Imprint has released a Model Context Protocol server that lets AI agents reach a member's rewards, virtual cards and transaction history. The release gives shoppers a clear checkpoint, because members get tools to review agent-initiated purchases and approve them. It is one of several launches in the same week that show loyalty, payments and retail platforms building the plumbing for agentic commerce.
What the Imprint Server Does
Imprint, a co-branded card and loyalty platform, built its MCP server so that compatible AI assistants can securely access three kinds of member information: rewards balances, virtual cards and transaction history. Model Context Protocol is an open standard that lets an assistant call outside tools in a consistent way. One connection can therefore serve many assistants, rather than requiring a custom integration for each.
The company says the release responds to a change in behavior. More consumers are asking AI assistants to manage their finances and make purchases, and brands need a way to meet those assistants on the assistants' terms. Imprint's framing is that the server helps brands adapt rather than being bypassed.
For shoppers, the practical effect is convenience. Loyalty points and transaction records that once lived only inside an app could become something an assistant can check on request. A question such as "Do I have enough rewards for this trip?" could be answered without the member copying figures by hand.
The Approval Step Is the Key Feature
The most important detail is the control, not the data access. Coverage of the launch says members get tools to review and approve purchases that an agent starts. In practice, the workflow looks like this:
- The assistant initiates a purchase on the member's behalf.
- The member sees the request and reviews it.
- The member decides whether to approve it, so final authority stays with the cardholder.
That pattern is spreading across agent commerce. The assistant can prepare a transaction, but a person confirms spending that matters. Reports do not describe Imprint's spending limits, the list of supported assistants, or pricing, so readers should check the company's documentation before relying on any specifics.
Part of a Wider Shift in Agent Commerce
Imprint's launch came alongside several other agent-facing products reported by industry trackers. Riskified introduced Agent Identity Risk Intelligence, which aims to identify the AI agent behind an order or customer service request and return a real-time risk signal. A Zendesk partnership extends that capability to refund, return and exchange requests.
LoyaltyLion added an MCP connection that lets compatible assistants query a brand's loyalty program configuration, performance, members and orders. For its branded agent, the company says brands review changes before they go live and an audit log tracks them. Meanwhile, VibeDM launched agentic shopping inside Instagram direct messages. A shopper can browse a product carousel, receive an AI try-on image from a selfie and check out in the same chat. The company launched in the United States and Europe with eight brands and says uploaded selfies are deleted after 48 hours.
Taken together, these products share a design theme: the agent acts, but identity checks, audit logs and human approvals sit alongside it.
Funding Behind the Launch
Imprint also raised $130 million in a round led by Khosla Ventures, which values the company at about $1.9 billion. Existing investors Hedosophia, Ribbit Capital, Spice and Thrive Capital participated, and Positive Sum joined as a new investor. The funding suggests investors see loyalty and payments data as valuable infrastructure for assistants that act for consumers.
Why It Matters
Agent commerce depends on three things. The assistant needs access to useful data. The merchant or platform needs to know which agent is acting. The consumer needs a reliable brake. Imprint's launch addresses the first and third, while Riskified's approach addresses the second. The releases show the industry moving past demos toward products that handle real money.
There are also open questions. Coverage does not explain how an agent authenticates to the server, how long access lasts, or what happens when a member approves a purchase that later turns out to be unwanted. Those details will decide whether the approval step is a strong safeguard or a formality. Developers building on MCP should expect every connected tool to ask the same questions.
For readers, the practical lesson is to treat agent access as a permission decision. Before connecting an assistant to a card or account, check whether approvals are required, whether spending limits exist, and whether a record of actions is kept. Those questions matter more than which assistant is most capable.
