AgentCard: Virtual Visa Cards for AI Agents (2026)
AgentCard gives AI agents prepaid virtual Visa cards via MCP. Free tier caps cards at $50. Works with Claude Code, Cursor, Cline, and 40+ MCP clients. YC S2026.
AgentCard is an MCP server that gives AI agents their own single-use virtual Visa cards. You approve each card, the agent creates it and pays with it, and it self-destructs after one use. Free plan: $50 per card. Basic plan: $500 per card. Made by Tomorrowland Beta (YC S2026). Works with Claude Code, Cursor, Cline, and any MCP-compatible agent.
AgentCard is an MCP server by Tomorrowland Beta (YC S2026) that lets AI agents issue single-use prepaid virtual Visa cards. Agents call create_card with an amount in cents, receive full card credentials, and pay anywhere Visa is accepted. The free tier caps cards at $50; Basic raises the limit to $500. Works with Claude Code, Cursor, Cline, Windsurf, and 40+ MCP clients.
Maker: Tomorrowland Beta · Protocol: MCP · Auth: token
Compatible agents: Claude Code, Claude Desktop, Cursor, Cline, Windsurf, OpenClaw, Any MCP HTTP client
Required runtime: Node.js >= 18 (for agent-cards CLI), Stripe-connected payment method (for card funding)
About AgentCard
AgentCard is an MCP (Model Context Protocol) server built by Tomorrowland Beta, a Y Combinator Summer 2026 company based in San Francisco, that gives AI agents the ability to issue and spend prepaid virtual Visa cards. Co-founded by Karen Serfaty, who previously led Atlas (an expense card acquired by Remote.com), and Felipe Abello, AgentCard launched publicly in early 2026 and demoed live at MCP Night 4 in May 2026, where an agent paid for an OpenAI subscription entirely from the terminal. The service exposes more than 20 MCP tools covering card creation, balance checks, transaction history, payment method management, checkout autofill, and a conversational buy tool for natural-language purchasing. AgentCard runs as a hosted HTTP MCP server at https://mcp.agentcard.sh/mcp. The Personal tier authenticates via OAuth magic-link email, storing a JWT as the AGENT_CARDS_JWT environment variable. The Companies tier uses sk_test_* and sk_live_* API keys for platforms issuing cards to multiple end-users at scale. All card credentials are encrypted at rest with AES-256-GCM and only decrypted when an agent explicitly calls get_card_details. A human approval checkpoint fires before any card is minted or any charge is authorized, so agents can propose spending but cannot execute it unilaterally. AgentCard works with any MCP-compatible agent, including Claude Code, Claude Desktop, Cursor, Cline, Windsurf, and OpenClaw, and the company claims compatibility with over 40 MCP clients in total. Developers add one JSON config block pointing at the MCP server URL, install the agent-cards npm package, and call get_instructions to have the agent load its own usage guide. Ideal workflows include automated SaaS subscription renewal, API credit top-ups, one-off tool purchases, and e-commerce checkout flows where the agent needs to pay with a card scoped to the exact transaction amount. AgentCard offers a free tier with a $50 maximum per card and a Basic tier with a $500 maximum per card. There are no monthly subscription fees on the free plan; card funding goes through Stripe holds so money is only captured when the card is used. Stripe's standard processing fees apply on top. The Companies tier pricing is available on request for platforms issuing cards at scale. AgentCard is in active development as of June 2026, with the product updates page showing weekly shipping of new features including hold-based funding, real-time authorization webhooks, automatic stale-card cleanup, and a built-in support chat tool. The CLI auto-upgrades itself on each run. The team is building toward fully autonomous agent purchasing workflows that can handle multi-item cart checkout, merchant linking, order tracking, and delivery address management entirely via MCP.
Key Features
- MCP-native card issuance: Agents call create_card with an amount in cents and receive a live Visa card number, CVV, and expiry within seconds, with no custom HTTP integration code required.
