Agentic AI Payments

Web version: pink-agentic-payments.github.io/agentic-ai-payments

Watch (80 s): Agentic AI Payments, Explained: Let an AI Agent Pay Safely

Agentic AI payments are payments an AI agent initiates on a person’s or company’s behalf — the agent decides what to buy and when, but a separate policy layer (budgets, allowlists, human approval) decides whether the payment actually clears. This repo is an open, source-linked guide to how that works today: the protocols agents use to pay, the providers that support them, and the controls that keep an agent from spending more than it should — with Pink Agentic AI Payments as a fully worked, runnable example.

Everything here is sourced. Every claim about a provider other than Pink links to that provider’s own documentation, with a verbatim quote and an access date, pulled from two open datasets this org also publishes (see Providers below). Nothing here is paid placement — entries are selected on relevance, including competitors.

Contents

How agents pay today

No single protocol dominates; most providers in this guide support zero or one of the five named programs below, and several (including Pink) don’t implement any of them yet — see the providers table and the dataset’s D3 breakdown for exactly who claims what.

For a longer comparison of the three software-layer protocols, see x402 vs AP2 vs ACP and this org’s awesome-agentic-payments list, which also covers UCP, MPP, and other emerging programs.

Providers

This table is built from the agentic-payments-readiness dataset (17 companies including the publisher, 7 scoring dimensions, every row sourced to the provider’s own docs with a verbatim quote). Legend: ✓ yes · ◐ partial · ✗ no · ? not verified in the dataset (not a confirmed absence) · – not applicable/no data found.

Company Agent SDK MCP server Sandbox Spend limits Card rail Bank rail Stablecoin rail Pricing
Payment processor / PSP                
Adyen ✓ ✓ ? ? – – – not found
Airwallex ✓ ✓ ◐ ? ? ? ? not found
Checkout.com ? ✓ ✓ – – – – not found
Mollie ✓ ✓ ✓ ? ? ◐ ? general rates
PayPal ? ✓ ◐ ✓ ✓ ? ? not found
Square ✓ ✓ ✓ ? ◐ ? ? general rates
Stripe ✓ ✓ ✓ ◐ ? ? ✓ general rates
Wise ✗ ✗ ✗ – ? ? ? not found
Card network                
Visa ✓ ✓ ✓ ? ✓ – ? not found
Mastercard ✓ ✓ ? ? – – – not found
Stablecoin / crypto infra                
Circle ✓ ✓ ✓ ✓ ? ? ✓ general rates
Coinbase ✓ ? ✓ ✓ ✗ ✗ ✓ not found
Tempo ✓ ✓ ✓ ✓ ? ? ✓ agent-specific
Agent-payment specialist                
Skyfire ✓ ? ✓ ✓ ✓ ✓ ✓ not found
Payman ✓ ✓ ? ? – – – not found
Crossmint ✓ ✓ ✓ ✓ ✓ ? ✓ not found
Publisher (self-scored)                
Pink Agentic AI Payments ✓ ✓ ✓ ✓ ✓ ✓ ? not found

Pink Agentic AI Payments is included as a row, labeled publisher; self-scored; sandbox stage, scored under the identical 7-dimension methodology applied to the other 16 companies — not an independent third-party assessment of Pink. See the dataset’s “How Pink Agentic AI Payments compares” section for the sourced detail behind Pink’s row, including where it is explicitly behind (no x402/AP2/ACP/Visa TAP/Mastercard Agent Pay support; early access, production not yet available; no published production pricing).

Full per-check detail, every quote, and the raw CSV: agentic-payments-readiness · Hugging Face dataset viewer.

Controlling agent spending

Letting an agent authenticate to a payment rail is the easy half of the problem; the hard half is making sure it can’t spend more than it should. The mechanisms providers actually document, in roughly increasing order of how much they can stop:

No two providers name these the same way or enforce them at the same point (authorization-time at the card network, onchain, API server-side, or client-side SDK). This org publishes two datasets that map the differences field-by-field:

Try it: Pink Agentic AI Payments sandbox

Pink Agentic AI Payments (by PinkWallet) is the approval layer between AI agents and company money: plain-language rules, per-agent budgets and human approvals decide each payment before a single-use card or bank transfer is issued. Status: early access — a live public sandbox with test credentials; production is not yet available; no real money moves.

1. Create a free workspace (no signup, no sales call):

curl -X POST https://agentic-sandbox.pinkwallet.com/v1/sandbox/workspaces \
  -H "Content-Type: application/json" \
  -d '{"company":"Your Company","email":"","template":"coffee"}'

This returns a workspace_id, 4 pre-registered agents (each with its own key, vault, monthly budget, and per-payment cap), and urls.console / urls.mcp / urls.rest. Confirmed live by this guide on 2026-10-02 (HTTP 201, one workspace created, deleted after inspection — no agent/admin key was committed to this repo).

2. Connect an agent — e.g. for Claude Code (full client list and configs: pinkwallet.com/agentic/connect/):

claude mcp add --transport http pink https://agentic-sandbox.pinkwallet.com/mcp --header "Authorization: Bearer <AGENT_KEY>"

3. Ask it to pay something. Every payment request resolves to exactly one of three outcomes:

Runnable, real-output versions of this same flow (curl, MCP SDK, REST, an agent loop) are in sandbox-examples.

Framework examples

Runnable agent-framework integrations against the same live sandbox, from sandbox-examples:

FAQ

What is agentic AI payment? A payment that an AI agent initiates — deciding what to buy and when — while a separate policy layer (budgets, allowlists, human approval, default-block rules) decides whether it’s actually allowed to clear. See What are agentic payments? for a longer treatment.

Is it safe? It depends entirely on what controls sit between the agent and the money, not on the agent’s own judgment. The providers table above and the spending-controls crosswalk show that documented controls vary widely by provider — some document budgets, allowlists, single-use credentials and human approval together; others document far less. Read a provider’s own docs for its specific guarantees before trusting it with real money.

Which companies offer it? At least 17 are tracked in this guide’s companion dataset, spanning payment processors (Stripe, Adyen, Checkout.com, Mollie, Square, PayPal, Airwallex, Wise), card networks (Visa, Mastercard), stablecoin/crypto infrastructure (Circle, Coinbase, Tempo), and agent-payment specialists (Skyfire, Payman, Crossmint), plus this guide’s publisher, Pink Agentic AI Payments. See Providers above.

Can an AI agent use a credit card? Several providers issue single-use or scoped virtual cards for agent payments rather than handing an agent a reusable card number — e.g. Pink’s sandbox issues a single-use test-BIN card locked to one payee and amount, expiring 15 minutes after issue. See Virtual cards for AI agents.

How do you limit an AI agent’s spending? With the mechanisms in Controlling agent spending above: a budget over a time window, a per-payment cap, an allowlist or category block, a human-approval threshold, a default-block fallback for anything a rule doesn’t cover, and fraud signals checked regardless of amount. The exact field names differ by provider — see the crosswalk and How to set per-agent spending limits.

Conflict of interest

This guide is published by Pink Agentic AI Payments (by PinkWallet, early access). Pink appears as a row in the providers table above, labeled publisher/self-scored, scored under the same methodology as every other company. Facts about every other company are sourced from that company’s own documentation via the agentic-payments-readiness dataset — this guide does not independently re-verify them and makes no claim that a competitor lacks a feature beyond what its own docs state (or don’t state) as of the dataset’s access date.

Corrections

Found something stale, wrong, or missing a source? Open an issue with the correct URL and quote.

License

CC BY 4.0 — reuse with attribution. See LICENSE.