Mastercard Agent Pay is a network-level program that lets AI agents charge a Mastercard credential without ever touching the underlying card number. Mastercard unveiled Agent Pay on April 29, 2025 with Microsoft, IBM, Salesforce, and Checkout.com as launch partners. Citi and US Bank cardholders entered the pilot in September 2025; the rollout completed across all US Mastercard cardholders in November 2025. The product runs on a new credential primitive called Mastercard Agentic Tokens, which bind a tokenized payment credential to a specific agent identity with per-session and per-merchant limits. This article walks through what Agent Pay is, how the Agentic Tokens primitive works, what shipped through the September 2025 pilot and the October 27, 2025 PayPal framework, and how Agent Pay compares to Stripe Link Agents and Visa Intelligent Commerce.
​

Mastercard Agent Pay is the network's program for letting AI agents transact on cardholder credentials without holding card numbers. The mechanism issues an Agentic Token bound to an agent identity and scoped per session and per merchant. Agent Pay launched April 29, 2025, piloted with Citi and US Bank in September 2025, and reached all US cardholders in November 2025.
​

The product was announced from Mastercard's April 29, 2025 unveiling, with Microsoft, IBM, Salesforce, and Checkout.com named as launch ecosystem partners. The framing was deliberately network-scale: every Mastercard cardholder, every Mastercard merchant, every Mastercard issuer would eventually get a path to agent-initiated payments without changing the underlying processor stack. That ambition matters because Mastercard processes roughly 159 billion transactions annually, and dropping a new credential layer across that surface is what separates Agent Pay from a startup-scale agent wallet.
​

Agent Pay sits at the credential and authorization layer. It does not own the agent surface (Microsoft Copilot, ChatGPT, Gemini, and dozens of others continue to compete there), and it does not own the merchant integration (acquirers and processors like Stripe, Adyen, and Checkout.com retain that relationship). What Mastercard owns is the leg between the cardholder's bank and the merchant's processor, and Agent Pay extends that leg with agent-aware tokens. A useful comparison: ChatGPT Instant Checkout uses Stripe's Shared Payment Token to scope card credentials per merchant; Agent Pay uses Mastercard's Agentic Tokens to do something similar, but issued by the network rather than the processor.
​

The product is built around three constraints that came out of two years of network-side research. First, the user must remain in control of the credential at all times — the agent never holds the card number, only a scoped token. Second, every transaction must produce a clear identity trail (which agent, on which user's behalf, with which authorization). Third, the agent's authority must expire — sessions are short, limits are explicit, and a misbehaving agent loses access without the user changing their card. The same constraints show up in the broader landscape covered in agent spend controls.
​

An Agent Pay transaction binds an Agentic Token to an agent identity and routes that token through Mastercard's existing authorization network. The user enrolls an agent through their issuing bank's app, the bank requests an Agentic Token, the agent presents the token at checkout, and Mastercard's network checks the token's session, limit, and merchant scope before authorizing.
​

The flow has five steps, documented across Mastercard's launch announcement and the September 2025 issuer pilot release. The cardholder enrolls an agent through an issuing bank that supports Agent Pay (Citi and US Bank were first; the program is expanding through 2026 globally). The bank requests an Agentic Token from Mastercard, scoped with the user's chosen parameters: maximum spend per session, allowed merchants or merchant categories, session lifetime. Mastercard mints the token and delivers it to the agent surface (Copilot, Claude, a custom agent). When the agent reaches a checkout step, it presents the token as the payment credential. The merchant's processor sends the authorization request through the normal Mastercard rail, and the network validates the token's scope before approving or declining the transaction.
​

The crucial design choice is that Agentic Tokens replace the card number in flight. The merchant's processor sees a Mastercard credential structured like any other tokenized payment, but the token resolves at the network layer to the user's underlying card plus the agent identity that authorized the charge. From the merchant's perspective there is no integration burden; the processor handles a Mastercard token. From the issuer's perspective there is one new validation step: confirm that the token's scope matches what the user authorized at enrollment time.
​

Where Agent Pay differs from a traditional virtual card or single-use token is the identity binding. The Agentic Token is not just scoped to an amount — it is scoped to a specific agent. If the same user has separate agents in Copilot and in a custom shopping agent, each gets its own token. If the user revokes an agent's access in their bank's app, only that agent's tokens become invalid; the others continue working. This is what makes Agent Pay distinct from older programmable-card products like Lithic or Stripe Issuing, which scope credentials to a card or a session but not to a software agent identity.
​

The dispute story matters too. Because the agent identity is encoded in the token, charge disputes can be attributed to a specific agent rather than just to the user. Mastercard's network can flag patterns of agent-initiated fraud (e.g., an agent that consistently triggers chargebacks on a specific merchant) without forcing the issuer to deactivate the cardholder's underlying account. That attribution is the foundation of what Mastercard called its Trust Center for Agentic AI in the September 2025 investor release.
​

Mastercard Agentic Tokens extend the existing Mastercard Digital Enablement Service (MDES) tokenization standard with two new fields: an agent identifier and a session-scope object containing limits, merchant categories, and an expiry. The format remains compatible with Mastercard's normal authorization rails, so merchants and processors accept Agentic Tokens without integration changes.
​