- Human approval checkpoint: Every card creation and every charge requires explicit human sign-off before funds move, so agents can propose spending but cannot execute it unilaterally.
- AES-256-GCM credential encryption: Full card credentials (PAN, CVV, expiry) are encrypted at rest and only decrypted when an agent explicitly calls get_card_details, reducing credential exposure in logs or memory.
- 20+ specialized MCP tools: The server exposes tools for card lifecycle management, balance checks, transaction history, payment method management, plan upgrades, checkout autofill, and natural-language purchasing via the buy tool.
- Visa acceptance worldwide: Each issued card is a prepaid Visa accepted at any online merchant worldwide that takes Visa, with no geographic or merchant category restrictions beyond the card's fixed spend cap.
- Test and live mode switching: Developers switch between sandbox and production environments via the set_mode MCP tool, enabling full end-to-end testing before committing real money.
Use Cases
- Automated SaaS subscription renewal: An agent monitors a subscription expiry date, calls create_card with the exact renewal price, then fills the card at the merchant checkout so the developer never shares their real payment credentials.
- API credit replenishment: A coding agent tracks its own remaining API credits, issues a $20 card when the balance drops below a threshold, and tops up its provider account without any human touchpoint beyond the initial approval.
- Natural-language e-commerce checkout: A shopping agent calls the buy MCP tool with a plain-English description such as 'buy a 6-month gym membership', browses linked merchants, builds a cart, and completes checkout with a freshly issued card scoped to the exact cart total.
- One-off software license procurement: A developer agent identifies a missing dependency or tool, purchases a $10 license, and installs it in one terminal session without the developer ever entering payment details manually.
Install
npx agent-cards setup-mcp
Requirements
- Node.js 18 or later installed (npm comes with Node)
- Run agent-cards login to authenticate via magic-link email and generate your AGENT_CARDS_JWT
- Add a payment method via the setup_payment_method MCP tool or Stripe Checkout before creating any cards
- Add the MCP server to your agent config: URL https://mcp.agentcard.sh/mcp, transport HTTP, Authorization: Bearer <AGENT_CARDS_JWT>
Actions
Create Card
Issues a new single-use prepaid virtual Visa card with a fixed spend limit, holding the amount on your saved payment method.
{
"name": "create_card",
"arguments": {
"amount_cents": 2500
}
}amount_cents(number) — required: The card's spend limit in US cents. For example, 2500 = $25.00. Free plan maximum is 5000 (=$50); Basic plan maximum is 50000 (=$500).
List Cards
Returns all virtual cards on the account with their IDs, last four digits, expiry dates, balances, and statuses.
{
"name": "list_cards",
"arguments": {}
}Check Balance
Retrieves the current balance of a specific card without exposing the full card credentials.
{
"name": "check_balance",
"arguments": {
"card_id": "card_abc123"
}
}card_id(string) — required: The unique ID of the card to check, as returned by create_card or list_cards.
Get Card Details
Decrypts and returns the full card number (PAN), CVV, and expiry date needed to complete a payment. May require human approval before returning credentials.
{
"name": "get_card_details",
"arguments": {
"card_id": "card_abc123"
}
}card_id(string) — required: The unique ID of the card whose credentials to retrieve.
Close Card
Permanently closes a virtual card. This action is irreversible and any remaining balance is released back to the payment method hold.
{
"name": "close_card",
"arguments": {
"card_id": "card_abc123"
}
}card_id(string) — required: The unique ID of the card to permanently close.
List Transactions
Retrieves the full transaction history for a specific card, showing merchant name, amount, status, and timestamps.
{
"name": "list_transactions",
"arguments": {
"card_id": "card_abc123"
}
}card_id(string) — required: The unique ID of the card whose transactions to retrieve.
Setup Payment Method
Opens a Stripe Checkout session to save a new payment method (credit or debit card) to the account, which is used to fund future virtual cards.
{
"name": "setup_payment_method",
"arguments": {}
}List Payment Methods
Returns all saved payment methods on the account with their brand, last four digits, expiry, and which one is set as the default.
