Guide

Virtual cards for AI agents: what limits actually exist

Almost every option lets you cap a card. Fewer can lock it to one merchant, close it after one charge, or ask your code before each purchase. Here is what nine real options enforce, from their own pages.

By Agent Lane Editorial · Page dated 19 October 2026

Why give an agent its own card?

A mistake, a bad instruction or a hijacked tool can only spend what one card allows, and you can freeze that card without touching your real one. It is the payment version of a dedicated agent wallet.

Which limits does each option enforce?

Listed by fit for a one-person or small business, not by any deal with us. A dash means the provider’s page doesn’t say.

Which one fits you?

• Testing one agent with your own money: Privacy Personal. It is free, and single-use or merchant-locked cards keep a test from turning into a subscription.
• A business that wants the agent to fetch its own card: Mercury agent cards draw the clearest line between what the agent and the person can do. Mercury states you are liable for what the agent buys, so treat the limit as your maximum loss.
• Your company already runs on Ramp or Brex: issue one virtual card per agent with a vendor lock and a low limit there, and keep admin logins and API tokens away from the agent.
• You want code to approve every purchase: Stripe Issuing or Lithic. Both send each authorization to your endpoint in real time, and Stripe’s standard program needs no sales call to start.
• You are building agent purchasing into an app: AgentCard, Naïve and AgentWallet are agent-first, but check the price and the per-card cap before you rely on them.

One test settles most choices: can the agent raise its own limit, unfreeze a card, or open a new one? If it can, the limit is a suggestion rather than a control.

What about Visa and Mastercard’s agent programs?

Visa Intelligent Commerce and Mastercard Agent Pay are network programs for issuers, merchants and developers, not cards you sign up for. They reach you through a card provider. Stripe, for example, says an agent can use a Visa or Mastercard agentic network token created from an Issuing card. Visa says its program is still being deployed.

Setup checklist

1. Pick the cap first. Decide the most you would accept losing on this one job. The agent spend-cap calculator gives a per-transaction and daily starting point.
2. One card per agent and job. A separate card for each task lets you freeze one without touching the rest.
3. Set the limit before the agent sees the number. Per transaction plus a daily or weekly window, not an open monthly cap.
4. Lock it to where it should spend. Merchant-locked, single-use, or allowed categories, where the provider offers them.
5. Keep admin power with a person. Card details go in the agent’s secret store or the provider’s vault, never in chat. The agent never gets your admin login or secret API key.
6. Test small, then review. Run one small purchase first, check the receipt, then review weekly and freeze the card at the first odd charge.

What a card limit won’t do

• Judge whether a purchase was a good idea. For anything new, keep ping, don’t spend.
• Always catch late add-ons. Stripe notes tips and fees can post later and push spending past a limit.
• Replace the other layers. A card sits inside your AI agent spending limits: monthly budget, session ceiling and daily max.

Want it as a printable checklist? The AI Spend-Cap Setup Guide covers virtual cards, caps and what never to give an AI.

Common questions

What is a virtual card for an AI agent?

A card number that exists only online and is issued for one agent. It has its own limit and can be frozen or closed alone, so a mistake can cost no more than that card allows.

Can the agent raise its own limit?

It depends on the setup. Mercury says agents cannot create agent cards, raise limits or unfreeze them. With a card API such as Stripe Issuing or Lithic, anything holding your secret API key can change limits, so the agent must never hold it.

Does a real-time approval webhook mean I approve each purchase by hand?

No. Stripe waits 2 seconds for an answer and Lithic 6 seconds, so your own code decides, not a person. If you want a human yes or no before the agent spends, have it ask first; see ping, don’t spend.

Is a prepaid card the same thing?

Not quite. A prepaid card spends money loaded onto it. Most cards here draw from your account per purchase and stop at the limit you set. Either way, the limit is what stops the agent.

Sources

Checked 19 October 2026 on each provider’s own site. Prices and limits change.

• Privacy — pricing and plan comparison
• Privacy — Who can use Privacy?
• Privacy — AI agents (MCP and CLI)
• Privacy Developer API — Transaction webhooks
• Mercury — Agent cards
• Mercury — Eligibility and requirements
• Mercury — pricing
• Ramp — virtual cards
• Ramp — pricing
• Ramp — finance for agents
• Brex — corporate cards
• Brex — pricing
• Stripe — Issuing for agents
• Stripe — Issuing spending controls
• Stripe — Issuing real-time authorizations
• Stripe — Issuing lifecycle controls
• Stripe — Issuing for your business (standard program limits)
• Stripe — How Issuing works (availability)
• Stripe — pricing (Issuing)
• Lithic — Spend limits
• Lithic — Auth Stream Access (ASA)
• Lithic — Create card API (card types)
• Lithic — Merchant category codes in authorization rules
• AgentWallet — virtual card
• Naïve — cards
• Naïve — pricing
• AgentCard — pricing
• Visa — Visa Intelligent Commerce
• Mastercard — Agent Pay

Related guides

• How to test an AI agent before real access
• AI agent spending limits
• Should you give an AI agent your credit card?
• Can AI agents make purchases?
• Whitelist merchants only
• Never paste API keys in chat

All guides