Give agents spending authority you can revoke.
AgentPay issues scoped virtual cards, enforces your spending policy before the authorization is ever attempted, and puts a human in front of anything that matters.
- Card mint
- < 400ms
- Policy checks
- Pre-auth
- Interfaces
- API · MCP · CLI
AgentPay virtual card
Task card · DoorDash
•••• •••• •••• 4821
Limit
$40.00
Expires
12/30
Agent
agt_ops
A card is a policy object, not a number
Every card carries the agent that owns it, the merchant it may touch, the ceiling it may reach, and the approval route it must clear. The credential is the last thing created — and the first thing revoked.
import { AgentPay } from "@agentpay/sdk";
const agentpay = new AgentPay(process.env.AGENTPAY_SECRET_KEY);
const card = await agentpay.cards.create({
amount_cents: 4000,
currency: "USD",
card_type: "single_use",
agent_id: "agt_ops",
merchant_lock: "DoorDash",
approval_policy_id: "pol_finance_100",
metadata: { project: "Team lunch" },
});
// => { id: "card_123", status: "active", last_four: "4821" }Spending authority, scoped to the task
Limits, merchant locks, expiry windows, and approval thresholds are evaluated before an authorization is attempted — never reconciled after the fact.
Single-use virtual cards
A card that exists for one task, one merchant, and one amount — then closes itself on the first successful capture.
Human approval, in context
Route anything above a threshold to an approver in-app, by email, or in Slack, with the itemised cart and agent rationale attached.
Budgets that cascade
Per-transaction, daily, weekly, monthly, per-project, and company-wide ceilings that stack instead of competing.
Risk built for agents
Retry storms, duplicate carts, merchant mismatch, and threshold-shaving freeze a card before the loss, not after the statement.
Policy without code
Compose when/then rules across agent, merchant, category, amount, geography, and risk score in a visual builder.
Funding you control
Fund from a card, bank account, or prepaid wallet, and reserve balance against every outstanding card commitment.
From prompt to confirmed order, card issued mid-flight
The agent gathers requirements, builds a cart, prices the all-in total, clears policy, collects approval when required — and only then receives a credential.
- 01
Understand the ask
“Order sushi from my usual place for under $40.” The agent resolves address, preferences, and constraints before touching money.
- 02
Build and price the cart
Items, taxes, delivery, and service fees collapse into one all-in total — no surprise capture above the quoted amount.
- 03
Evaluate policy
Budgets, merchant locks, velocity, and risk signals run in-line. Above threshold, an approver is paged with the full cart.
- 04
Issue and check out
A merchant-locked, single-use card is minted for the exact amount, used once, and closed on capture.
$ npx agentpay setup
✔ Logged in as maya@northwind.co
✔ Organization "Northwind" created (test mode)
✔ API credentials issued — sk_test_••••4f2a
✔ MCP server registered for Claude Desktop
$ agentpay cards create --amount 40 --merchant DoorDash
card_9fJ2 active $40.00 •••• 4821Connect an agent the way it already works
A REST API for your backend, an MCP server for Claude, ChatGPT, and Cursor, and a CLI for everything you would rather do in a terminal.
Built for agent behaviour
Risk scoring that understands machines: retry loops, duplicate carts, and purchases engineered to land just under an approval threshold.
Credentials never sit in logs
Reveal requires reauthentication and returns a short-lived token. Full PANs and CVVs are never stored in the application database.
Test mode by default
Simulated cards, balances, and merchant checkouts so the whole flow is provable before a single real dollar exists.
Free until your agents start spending
Developer test mode is free forever. Production plans scale with agents, policies, and transaction volume.
Give your agents a card, not your card
Spin up an organization in test mode, issue a simulated card, and connect an agent over MCP in under five minutes.