{
"name": "list_payment_methods",
"arguments": {}
}Get Plan
Displays the current plan (Free or Basic), the per-card spending cap, monthly card usage count, and available upgrade options.
{
"name": "get_plan",
"arguments": {}
}Buy
Conversational purchase tool that accepts a natural-language description of what to buy and handles merchant lookup, cart building, and checkout end-to-end.
{
"name": "buy",
"arguments": {
"description": "Buy a 1-month ChatGPT Plus subscription for $20"
}
}description(string) — required: Plain-English description of the purchase to complete, including the item, quantity, and optionally the target merchant.
Pay Checkout
Creates a new virtual card and immediately fills it into an active browser checkout form in a single atomic operation, minimising the window in which credentials are exposed.
{
"name": "pay_checkout",
"arguments": {
"amount_cents": 4999
}
}amount_cents(number) — required: The exact amount in cents to load onto the card and charge at the checkout form currently open in the browser.
Get Instructions
Returns the current usage guide for AgentCard MCP tools. Call this at the start of any agent session to ensure the agent uses the most up-to-date workflow.
{
"name": "get_instructions",
"arguments": {}
}Get Funding Status
Polls the status of a card creation after the Stripe Checkout payment step, confirming when the card is funded and ready to use.
{
"name": "get_funding_status",
"arguments": {
"card_id": "card_abc123"
}
}card_id(string) — required: The card ID returned by create_card, used to poll whether the Stripe payment has settled and the card is active.
x402 Fetch
Performs an HTTP request and automatically handles HTTP 402 Payment Required responses by issuing a virtual card and paying the requested micropayment amount without human intervention.
{
"name": "x402_fetch",
"arguments": {
"url": "https://api.example.com/premium-endpoint",
"method": "GET"
}
}url(string) — required: The URL to fetch. If the server returns HTTP 402, AgentCard automatically pays and retries.method(string): HTTP method: GET, POST, PUT, DELETE. Defaults to GET.body(string): Request body for POST/PUT requests, as a JSON string.headers(object): Additional HTTP headers to include in the request.
How to Invoke
Exposed as 20+ MCP tools (create_card, list_cards, check_balance, get_card_details, buy, and more) called by any MCP-compatible agent after registering the HTTP server at https://mcp.agentcard.sh/mcp with JWT Bearer auth.
Pricing
Free plan: $50 max per card, no monthly fee. Basic plan: $500 max per card, upgrade via Stripe Checkout. Companies tier pricing on request. Stripe transaction processing fees apply to all card funding.
Strengths
- Zero integration code required: any MCP-compatible agent (Claude Code, Cursor, Cline, Windsurf) can start issuing cards with one config block and npm install -g agent-cards.
- Spend-cap safety by design: cards are backed by Stripe payment-method holds, so the maximum possible loss from any runaway agent is the card's loaded amount with zero overdraft risk.
- Human-in-the-loop gates: approval checkpoints before each card mint mean developers retain full financial control even as agents operate autonomously across multi-step workflows.
- The only virtual card product in 2026 built natively on MCP, which means agents use standardized tool-calling instead of bespoke REST wrappers.
Weaknesses
- Free tier $50 per-card cap is low: a single $49/month SaaS subscription or a mid-range API credit top-up hits the ceiling and requires a Basic plan upgrade.
- Human approval adds workflow friction: the required approval step before each card creation breaks fully autonomous pipelines and cannot be pre-authorized for specific merchants or amount ranges.
- Beta status as of June 2026: documentation warns to expect rough edges and the API surface has been actively changing, including wallet removal and hold-based funding introduced mid-cycle.
- Self-hosted Docker setup requires managing JWT auth and the AGENT_CARDS_JWT environment variable yourself, adding operational overhead compared to the one-command hosted path.
Frequently Asked Questions
What is AgentCard and what does it do?
