Cards your agents can spend but never widen
Every agent gets its own virtual card under a signed, revocable mandate: merchant locks, budgets that fail closed, expiry, and a decision trace for every purchase the agent makes or is stopped from making.
Every public demo shows an agent succeeding. Nobody shows the agent being stopped. An issuer's job is the second one, so that is what we built.
A first-class subject with its own API credential, scoped to a programme. Not a cardholder, never a KYC subject, never holding money. One agent can serve many principals; one principal can run many agents.
The signed, immutable, revocable authority: principal × agent × constraints. Merchant allow-list, lock-on-first-use, MCC and country rules, per-transaction cap, total budget, validity window. Its grammar mirrors Google's AP2 payment mandate, so the adapters are field mappings.
The agent mints its own virtual card under the mandate with one call, single-use or n-use. It cannot widen the envelope, and it never sees a PAN, only a card reference.
Every authorization is checked against the mandate on the hot path, inside the authorization budget. The budget is exact and fails closed. Every decline carries the reason and the mandate clause that produced it.
The principal, or their finance team, creates the mandate. The agent authenticates with its own credential, mints a card under the mandate and pays. The platform does the rest, at authorization time, in milliseconds.
POST /v1/mandates { "principal_id": "ch_8k2m", "agent_id": "agent_procure_01", "constraints": { "merchant_allow": ["aws", "vercel", "openai"], "merchant_lock": "first_use", "mcc_allow": ["5734", "7372"], "per_transaction_max": { "amount": "500.00", "currency": "EUR" }, "total_budget": { "amount": "4000.00", "currency": "EUR" }, "not_after": "2026-10-31T23:59:59Z", "max_authorizations": 40 } } // 201 · signed (JWS) · revocable · enforced at authorization POST /v1/mandates/mnd_4c9a/cards // called by the agent itself // 201 · card_7q2e · single_use · PAN never returned
Merchant allow-lists matched on acceptor ID, never on a name an agent could spoof. Lock on first use for the agents that don't know the merchant in advance.
Cumulative spend is derived from the authorization rows themselves, so a budget of €4,000 is €4,000, not approximately €4,000. Exceed it and the authorization fails closed.
A token and a cryptogram are not proof of consent under UK PSRs 2017 reg 75. A signed mandate is. It is portable evidence any third party can verify, kept with the decision trace and the scheme's agentic indicator.
Revoke a mandate, or disable an agent and every mandate it holds, atomically. Live usage per mandate through the API and the console.
Built alongside Visa Intelligent Commerce, Mastercard Agent Pay and the EMVCo agentic framework: agent identity signals decoded and retained, passkey authentication for the principal, tokenization to wallets.
Agent traffic is tagged, so behavioural baselines don't misfire on a bot that buys at 3am from three merchants. Correlated-fraud signals across agents feed the same in-line score.
An agent that buys cloud capacity, API credit and tooling for a company, from a fixed list of vendors, inside a monthly budget. The finance team sets the mandate; the agent never asks for a card number.
Flights and hotels bought under a per-trip mandate: MCC-limited, capped, expiring the day after return, with a single-use card per booking.
A consumer's assistant reorders household goods or grabs a deal inside a weekly allowance, with the merchant locked on first use and every purchase visible in the banking app.
Vehicles paying for charging, devices buying data, workloads buying compute: programmatic principals, programmatic limits, and a ledger entry for each.
Each of these runs on a disposable virtual card: single-use or n-use, merchant-locked, time-boxed, minted at the moment of purchase and closed the moment it settles.
Commerce where an AI agent selects, negotiates and pays on behalf of a person or a business. The payment still runs on card rails; what changes is that the credential is bound to an agent and the authority to spend is captured explicitly, in a mandate, rather than assumed from possession of a card.
The controls are the same primitives. The difference is the mandate: a signed, immutable record of who authorised which agent to do what, that the agent cannot widen, that can be revoked in one call, and that survives as evidence when a purchase is questioned.
No. The agent mints a card under its mandate and receives a card reference. The PAN stays inside the platform's PCI scope and is delivered to the payment surface through tokenization.
Card authorization has a budget of a few hundred milliseconds, so the human sets the policy ahead of time and the platform enforces it at authorization. Anything outside the policy is declined with a reason the person can read, and the mandate can be widened for next time.
Any agent that can call a REST API. Mandate constraints follow the AP2 payment mandate grammar, and agents driven through MCP, OpenAI's Agentic Commerce Protocol or a custom runtime integrate the same way: authenticate as the agent, mint a card under a mandate, spend.
The same platform, used differently by different teams. Each card names the segment, the outcome, and the products that carry it.
Every agent gets its own virtual card under a signed, revocable mandate: merchant locks, budgets that fail closed, expiry, and a decision trace for every purchase the agent makes or is stopped from making.
Bring an agent and one thing it should never be allowed to buy. We show you the mandate that stops it, live, in the call.