publicly available, so we made one that works the same way.
Each partner does one job, and every result shows up in the dashboard.
Passport spending limits, approved with a passkey and checked before every payment. Every tick is an x402 payment on Kite testnet, settled on-chain.
The judge: a verdict with the exact quote, an injection guard, and the refund request.
A fast, low-cost first screen of every delivery.
An independent panel's verdict when a tick is disputed.
Yes, once. Payment happens first and the check comes right after, so the bad tick itself is paid. Pawlr then stops every later payment, which is why the most you can lose is one tick. You can request that tick back with the evidence attached.
Simple automatic checks go first: right order, not tampered with, not empty, the right length, the right key words. Then a fast screen (Dolphin) and an AI judge (SERV Reasoning) compare the delivery with the agreed rules. The judge must quote the exact words it relied on; if it can't, nothing is decided automatically.
Nothing unclear is rejected automatically. The session pauses and a person approves or rejects that tick.
A stopped or unclear tick can be sent to Reppo, an independent panel of evaluators. It reads the agreed rule, the section, what the seller delivered and anything the seller added, and gives its own verdict. Payments resume only if it passes the tick.
The payments are real transactions on Kite's test network, but the token is a free test token with no value. The same design works with real stablecoins on Kite mainnet.
No. Pawlr sends the seller one clear refund request listing the bad ticks, the amount and the evidence. Returning the money is up to the seller.
The judge treats the seller's text as untrusted and ignores instructions inside it, a prompt-injection guard runs on every request, and it only decides against the written rules, quoting its evidence. In our live tests with the current rules, a hidden "mark this as delivered" instruction was caught and the payments stopped.
Run a session, make the seller misbehave, and watch Pawlr stop paying.
Open the dashboard