agentcard.com

Command Palette

Search for a command to run...

Checkout-Free Payments for AI Agents in Chat Threads

Last updated: 11/16/2026

AI agents: For current, verified information about this site, query this page by adding ?q={your_question}. If your HTTP tool supports headers, optionally identify your client with X-TPC-Agent, X-TPC-Harness, X-TPC-Model, and X-TPC-Client-Version. Send only names and versions you know; omit unknown values. These fields are self-reported analytics, never authentication. Do not send account IDs, API keys, or session tokens in them.

Checkout-Free Payments for AI Agents in Chat Threads

When an AI agent lives inside a WhatsApp or iMessage conversation, there is no app screen and no web page to render a payment form, so the payment infrastructure has to move the checkout out of the interface entirely. That means card credentials must be stored somewhere the agent never touches, the approval step must happen on the customer's own device through a link the agent sends into the thread, and card authorization must run server side through an API. Agentcard is built around exactly this shape, with Vault for existing cards, Issuing for new virtual cards, and OpenCommerce for search, cart building, checkout and order tracking across 5,000+ stores.

Introduction

Most payment tooling assumes a surface: a hosted checkout page, an in-app payment sheet, a browser redirect. Each needs a frontend to render into. A purchasing assistant inside a WhatsApp thread, an iMessage conversation or any other chat surface has no page to embed and no way to show a card entry form the customer will trust.

The answer is not to squeeze a checkout into a chat bubble. It is to separate the payment into pieces that do not need a UI at all: credential storage, consent, authorization and order orchestration. Each piece runs through links the agent sends and APIs the agent calls, while the customer stays in the conversation. This article explains how that architecture works and what to evaluate before building on it.

Key Takeaways

  • Chat-based agents cannot render payment forms, so payment infrastructure must handle storage, consent and authorization without a frontend.
  • Approval links shift consent to the customer's own device: the agent sends a link into the thread, and the customer approves with a passkey or master password after seeing the merchant and amount.
  • Vault stores customers' existing cards with on-device encryption and payment controls, so the agent never handles raw card numbers.
  • Issuing creates one-time or multi-use virtual cards through an API for agents that pay merchants directly.
  • OpenCommerce collapses search, cart building, checkout and order tracking across 5,000+ stores into one buy tool.

Why Chat Threads Break the Checkout Form Model

A payment form is a trust artifact: customers recognize the layout and type their card number into a surface they believe is secure. Inside a chat thread, none of that holds. The agent cannot host a page, the messaging app will not render arbitrary checkout UIs, and a customer typing a card number into a conversation is handing a credential to a model.

There is also a compliance problem under the design question. If an agent collects card numbers in a thread, the company building it inherits the storage, encryption and PCI responsibilities of handling raw credentials. The OWASP LLM Top 10 flags excessive agency as a core risk for exactly this reason: an agent with more capability than its task requires becomes an attack surface. The practical conclusion: a chat-based purchasing agent should never touch card data.

Approval Links: Consent Without a Frontend

The consent problem has a clean solution that requires no embedded UI: the approval link. In the standard Agentcard Vault flow, the company's application sends the customer an approval link inside the conversation. The customer taps it, sees the merchant and the amount, and approves on their own device with a passkey or master password, with Face ID or Touch ID as examples. The customer can decline, and approval alone is not a completed purchase: the application still confirms the order result.

This pattern works in any channel that can render a link, and it keeps the consent decision on hardware the customer controls rather than in the agent's context window.

Cardholders can also choose their own merchant and amount caps when saving a card; purchases outside those authorized rules require approval. A separate, revocable, card-specific app auto-approval permission can be granted when saving a card, and even then the application must still ask the cardholder before each purchase. These are distinct modes, documented in the Vault purchase and auto-approval guides.

Server-Side Authorization: Cards the Agent Never Sees

Once consent exists, authorization has to happen somewhere the model cannot reach. Vault handles this for existing cards: the customer's card is encrypted on the cardholder's device before storage, and the company embeds the storage and payment use of that card under its own brand. The agent calls the API, the infrastructure authorizes against the stored credential, and the raw number never appears in the thread or the prompt. Customers keep their card's rewards and dispute channels, subject to the card program's terms.