AgentCard is an MCP (Model Context Protocol) server built by Tomorrowland Beta, a Y Combinator Summer 2026 company based in San Francisco, that gives AI agents the ability to issue and spend prepaid virtual Visa cards. Instead of embedding your real payment credentials inside an agent's context window, you give the agent access to AgentCard's MCP tools and it creates a new single-use card for each task. Each card has a fixed dollar amount, is accepted anywhere Visa is accepted worldwide, and is automatically cancelled after one authorized use. The server exposes more than 20 MCP tools covering card creation, balance checks, transaction history, payment method management, and a conversational buy tool for natural-language purchasing. All card credentials are encrypted at rest with AES-256-GCM and only decrypted when an agent explicitly calls get_card_details. A human approval checkpoint is required before any card is minted, so agents can propose spending but cannot execute it without sign-off. AgentCard is the only virtual card provider in 2026 built natively on MCP, requiring no custom integration code on the developer's side.
How do I install and set up AgentCard?
Install the AgentCard CLI globally with npm install -g agent-cards, which requires Node.js 18 or later already on your machine. After installation, run agent-cards login to sign in via magic-link email; this generates a JWT stored as the AGENT_CARDS_JWT environment variable that you will use to authenticate the MCP server. Add the MCP server to your agent's configuration by pointing it at https://mcp.agentcard.sh/mcp with HTTP transport and your JWT as a Bearer token in the Authorization header. Before creating any cards, add a payment method by calling the setup_payment_method MCP tool, which opens a Stripe Checkout page where you save a credit or debit card. The Personal tier is for individual developers managing their own agent's card; the Companies tier uses sk_test_* and sk_live_* API keys for platforms issuing cards to multiple end-users at scale. To test without spending real money, call set_mode with 'test' to switch to sandbox mode before creating any cards. Once live, call get_instructions at the start of each agent session so the agent loads the latest usage guide and workflow rules.
Which agents and LLMs support AgentCard?
AgentCard works with any agent or LLM client that connects to MCP servers over HTTP, which in 2026 includes more than 40 clients. Verified integrations include Claude Code, Claude Desktop, Cursor, Cline, Windsurf, and OpenClaw, all of which support the MCP HTTP transport that AgentCard uses. Because the server follows the standard MCP spec with JSON-Schema-defined tool parameters, it is also compatible with any agent that implements MCP via a bridge, including custom agents built with the Anthropic or OpenAI SDKs. Setup for Claude Code requires adding one JSON block to your MCP config file that specifies the server URL and the JWT environment variable; the process is identical for Cursor and Cline. AgentCard does not tie you to a specific underlying LLM model, so agents powered by Claude, GPT, Gemini, or any other model can all issue and use cards. The tiny-agent-company GitHub organization also publishes a Claude Code skill that teaches agents the correct AgentCard workflow including safety rules and error handling. For platforms building multi-user products, the Companies tier lets you issue OAuth MCP credentials per end-user so each user's agent has its own card scope.
How much does AgentCard cost?
AgentCard has no monthly subscription fee on its Free plan, making it free to start for individual developers. The Free plan caps each individual virtual card at $50 maximum; you can issue multiple cards per month up to that per-card limit. The Basic plan raises the per-card ceiling to $500 and is available by calling the upgrade_plan MCP tool or running agent-cards plan from the CLI, which opens a Stripe Checkout to pay for the upgrade. Card funding works via Stripe holds: the amount is held on your saved payment method when the card is created and only captured when the card is used, so unused cards release their hold automatically. Stripe's standard card processing fees, typically around 2.9% plus $0.30 per transaction, apply on top of the card's loaded amount. The Companies tier pricing for platforms issuing cards to multiple users at scale is available on request from AgentCard. Monthly card usage and your current plan cap can be checked at any time by calling the get_plan MCP tool.
Is AgentCard open source?