Tokenization at Mastercard is not new. The network has been replacing primary account numbers (PANs) with tokenized representations across mobile wallets, recurring billing, and click-to-pay flows for over a decade through MDES. Agentic Tokens are the next iteration of that infrastructure, retrofitted with agent-identity awareness. The April 2025 unveiling positioned this as the first network-scale credential primitive built specifically for agent traffic, which is the part of the announcement that drew the most attention from issuers and acquirers.
​

The two new fields matter for very different reasons. The agent identifier is what makes attribution possible — every transaction produces a record of which software agent initiated it, in addition to which cardholder authorized the agent. That identifier is what makes the dispute story work. The session-scope object is what makes the user-control story work — a session can be limited to $200, valid for two hours, restricted to apparel merchants, and terminated by the user from their bank's app. Old single-use tokens did limits; they did not do per-agent limits.
​

From a developer's perspective, presenting an Agentic Token at checkout looks identical to presenting a normal Mastercard tokenized credential. The token resolves at the network layer; the merchant's processor sees a Mastercard authorization request and treats it like any other. This is why Mastercard could launch Agent Pay across its entire network without forcing a merchant migration. The processor relationships and the merchant integration patterns stayed the same; the network swapped in a more capable token format underneath.
​

The tokenization choice also explains why Agent Pay does not require a new agent surface. Mastercard does not ship an agent of its own. Instead, the credential is portable across whatever agent surface the user prefers — Microsoft was named at launch as a partner (with the Copilot agent line maturing through 2025-2026), and the program has expanded through partnerships with Salesforce's agent platform, IBM's agentic AI offering, and Checkout.com's processor surface. The same logic underpins how stablecoin-native agents identify themselves on different rails, covered in Stripe Link Agents.
​

Mastercard piloted Agent Pay with Citi and US Bank cardholders in September 2025 before completing the full US-cardholder rollout in November 2025. Microsoft was named as a launch partner; the Copilot agent surface evolved through 2025-2026, with the Copilot Checkout consumer experience launching January 2026 (PayPal/Microsoft) and active integration with Mastercard Agent Pay still in development as of April 2026. PayPal signed an Agent Pay Acceptance Framework on October 27, 2025. Mastercard joined Google's Universal Commerce Protocol coalition and expanded Start Path on January 20, 2026 to include a dedicated agentic commerce track.
​

The September 2025 pilot is documented in Mastercard's Q3 2025 investor release, which named Citi and US Bank as the first two issuers to give their cardholders access. The pilot ran for roughly two months. Issuer enrollment was the gating step: cardholders had to opt in through their bank's app, link a supported agent surface, and accept the new agent-permission terms. The same release announced expanded collaborations with Microsoft, Salesforce, and IBM, plus a dedicated Trust Center for Agentic AI to coordinate fraud signals and consent attestations across the rail.
​

The full US rollout completed in November 2025. By that point, every US Mastercard cardholder whose issuing bank supported Agent Pay could enroll an agent and start transacting. The expansion mechanism was issuer-by-issuer rather than network-wide flip — Mastercard activated the network capability, and individual banks turned the feature on for their cardholders as their app integrations shipped. Citi and US Bank stayed ahead through Q4 2025 and into 2026; other large US issuers followed at varying paces.
​

The October 27, 2025 PayPal × Mastercard Agent Pay Acceptance Framework was the most strategic partnership of the year. PayPal agreed to accept Mastercard Agent Pay transactions for PayPal users, with both networks committing to interoperability with major agentic commerce protocols. The framework commits both networks to compatibility with leading agentic commerce protocols. PayPal is a confirmed AP2 launch partner separately, which gives the Mastercard framework a clean path to compose with Google's mandate-signing layer.
​

Microsoft was named as a launch partner for Agent Pay in April 2025, and the Copilot agent surface evolved through 2025-2026 as Microsoft's broader Copilot product line gained checkout capabilities. Copilot Checkout, the consumer-facing checkout experience, launched in January 2026 via the PayPal/Microsoft partnership; active integration of Mastercard Agent Pay with the Copilot Checkout surface is in development as of April 2026. Salesforce's agent platform and IBM's agentic AI products carry similar Agent Pay integrations, though with smaller initial reach.
​

Mastercard also expanded Start Path on January 20, 2026 to include a dedicated agentic commerce track, bringing AI-native payment startups onto the network earlier in their lifecycle. The track was framed as a way to seed merchant-side innovation around the Agentic Tokens credential, with mentorship and integration support from Mastercard's developer team.
​

The protocol-coalition story matters too. Mastercard joined Google's Universal Commerce Protocol coalition alongside Visa, PayPal, Stripe, and American Express. UCP is the discovery-and-checkout layer agents use to interact with merchants, and the AP2 protocol it composes with handles the cryptographic-mandate authorization that Mastercard's Agentic Tokens slot into. The combined effect: Agent Pay has a place in the protocol stack regardless of which agent surface a cardholder ends up using.
​

Three large players each ship a distinct agentic commerce primitive in 2025-2026. Mastercard Agent Pay is network-issued tokens with agent-identity binding. Stripe Link Agents is a consumer wallet that delegates spend through real-time approvals and one-time-use cards. Visa Intelligent Commerce is a competing network program centered on the Trusted Agent Protocol and the Agentic Ready issuer pilot.
​

The three programs share a goal — let an AI agent charge a card without holding card credentials — but they sit at different layers of the agentic commerce stack. Agent Pay and Visa Intelligent Commerce are network-layer programs; Stripe Link Agents is a consumer-wallet product that runs on top of card networks. The table below summarizes the comparison along the dimensions buyers care about.

