Uploaded May 2026 | Updated September 2026, 3 days ago
It's not safe to share your credit card with your agent. Karen Serfaty, CEO and co-founder of AgentCard, walks through a debit card built for AI agents — one-time cards your agent can spend, that destroy themselves after a single use.
0:00 Intro — AgentCard is a debit card for agents
0:22 Why buying things with Claude is unsafe today
1:21 Create a one-time card in Claude with MCP
3:14 How one-time cards stay safe after exposure
3:34 Buy an OpenAI subscription with the agent
6:03 Try AgentCard and build crazy things
Filmed live at MCP Night. Karen demos the AgentCard MCP server end-to-end inside Claude: a human-approved card creation flow, then an agent-driven checkout that detects a URL, reads the amount on screen, mints a one-time card for that exact charge, and completes the purchase in the browser. The whole pitch hinges on a simple safety property — a card that destroys itself after first use is fine to expose to an LLM, fine to paste into a terminal, fine to hand to a browser-using agent. That property is what makes "let your agent buy things" go from reckless to routine. If you're building agents that need to transact and you've been hand-wringing about credentials, this is the demo to watch.
AgentCard: https://www.agentcard.sh
It's not safe to share your credit card with your agent. Karen Serfaty, CEO and co-founder of AgentCard, walks through a debit card built for AI agents — one-time cards your agent can spend, that destroy themselves after a single use.
0:00 Intro — AgentCard is a debit card for agents
0:22 Why buying things with Claude is unsafe today
1:21 Create a one-time card in Claude with MCP
3:14 How one-time cards stay safe after exposure
3:34 Buy an OpenAI subscription with the agent
6:03 Try AgentCard and build crazy things
Filmed live at MCP Night. Karen demos the AgentCard MCP server end-to-end inside Claude: a human-approved card creation flow, then an agent-driven checkout that detects a URL, reads the amount on screen, mints a one-time card for that exact charge, and completes the purchase in the browser. The whole pitch hinges on a simple safety property — a card that destroys itself after first use is fine to expose to an LLM, fine to paste into a terminal, fine to hand to a browser-using agent. That property is what makes "let your agent buy things" go from reckless to routine. If you're building agents that need to transact and you've been hand-wringing about credentials, this is the demo to watch.
AgentCard: https://www.agentcard.sh
