AgentCard Virtual Cards Skill
Give your AI agent virtual Visa cards via MCP tools: fund a wallet with Apple Pay/Google Pay, issue single- or multi-use cards, and shop and check out at merchants like DoorDash.
Evidence shows the skill is markdown-only instructions, no code execution; it embeds explicit safety rules (never show PAN/CVV unprompted, confirm before closing cards, ordering, withdrawals, card creation), relays payment links verbatim, and uses approval flows (202) and reversible operations (pause_card). Deductions: it drives real-money actions (cards, purchases, withdrawals, KYC) with limited rollback (close_card is irreversible); the remote MCP server is a black box that cannot be statically audited; publisher is unverified and attribution rests on a self-declared checksum mechanism. Not full marks.
Workflows are detailed and self-consistent: error-code handling (verification, KYC, funding gates, 202 approvals) is thorough, with a CLI fallback and known non-interactive-shell defects disclosed. Deductions: static review cannot reproduce execution; MCP tool signatures and references/setup.md are absent from the evidence; behavior of 44 tools is inferred from descriptions only; failure-feedback quality on edge inputs is unverified.
Trigger scenarios are clear in the description (create/manage virtual cards, fund wallet, shop on behalf, set limits); non-fit boundaries (no sandbox mode, limited supported merchants, regional availability caveats) are stated. Deductions: core function depends entirely on mcp.agentcard.sh and other overseas services with undisclosed mainland-China reachability risk; no Chinese-language support; thin evidence on merchant/region boundaries.
Documentation is well structured: frontmatter with version 1.5.0, license, compatibility; layered workflows with separate safety and error-handling sections; SECURITY.md, canonical checksum note, and install guides. Deductions: no changelog; maintenance responsibility and update path are only loosely pointed at a monorepo mirror; references/setup.md missing from evidence leaves hidden assumptions.
The skill claims a complete fund→card→spend→manage loop and adds procedural knowledge beyond bare MCP access (verification flows, error recovery, safety confirmations), giving clear marginal value. Deductions: static review cannot verify output usability; no execution evidence or real usage examples; cost/benefit versus alternatives is not argued.
Auditable primary material exists: MIT LICENSE, SECURITY.md, and a claimed canonical checksum location. Deductions: no test suite or CI covering key paths; the checksum mechanism and monorepo-provenance claims cannot be cross-verified in this static evidence; no third-party execution evidence.
- The skill drives real-money actions (card creation, purchases, withdrawals); users should confirm every transaction amount and destination before approving.
- Core function depends entirely on the overseas service mcp.agentcard.sh; mainland-China network reachability is unverified and may block use.
- close_card is irreversible and withdrawals take 1-3 business days executed by the vendor, so mistaken actions are costly to recover.
- Publisher is unverified; the self-declared checksum attribution mechanism is not independently corroborated.
- There is no sandbox/test mode; every card issued is live.
What does this skill do, and when should you use it?
agent-card-skill is a SKILL.md file that teaches AI agents to drive AgentCard's 44 MCP tools for managing virtual Visa cards. The user tops up a wallet with Apple Pay or Google Pay (funds held as USDC, spent in USD); the agent then creates single-use or multi-use cards, checks balances and transactions, shops and checks out, withdraws, and manages plans. First-time funding and card creation require one-time identity verification (KYC), with full workflows, error handling, and safety rules built in. It works with Claude Code, Cursor, and 70+ other agents, and every card is live, spending real wallet money.
The skill is procedural knowledge that, once loaded, lets the agent call AgentCard MCP tools: fund the wallet (add_funds returns a single-use payment link), complete phone/KYC verification (verify_phone, start_kyc with government ID photo plus browser face scan), create virtual Visa cards (create_card with cents amounts and an ai_labs scope preset that locks cards to AI merchants and earns TOKENBACK), query balances and card details (PAN/CVV gated behind approval, never shown unprompted), pause/resume/resize/close cards, list transactions per card or across all cards, shop via the buy tool in natural language at DoorDash and others (calling get_instructions first for the current guide), withdraw to bank accounts or USDC on Base, manage paid plans and payment methods, and open support tickets.
- Individual developers who want spending controls for an agent: create a fixed-amount single-use card that auto-closes after the first charge, capping exposure.
- Users paying for AI services like OpenAI or Anthropic: create a multi_use or ai_labs AI card restricted to AI merchants, accruing TOKENBACK rewards.
- People who want an agent to order food: place conversational DoorDash orders (plus Good Eggs and Rappi where available) via buy, with checkout only after the user confirms the cart and total.
- Users with subscriptions or recurring charges: issue a multi-use card that can be paused, resumed, or resized at any time.
- Teams managing agent spend from a company balance: card limits are resized server-side, and OAuth-connected apps can be audited with list_connections.
What are this skill's strengths and limitations?
- Full lifecycle coverage — funding, issuance, controls, transactions, shopping, withdrawals, plans, and support — with workflows for all 44 tools.
- Explicit safety rules: never show PAN/CVV unprompted; confirm before closing cards, placing orders, or withdrawing; relay payment links verbatim.
- Detailed error handling covering phone_verification_required, kyc_required, wallet_funding_required, deposit_confirming, and 202 approval states.
- CLI fallback when MCP tools aren't loaded, with an explicit ban on raw curl calls to internal API routes.
- MIT-licensed; the repo is a read-only mirror of the canonical skill, checksummed in agentcard.sh's well-known index.
- No test/sandbox mode — every card on the consumer MCP is live and spends real wallet balance.
- Hard dependency on an external MCP server and an AgentCard account (including KYC); tools don't work offline.
- Withdrawals are executed manually by the AgentCard team, typically taking 1-3 business days.
- Several CLI interactive commands crash in non-interactive shells, limiting agent-side usage.
- README claims compatibility with 70+ agents without per-platform verification; the repo shows no evidence of a test suite.
How do you install this skill?
1) Install the skill: npx skills add tiny-agent-company/agent-card-skill. 2) Install the CLI: npm install -g agent-cards. 3) Connect the MCP server: in Claude Code run agent-cards setup-mcp, or manually claude mcp add --transport http agent-cards https://mcp.agentcard.sh/mcp; for Cursor/Windsurf add an mcpServers entry pointing to https://mcp.agentcard.sh/mcp in .cursor/mcp. or .windsurf/mcp.. 4) Restart your agent session (tools only load after restart). If you have no account, run agent-cards signup. The README also provides a paste-ready multi-step setup prompt for your agent (including email-code login).
How do you use this skill?
After restarting, trigger it in natural language, e.g. 'Create me a $25 virtual card' — the agent reads SKILL.md and calls create_card(amount_cents: 2500). First use triggers OAuth sign-in; first funding/card creation walks through one-time phone verification and KYC (ID photo + face scan, valid for 60 days). For shopping, have the agent call get_instructions first, then use requests like buy('order a caesar salad from Zuni on DoorDash') conversationally. You can also use the CLI directly: agent-cards cards list, agent-cards balance, agent-cards fund --amount 50, etc. Note that interactive commands (cards create, wallet fund, signup, support) crash in non-interactive shells and must be run in your own terminal, not by the agent.
How does this skill compare with similar options?
Compared with just handing an agent API docs, this skill adds procedural knowledge: complete workflows, safety confirmation rules, and error-state handling. The architecture pairs a Skill (knowledge) with an MCP server (tools) — both are required. The source names no other specific competitors.