Payments your agent can't make without authority.
AI agents can move money now — and a payment tool doesn't ask whether the agent was allowed to spend. The Agent Payment SDK gates every payment behind a resolved admissibility verdict and a signed Evidence Pack: ceiling, scope and purpose are enforced before execution, and no payment clears without proof it was permitted. Strong Customer Authentication aware.
The problem
A payment tool will move money the instant an agent asks.
The tool executes; it never asks whether this agent, for this purpose, up to this amount, was actually authorised. That gap turns a helpful agent into an unbounded spender the moment a prompt goes wrong.
Unbounded spend
An agent with a payment tool can clear a transfer against a ceiling no one set. One reasoning slip and the limit is whatever the API allows, not what you intended.
Prompt-injected transfers
"Pay this invoice" arrives inside poisoned context and executes. Nothing between the model and the money asks whether the instruction was legitimate.
Undefendable disputes
When a payment is challenged, you can't show the authority that permitted it — so a chargeback or a regulator's question becomes a manual reconstruction.
The solution
Every payment passes a verdict before it moves.
The SDK sits between your agent and your payment rail. Each request is resolved against delegated authority — ceiling, scope, purpose, step-up — and only an admissible request is released, sealed with evidence anyone can verify.
Admissibility verdict first
Ceiling, scope and purpose are checked at the payment boundary. Over-ceiling, off-scope or unpurposed requests are refused — fail-closed, never "probably fine".
Signed Evidence Pack
Every executed payment emits a replay-provable Evidence Pack — the request, the verdict, and the authority behind it, verifiable from published keys alone.
Strong Customer Authentication aware
When a payment needs step-up, the SDK surfaces the SCA requirement instead of silently proceeding — so high-risk moves get a human, not a guess.
Why it pays off
Let agents pay — without betting the company on a prompt.
The return is control you can prove: fewer disputes, faster reconciliation, and an authority boundary between your model and your money.
For engineering
One gate in front of your existing payment provider. You keep your rails and your ledger — you gain a hard ceiling and scope check the model can't talk its way past. Fail-closed means a misconfigured limit blocks a spend instead of clearing one.
For finance & risk
Every agent payment carries its own defence: the verdict, the delegated authority, the purpose — sealed and verifiable. Disputes and audits become a lookup, and there is a real ceiling on autonomous spend.
Get started
Gate a payment in three lines.
Backed by the KYE Decision Engine™. The Evidence Pack format and verification surface are open — a counterparty can verify a payment's authority from published keys alone, with no service in the loop.
# TypeScript / JavaScript npm install @kye/agent-payment # Python pip install kye-agent-payment
import { PaymentGate } from "@kye/agent-payment"; const gate = new PaymentGate({ trustDomain: "acme" }); // Verdict resolves before a cent moves const v = await gate.authorize({ agent: "ops-bot", purpose: "vendor-invoice", amount: { value: 4200, currency: "USD" }, }); if (v.admissible) await gate.execute(v); // + Evidence Pack // else: refused with a reason — no payment, no silent spend
Let your agents pay — with a receipt a regulator would accept.
Admissibility verdict before execution, signed Evidence Pack after, SCA-aware throughout. Apache-2.0, no runtime lock-in.