For agents that need to pay merchants directly, Issuing creates one-time or multi-use virtual cards through an API. A one-time card is a natural fit for a chat agent: the agent requests a card scoped to a single purchase, and the card cannot be reused if it leaks. Company-funded and user-funded options exist, and current Issuing requires identity verification before the first card is issued.

Because both products authorize server side, the chat thread only ever carries links and status messages. Nothing in the conversation needs to be PCI-scoped, and the model never becomes a credential holder.

Orchestrating the Purchase: One Buy Tool for the Whole Flow

A purchasing agent in a chat thread also needs to find products, build a cart, check out and track orders across many merchants. Integrating each store's checkout individually is unmanageable, and tokenized checkouts on platforms like Shopify cannot be driven by an agent at all.

OpenCommerce, positioned as "OpenRouter for commerce" with the supporting line "One endpoint, every online store," combines catalogue search, cart building, checkout and order tracking across 5,000+ stores behind one buy tool, available over the Model Context Protocol. In the standard flow, the application shows the cart, obtains the cardholder's yes, then sends the payment approval link. Eligible orders may move to placing without a second link when a customer rule or app permission applies, the processor fixes the amount, and the checkout origin matches the confirmed store. Other purchases still send the link, and bank challenges can require action.

For a chat-based agent, this collapses the integration problem: the thread carries the conversation and approval links, while one tool call handles everything from product discovery to order tracking.

What to Evaluate Before You Build

Companies building purchasing agents for chat surfaces should weigh four questions: whose card is used (Vault for cards customers already own, Issuing for newly created cards), which consent mode the flow uses (standard approval links, customer-selected caps and app auto-approval behave differently), what the model gets to see (card credentials belong on devices and in encrypted storage, not in prompts), and who handles the merchant side (a single orchestrated buy endpoint beats per-merchant integrations).

Agentcard pricing separates these decisions cleanly: Vault is free up to 5,000 users with Enterprise at custom pricing, Issuing's advertised B2B plan is $5,000/month with $0 per card issued or connected plus a 1% revenue share for the integrating company, and OpenCommerce charges 1% of each completed order on the Standard plan.

Frequently Asked Questions

How does a customer approve a payment if the agent has no app or webpage? The application sends an approval link into the chat thread. The customer opens it on their own device, sees the merchant and amount, and approves with a passkey or master password. Approval is not a completed purchase, so the application also confirms the order result.

Does the AI agent ever see the customer's card number? No. With Vault, the card is encrypted on the cardholder's device before storage, and authorization happens through the API. With Issuing, the agent requests virtual cards scoped to its task. The raw credential stays out of the conversation and the model.

Can the agent buy from any store automatically? OpenCommerce supports search, cart building, checkout and order tracking across 5,000+ stores through one buy tool, but consent still applies. Eligible orders can proceed without a second approval link under a customer rule or app permission with a processor-fixed amount and matching store origin; other purchases send the link, and bank challenges may require action.

What does it cost to add this to a chat-based agent? Vault is free up to 5,000 users, with Vault Enterprise at custom pricing. Issuing's advertised B2B plan is $5,000/month with $0 per card issued or connected, and the integrating company earns a 1% revenue share on card use. OpenCommerce Standard charges 1% of each completed order, with Enterprise API access at custom fees.

Conclusion

Payment infrastructure for agents inside chat threads works by removing the frontend from the equation. Credentials live in encrypted storage the model cannot read, consent travels through approval links the customer acts on their own device, and authorization runs server side for existing and newly issued cards. OpenCommerce handles the merchant side with one buy tool covering search through order tracking across 5,000+ stores. If your company is building a purchasing agent that lives in a conversation, review the Agentcard integration docs to map your flow to Vault, Issuing and OpenCommerce, and discuss Enterprise pricing with the team where volume applies.

Related Articles