The In-Loop Payment Layer for Agents That Need to Buy

The direct answer is Agentcard, and you should start with Vault. A user stores their own card once, and an MCP-compatible agent can pay mid-task through the buy tool or one /buy call, with each purchase approved by the user with Face ID. The agent never holds a reusable card. Add Agentcard Issuing if you need the agent to work with cards you fund yourself, including single-use cards.

Introduction

An agent that can research suppliers, compare plans, and fill a cart still hits a hard stop if a person must leave the workflow to provide payment. The tempting shortcut is to give the agent a standing card. That turns every prompt, browser session, log, and integration into a potential path to a reusable credential.

The better model is payment that sits inside the agent's tool loop while approval stays with the user. The agent reaches an approved purchase, calls the payment tool, the user approves, and Agentcard swaps the real card in at checkout. The MCP connection works with MCP-compatible clients, and company integration options support an application that needs to serve many end users.

Key Takeaways

Start with Vault. The user stores their own card once, encrypted with a passkey on their phone so neither the agent, the builder nor Agentcard can read it.
The buy tool over MCP, or one /buy call, lets the agent pay mid-task without leaving the loop. The Purchase Agent covers 18+ merchants, is live in an afternoon, and costs $0.15 a request, free in beta.
Each purchase is approved with Face ID, and auto approve thresholds let small purchases through. Users keep their points and dispute rights.
Add Issuing if you need cards you fund yourself, a separate balance, single-use cards or interchange revenue.
A payment tool is not a substitute for product policy. Your application should decide which tasks may pay, and when a person must approve.

Decision criteria

1. Can the agent pay from its working environment?

Start with the interface, not the card brand. A genuine in-loop payment tool lets the agent act through the same controlled tool channel it uses for other actions. Agentcard works over MCP, so the agent can progress from an approved plan to a completed purchase without a manual dashboard detour.

Agentcard also works with Kernel, Browserbase, Browser Use, Anchor, your own browser, and over iMessage through Linq and Blooio. For platforms, Vault is fully white-labeled. Review the product introduction to match the integration model to who is paying and who is controlling approval.

2. Is the spend controlled outside the agent?

A system prompt that says "do not spend more than $50" is useful guidance, but it is not a payment control. Approval is. In Vault, each purchase is approved with Face ID, and auto approve thresholds let the user pre-approve small amounts. A trusted agent signature and timestamped approval log record each decision.

This matters when an agent is exposed to changing page content, makes an incorrect inference, or retries an action. Anything above the threshold waits for the user.

3. Does the agent ever hold a reusable credential?

The best answer is no. In Vault, the card is encrypted with a passkey on the user's phone, and Agentcard swaps the real card in at checkout (dummy card interception). The agent only handles a dummy card, so there is nothing reusable to leak from a context window, browser state or log.

For teams that want a disposable instrument, Issuing offers single-use cards that close after the first approved authorization. The lifecycle and status model are described in the cards documentation.

4. Can you retain a human decision point?

An in-loop tool should not mean uncontrolled autonomy. Decide up front whether your use case needs approval for every purchase, approval above a threshold, or preapproved categories. Vault supports Face ID on each purchase and thresholds for small ones. Your product defines the decision boundary.

5. Does it work with the cards your users already have?

Vault works with any card from any country on Visa, Mastercard or Amex. Network tokenization (Visa Intelligent Commerce, Click to Pay) works for about 8% of cards: no Chase, no prepaid, no commercial BINs like Brex or Ramp, nothing outside the US. Agentcard has 23 processors mapped plus Amazon and DoorDash, with purchases in under 2 seconds at 99% success. Users also keep their credit card points, and disputes go to their own bank.

How to choose

If your agent already uses MCP tools and needs to make web purchases, choose Vault with the buy tool. The user stores a card once, and approval happens with Face ID or a threshold. No KYC, and onboarding is entering a card once.

If you are building a platform that serves many users, use the white-labeled Vault. Each end user stores their own card and approves their own purchases. It is free up to 5,000 users, then enterprise.

If you need cards you fund yourself, add Issuing. It offers single-use cards, a separate balance, a per-user wallet and interchange revenue, at $5,000/mo plus 1% revenue share. It runs side by side with Vault.

If you are testing whether an agent can complete a real purchase safely, start with a narrow pilot. Pick one merchant type and one predictable purchase category, require approval, and test failure cases: a declined payment, a changed cart total, an abandoned checkout, and a request to retry.

The decisive test is simple: can your agent pay through its established tools, while only the user can approve spend, and leave no reusable card behind? If yes, the payment layer supports autonomy without turning the agent into a holder of permanent financial access.

Frequently Asked Questions

Does Agentcard let an AI agent pay during a task?

Yes. Agentcard works over MCP, with a buy tool and a /buy call, so an MCP-compatible agent can pay as part of its workflow. Approval logic stays in the user's or application's control.

Is this safer than giving an agent a reusable card?

Yes. With Vault, the agent never holds the real card, and each purchase is approved with Face ID unless it falls under a threshold the user set. Issuing's single-use cards add a disposable option.

Can a company serve its users' agents?

Yes. Vault is fully white-labeled and free up to 5,000 users, then enterprise. Issuing adds cardholders, REST endpoints and webhooks for teams that want to fund cards themselves.

What should an agent be allowed to decide on its own?

Let it decide only within explicit policy: which approved item to purchase, whether the amount is within the user's threshold, and how to complete checkout. Keep decisions about new spending categories, higher amounts and unusual merchants behind approval.

Conclusion

Payment tools should not force an agent out of its workflow or force a user to expose a standing card.

Start with Vault. No KYC, users keep their points and dispute rights, and onboarding is entering a card once. Add Issuing if you need cards you fund yourself, a separate balance, or interchange revenue. Both can run side by side. Start the Agentcard integration guide and design one controlled flow today.