AP2 & MPP — “Ask-Permission-First” Payments vs. “Run-a-Tab” Payments (3/5)
If last part’s x402 was a vending machine, these two approaches are closer to a “spending allowance” and a “bar tab.”
Recap: who’s holding the wallet?
With x402 last time, the AI held its own wallet and paid right there on the spot — like dropping a coin in a machine. Free and fast, but also a little nerve-racking, right? What if the AI spends my money however it likes?
That worry is exactly what today’s two approaches address. The key difference between them comes down to one thing: who controls the money, and how.
AP2 = the allowance limit your parent set
The first, made by Google, is AP2. Think back to your childhood allowance and it clicks.
A parent says: “You can use this card for up to $10 a day, and only at the convenience store.”
Within those bounds the kid spends freely — but can’t cross the line. Go over $10 or try to pay somewhere that isn’t the convenience store, and the card is blocked.
That’s AP2. A person signs off in advance on a permission (the technical term is a mandate) that says “you may spend up to here,” and the AI only pays within those bounds. The moment it tries to go further, it’s stopped.
The upside is peace of mind. You can later check exactly what the AI spent and where, and for a company, it’s easy for the finance team to audit. That’s why AP2 has more than 60 companies behind it, including Mastercard and PayPal. It’s especially strong in enterprise settings.
MPP = running a tab at a bar
The second, built by Stripe and Tempo, is MPP. This time, picture a bar.
When you sit down, you don’t pay for each drink separately. Just say “open a tab,” and your drinks pile up on the ledger, to be settled all at once when you leave.
That’s precisely how MPP works. While an AI keeps using a service bit by bit, the charges accumulate, and everything settles in one go when the session ends. Far more efficient than exchanging a payment signal every single time — especially when the AI works on one task for a while and uses a service repeatedly.
MPP has some heavyweight names attached: Stripe, Visa, and Mastercard, plus AI companies like Anthropic and OpenAI as partners.
The key point: these aren’t rivals — they’re different layers
Here’s a common misconception to clear up. It’s tempting to ask, “x402 vs. AP2 vs. MPP — which one wins?” But they’re actually not fighting each other. Each fits a different situation.
| Situation | Best fit | Analogy |
|---|---|---|
| One-off, on-the-spot payment | x402 | Vending machine |
| Delegating within a preset limit | AP2 | Spending allowance |
| Long use, settle later at once | MPP | Bar tab |
In practice, the emerging picture is one where all three get mixed and matched to fit the moment — vending-machine style for a quick data purchase, allowance style for big spending, bar-tab style for long tasks.
But there’s a big hole here
All three handle “the AI spending money” well. But what about this situation?
The AI paid — and what it received was defective. A person would demand a refund, but a blockchain payment, once sent, is hard to reverse.
Solving that headache is the subject of the next part: escrow.
This part in one line
AP2 = an “allowance” that spends within a preset limit. MPP = a “bar tab” that piles up and settles at once. Together with x402, the three split by situation.
This is Part 3 of ChainLab’s series, “Agent Payments, Explained Simply.”
