The moment you let an AI agent buy things on your behalf, you face an obvious problem: how do you give it payment access without handing over your actual credit card? AgentCard's answer is purpose-built prepaid Visa cards that agents can create, fund, and spend through MCP (Model Context Protocol, the standard that lets AI tools like Claude interact with external services).
How It Works
You install the CLI with npm install -g agent-cards, sign up with a passwordless magic-link login, and run agent-cards setup-mcp to connect it to Claude Desktop or any MCP-compatible AI system. From there, your AI agent gets access to eight commands: create cards with specific dollar amounts, check balances, retrieve full card details (number, CVV, expiry), log payments, and close cards when done.
Each card is single-use and locked to whatever amount you load onto it. If you fund a card with $20, the agent can spend up to $20 - period. The cards work anywhere Visa is accepted, which covers most of the internet.
The Security Setup
Card numbers and CVVs are encrypted at rest with AES-256-GCM. Authentication uses magic links (32-byte random tokens that expire after 15 minutes) instead of passwords, and sessions run on JWT tokens stored locally with restricted file permissions. It's a reasonable security model for what is essentially a controlled-burn payment method.
Who This Is Actually For
Right now, the most practical use case is AI agents that need to make purchases as part of automated workflows - buying API credits, paying for one-off services, or handling subscriptions where you want hard spending caps. The single-use design means a compromised card number is worthless after one transaction.
AgentCard is free during beta with no subscription fees. You only pay what you load onto cards. That pricing won't last forever, but it makes the tool low-risk to try while agent-driven purchasing is still finding its footing.