Yes, AgentCard's MCP server is open source under the MIT license. The source code is published at https://github.com/radiahq/mcp by the Radia team behind agentcard.sh. The MIT license means you can inspect the code, modify it, and self-host the server on your own infrastructure via Docker. Self-hosting gives teams with strict data-residency or compliance requirements full control over where card credentials are decrypted and processed. The agent-cards npm CLI package is also published openly on npm. Payment processing still routes through Stripe under the hood, which means PCI compliance for card data in transit inherits from Stripe's certified infrastructure regardless of whether you self-host the MCP layer. Developers who want to run the server locally can clone the repo, run npm install, and start it with npm run dev for stdio mode or npm run dev:http for HTTP mode on port 3002.
What are the best alternatives to AgentCard for AI agent payments?
The closest alternatives to AgentCard for AI agent spending are Privacy.com, Stripe Issuing, and Mastercard's agent toolkit. Privacy.com lets you generate virtual card numbers via a browser extension or REST API, but in 2026 it has no native MCP integration and requires custom code wrappers for each agent runtime, adding setup overhead. Stripe Issuing is a developer-grade API capable of creating virtual Visa and Mastercard numbers programmatically at scale, but it targets fintech builders, requires a compliance review and Stripe Treasury onboarding, and is not designed for individual developers wanting a five-minute setup. Mastercard published an MCP-based agent payment toolkit aimed at enterprise payment platforms, but it is not publicly accessible without a partnership agreement. Ramp and Brex offer virtual corporate cards for human employees but neither exposes an MCP server and both require full company onboarding. For agents that only need to pay for sub-dollar services in real time, an x402 micropayment wallet is lighter than a full Visa card issuance flow. AgentCard is the only product in 2026 built specifically for agents on MCP, which gives it a significant setup speed advantage over any REST-only competitor for developers already using Claude Code or Cursor.
Who is AgentCard best for?
AgentCard is best suited to developers building autonomous AI agents that need to buy things on the internet: purchasing API credits, renewing SaaS subscriptions, ordering cloud resources, or completing checkout flows without the developer manually entering card details at each step. The free tier is well-suited to solo developers and hobbyists experimenting with agent purchasing up to $50 per card, covering most single SaaS trial signups and API credit bundles. The Basic tier at $500 per card covers the full range of software subscriptions, mid-range API credit purchases, and small-scale procurement that a coding or ops agent would handle in practice. It is NOT the right tool for agents that need to make high-volume micro-purchases below a few dollars, where x402 payments or native API billing are faster and cheaper. Compliance-sensitive teams in healthcare, finance, or other regulated industries should verify with AgentCard whether the hosted card infrastructure meets their specific regulatory requirements before building production workflows. Companies building multi-user agentic products, where each end-user's AI agent needs its own capped budget, are the primary audience for the Companies tier. Developers whose agents already run inside Claude Code or Cursor will find the setup especially fast because AgentCard ships one-click MCP config examples for both.
How does AgentCard compare to Stripe Issuing for AI agent payments in 2026?
AgentCard and Stripe Issuing both issue prepaid virtual Visa cards, but they target very different audiences and require very different setup investments. Stripe Issuing is a full card-program API built for fintech companies: it requires applying for Stripe Treasury access, passing compliance review, and writing significant backend code to create cardholders, fund balances, and handle webhooks; onboarding takes days to weeks. AgentCard is aimed at individual developers and takes under 10 minutes to set up: one npm install, one login, and one MCP config block. AgentCard exposes its functionality through more than 20 MCP tools that any compatible agent can call natively; Stripe Issuing is REST-only and requires custom MCP wrappers before an agent can use it. Per-card spend limits on AgentCard top out at $500 on the Basic plan, whereas Stripe Issuing supports configurable limits up to thousands of dollars per card, making it more suitable for high-value procurement workflows. AgentCard requires human approval before each card is minted, which is a safety feature but adds latency; Stripe Issuing lets developers set authorization rules programmatically with no per-card human gate. For individual developers and small teams who want to get an agent spending online today with zero compliance overhead, AgentCard is the faster path. For companies building a card program into a production product at scale with custom spending rules, Stripe Issuing offers more flexibility despite the higher setup cost.