Reading the table: Mastercard Agent Pay's strength is breadth. It rides the existing Mastercard rail, so any merchant whose processor accepts Mastercard credentials accepts Agent Pay tokens with no integration work. The trade-off is that the user has to use a Mastercard issued by a bank that has shipped its Agent Pay onboarding flow — narrower than the Mastercard footprint at large during the rollout window.
​

Stripe Link Agents' strength is consumer surface. Link has more than 250 million users globally per Stripe Sessions 2026, and the wallet can present any card the user has saved, not just Mastercards. Real-time approvals through the Link app give users per-transaction control that Mastercard's session-scope model does not match. The trade-off is that Link Agents is a wallet, so it requires the consumer to use Link rather than enrolling through their issuing bank.
​

Visa Intelligent Commerce solves a similar problem on the Visa side, with a different identity model. Visa's Trusted Agent Protocol signs the agent's identity into HTTP request headers, and merchants verify the signature against Visa's directory. That is a complementary capability to a tokenized credential — Visa's identity layer can sit alongside any card network's payment layer, which is why the protocol composes rather than competes with ACP and AP2.
​

For most issuers and most large merchants, the practical answer in 2026 is "support all three." A Mastercard cardholder might checkout via Copilot using an Agentic Token; a Visa cardholder using a different agent surface might present a TAP-signed request that resolves through the merchant's normal Visa processor; a consumer using Stripe Link might delegate per-transaction approvals through the Link app. The protocol layer (UCP, ACP, AP2) is what makes those three flows present a uniform interface to the merchant, even though the credential leg differs across each. That convergence shows up in the broader agent payment protocol comparison.
​

Without an agent identifier, the network cannot distinguish an agent-driven charge from a user-driven one, cannot attribute disputes to a specific agent, and cannot revoke one agent without revoking the card. Embedding agent identity in the Agentic Token makes per-agent revocation, fraud attribution, and consent provenance work at the network layer.
​

This question matters because the older virtual-card pattern (Lithic, Stripe Issuing, Marqeta) already lets a user create a single-use card with a spend limit. What it does not do is record which software agent initiated each charge. If a user has three agents bound to one card and one of them starts triggering chargebacks, the issuer cannot revoke that agent's access without revoking the card. With Agentic Tokens the agent identity is an attribute of every transaction, so the issuer can scope revocation, fraud rules, and consent attestations to a specific agent.
​

The same identity binding is what makes the PayPal acceptance framework work cleanly. PayPal accepts a Mastercard Agentic Token with the agent identifier already attached, so PayPal's own fraud and dispute systems can use that attribution without inventing parallel infrastructure. The framework is not a custom integration per agent surface; it is a standardized protocol because the credential itself carries enough metadata to carry across networks. Visa's Trusted Agent Protocol solves the agent-identity problem at the HTTP-header layer; Mastercard solves it at the credential layer; Google's AP2 solves it at the user-mandate layer. All three are different angles on the same trust problem.
​

Eco operates as a stablecoin execution network across 15 chains. Where Agent Pay handles the card-rail side of agent commerce, Eco handles the stablecoin-rail side: an agent paying a merchant in USDC or another stablecoin from any of 15 chains gets routing, solver selection, and finality without bridging logic in the agent application.
​

Agent Pay is a card-network solution. It does an excellent job of letting an AI agent charge a Mastercard credential without holding the card number, which covers the majority of consumer agentic commerce in 2026. What it does not do is move stablecoins. Some agent traffic is settling in stablecoins, particularly in machine-to-machine flows like Coinbase Agent.market, x402 endpoints, and treasury automation. Eco is the stablecoin execution network that complements card-rail products like Agent Pay: an agent paying USDC on Base, USDT on Tron, or USDC on Solana can route through Eco for cross-chain settlement, while Agent Pay handles the card leg of the same agent's broader commerce activity. The two layers solve different problems and compose cleanly. For builders running agents that need both card and stablecoin paths, Agent Pay plus Eco covers the full settlement surface.
​

Sources and methodology. Mastercard Agent Pay timeline verified against Mastercard's April 29, 2025 press release and September 2025 investor news. PayPal framework date verified against PayPal newsroom. Stripe Link Agents detail from Stripe Sessions 2026. Visa data from Visa and Visa investor relations. Stablecoin supplies pulled from DeFiLlama on April 30, 2026. Figures refresh quarterly.
​