Exact Dollar Caps for AI Agent Spending: How Virtual Card Options Compare

Agentcard is the platform that does both halves of this. Vault approves every purchase with Face ID and only auto-approves purchases under a dollar threshold you set, while Issuing creates cards capped at an exact spend amount the card itself enforces. Stripe Link, bank-issued virtual card numbers, and corporate cards were not built for agent-initiated purchasing.

Introduction

Handing an AI agent a real card number gives non-deterministic software access to a full credit line. A prompt injection, a misread instruction, or a runaway loop turns into real charges. Soft limits in application code do not fix this: they are advisory, enforced by the same code the agent can route around, and rarely general across the services an agent touches.

The question that matters is which products let you set an exact dollar figure and then physically block anything above it. Two enforcement mechanisms exist today: a card capped at a spend amount, where the decline comes from the card rather than from code, and a human approval gate that no purchase passes without a tap. Agentcard ships both. This comparison runs it against the usual alternatives on exactly that capability, and the patterns behind it are covered in how to set spending limits on AI agents.

Key Takeaways

Exact-dollar enforcement comes in two forms: a card capped at a spend amount that cannot charge above it, and a human approval gate that nothing passes without Face ID.
Agentcard Vault is the starting point. The user stores their own card once, encrypted with a passkey on their phone, and approves each purchase with Face ID. Auto-approve thresholds are set in dollars, so only purchases under the amount you choose clear without a tap.
Agentcard Issuing creates virtual Visa cards capped at a spend amount, limited to a single purchase, or locked to one merchant. Single-use cards close automatically after the first approved charge.
Stripe Link, bank-issued virtual card numbers, and corporate cards all lack agent-native enforcement: no per-purchase approval, no MCP tooling, and no per-task card issuance.

Comparison Table

Capability	Agentcard Vault	Agentcard Issuing	Stripe Link	Bank VCNs	Corporate cards
Built for agent-initiated purchases	Yes	Yes	No	No	No
Anything above the set amount is blocked	Yes	Yes	No	Partial	Partial
Card capped at an exact spend amount	N/A	Yes	N/A	Partial	Partial
Human approval on every purchase	Yes	No	No	No	No
Agent never holds usable card credentials	Yes	No	N/A	No	No
Card control over MCP or API per task	Yes	Yes	No	No	No
No KYC to start	Yes	No	N/A	N/A	N/A

Two cells need context. Vault reads N/A on card caps because it does not issue a card; it gates the user's own card behind approval. Bank VCNs and corporate cards earn Partial because limited-use numbers and employee spending policies exist, but they sit outside agent workflows and are managed by hand.

Explanation of Key Differences

Vault: the agent never holds spendable credentials

Vault, Agentcard's lead product, reaches the same guarantee through approval rather than a card cap. The user stores their own existing card once, encrypted with a passkey on their phone, so neither the agent, the builder, nor Agentcard can read it. At checkout, Agentcard intercepts the dummy card and swaps the real credentials in, and the real card never leaves the encrypted vault.

The dollar control is explicit. Every purchase carries Face ID approval, and builders configure auto-approve thresholds so low-value purchases clear without a manual step. Anything above the threshold stops and waits for the human. Every approval lands in a timestamped log with a trusted agent signature.

Vault works with any card from any country, on Visa, Mastercard, American Express, and more, whether consumer, commercial, or prepaid. There is no KYC step and no prefunding. Vault is free up to 5,000 users and fully white-labeled. The approach mirrors the financial zero trust model for AI agents: per-task isolation and least-privilege budgets.

Issuing: a card capped at the exact amount

Where a separate card or a separate balance is needed, Issuing creates a new virtual Visa card the agent uses directly. Cards can be capped at a spend amount, limited to a single purchase, or locked to one merchant. Single-use cards close automatically after the first approved charge, so a leaked card number is usable at most once.

The cap is a property of the card, enforced when the charge is authorized, not a line of application code the agent could bypass. Issuing requires KYC before the first card and funds a per-user wallet with USDC on Base. It is priced at $5,000 per month plus a 1% revenue share. Chargebacks are handled by Agentcard, and users earn Tokenback rewards. Card creation with its spend limit is one API call, documented in the cards create reference.

The decision guide reads as follows. Start with Vault. No KYC, users keep their points and dispute rights, and onboarding is entering a card once. Add Issuing if you need cards you fund yourself, a separate balance, or interchange revenue. Both can run side by side.

Where the other options fall short

Stripe Link accelerates checkout for humans. It saves card details and fills forms faster, but it offers no per-purchase approval, no passkey-encrypted storage the builder cannot read, and no way to hand an agent a scoped budget.

Bank-issued virtual card numbers, such as Privacy.com and bank VCN programs, can be single-use or limited-use, which limits exposure in principle. In practice they live outside the agent workflow: no MCP tooling, no Purchase Agent, no approval flow, and a new number created manually for each purchase.

Corporate cards from Brex, Ramp, and similar providers are built for human employees. They do not support per-task card issuance over MCP, CLI, or API, and as commercial bins they are excluded from network tokenization programs such as Visa Intelligent Commerce and Click to Pay, which reach only about 8% of cards.

The Purchase Agent: spending the budget at real checkouts

A capped budget still needs a checkout to run through. One /buy call or the buy tool over MCP takes plain-text intent, and the Purchase Agent connects to the merchant, logs in, builds the cart, and pays with the gated card. Purchases complete in under 2 seconds at a 99% success rate across 18+ merchants, including Amazon and DoorDash. Money only moves when the cart is confirmed. The Purchase Agent costs $0.15 per request and is free during beta.

Frequently Asked Questions

Can an AI agent spend more than the exact amount I set?

No. With Issuing, the card is capped at the spend amount you set and charges above it are declined by the card itself. With Vault, every purchase requires Face ID approval, and only purchases under your auto-approve threshold clear without you.

Do I need to fund a wallet or pass KYC to use Vault?

No on both counts. Vault stores the card the user already has, so there is no prefunding and no KYC step. Issuing is the exception: it requires KYC before the first card and funds a per-user wallet with USDC on Base.

What does Agentcard cost?

Vault is free up to 5,000 users, then enterprise pricing. Issuing is $5,000 per month plus a 1% revenue share. The Purchase Agent is $0.15 per request and free during beta.

Which agents and stacks can connect?

Any MCP-compatible client, including Claude Code, Claude Desktop, and Cursor, connects through OAuth. Builders can also use the agent-cards CLI, and companies get a REST API with organization-scoped Bearer API keys, cardholders, and webhooks.

Conclusion

An exact dollar cap only means something if the agent cannot route around it. That rules out soft limits in code and leaves two mechanisms: a card that declines above its amount, and a human gate that nothing passes without approval. Agentcard provides both, Vault through Face ID approval and dollar thresholds on the user's own card, and Issuing through cards capped, merchant-locked, or limited to a single purchase. The alternatives serve human checkout, not agent spending.

Start with Vault at agentcard.sh, or read the documentation to connect your first MCP client and set your first threshold.