Everything on the other side of that connection
Grouped by the question each one answers, not by vendor. A vault is not an alternative to a wallet and a store is not an alternative to a card — reading them as one flat list is what makes this space feel impossible.
What decides to spend?
Something has to read the intent and choose to act. Every layer below plugs into it.
Reads your list, works out what needs buying, and spends inside the limits you set.
It acts on what you write down, so the list is the interface and the limits are the contract.
Where does the money come from?
Six rails, and they are not six points on a line. They differ in who holds the money, how a spend is bounded, and how much you set up once.
Card and crypto in one agent wallet, with a real account number for bank deposits and automatic top-ups across chains when a purchase settles somewhere you are not funded.
You verify your identity and set up the card once; from then on the agent spends within your limits, and the cross-chain top-up runs on its own when a checkout needs a chain you are not funded on.
Save one card with no ID check; the agent charges it off-session afterwards. The widest acceptance on this page and the least setup.
Verified end to end in Stripe test mode — an off-session charge with nobody present, a replayed idempotency key returning one payment rather than two, and 3DS surfacing as authentication_required. The trust you extend is “charge my card”, so this is the rail to give the tightest caps.
Your real Visa or Mastercard, but every purchase gets a single-use number capped by limits you set — built for scoped recurring spend like $50 a month at one merchant.
Runs against Crossmint staging, so a green dot here means the card path works in staging. A cap is enforced at the network, which is why an over-charge simply declines rather than needing a refund.
A crypto wallet the agent funds and spends directly, with no approval per purchase.
Deploys on Base Sepolia, which proves the whole flow end to end at zero risk. Purchases that settle on Solana route through your Sumday wallet instead.
Connect a PayBox vault you already own over OAuth. Sumday reads its balances and can ask it to release a credential you stored — you approve with a passkey, and the keys stay inside MoonPay's enclave.
Connect, read balances and release an approved secret today; MoonPay puts payment cards in Phase 2, and the card API is already built here waiting for them. Signing, swaps and x402 complete inside PayBox's in-chat window, which an MCP host draws — so those belong to an MCP client rather than a web app.
Connect a wallet you already control. The agent proposes each spend and you sign it yourself.
Your signature is the approval step, which makes this the rail for spends you want to see before they happen.
Where can it buy?
Finding something and buying it are different jobs. These reach real merchants, and each pairs with a rail underneath to complete a purchase.
Places real orders end to end — no Amazon account, no password, no saved cart. Billed per search and per order over x402.
Settles in USDC on Solana and bills inside the HTTP request itself over x402, so it pairs with a wallet rather than a checkout page. Your Sumday wallet covers both, and bridges automatically if you are funded on another chain.
First-party product search across Shopify merchants, straight from Shopify's own catalog API rather than a scraper.
Search is free and needs no rail; checkout completes through whichever rail you connected.
Finds and vets vendors, tools and service providers before your agent commits to one.
A discovery step that runs inside the agent's skill, so it costs nothing and needs no rail connected.
Maps, enrichment, data lookups — paid per request as the agent uses them, with no subscription per provider.
Billing happens inside the request over x402, so it pairs with a wallet that can answer a 402 rather than with a saved card.
141 services and 1,271 paid endpoints an agent can call and settle for in the same request — Anthropic, OpenAI, Gemini, Perplexity, Exa, Firecrawl, Tavily, Browserbase, Google Maps and more.
A different protocol from x402, and a broader one: payment-method agnostic, so cards through Stripe and stablecoins through Tempo both work, with receipts and idempotency built into the spec. Visa has published a card-native extension. We ran it end to end on testnet in under two minutes; reaching production services needs a funded Tempo mainnet wallet.
Checkout across participating merchants rather than paying us directly — a different product from Link, and the one most people mean by “Stripe for agents”.
Agent-side access is in private preview; we are on the waitlist and will wire it up when it opens.
Place a Polymarket order straight from a line on your list.
Built and waiting on two upstream things: Polymarket's terms currently scope trading away from U.S. persons, and a Sponge deploy-tx fix is open. It turns on when both land.
What holds the credentials?
You never pick this. It arrives underneath whatever rail you chose, which is exactly why it is worth seeing.
Card tokenization and PCI-scope infrastructure. It holds credentials, not money.
It arrives as part of the rails above rather than as a choice beside them — it is already in this tree underneath Crossmint, and documented underneath PayBox. Reaching it directly would mean running your own card vault and the PCI scope that comes with it.
How does the other side know it is you?
The part nobody demos. An agent that can identify itself gets treated as a known caller rather than as anonymous traffic.
Cryptographically signs your agent's requests so a Cloudflare-protected site can recognise it as a known caller and let it through on purpose.
Built and waiting on one step of ours: publishing our signing keys.
One sign-in tying you to your agent's wallet, so a rail you connect is scoped to you and not to the app.
Every rail you connect hangs off your session, which is what makes a connected provider yours rather than shared.
Can my agent do this?
The useful question is never which rail — it is a sentence about something you want done. These are the ones we get asked most, answered with the actual combination, what each piece contributes, and where we checked.
“Book me a flight to Lisbon under $600.”
Works, with one piece addedYes — Expedia's MCP server finds and holds the itinerary, and a rail you already connected pays for it.
Expedia runs a hosted MCP endpoint at expedia.com/mcp exposing flights, hotels, cars and packages to agents, and Expedia Group publishes an open-source travel-recommendations MCP server alongside it. The agent searches and assembles the itinerary there; the money is a separate concern, so the charge lands on whichever rail you connected. Pairing it with a single-use number capped at the fare means the booking cannot be charged for more than you approved.
Checked against Expedia Group developer hub· 3 August 2026“My company agent needs to pay this invoice before it's late.”
Works todayYes — the amount is known up front, which is exactly the shape a single-use capped card is for.
A fresh number capped at the invoice total means an over-charge declines at the network rather than becoming a refund conversation, and the card is spent once. For a bill that repeats, the same rail takes a per-merchant monthly cap instead — “$400 a month at this supplier” — so a duplicate invoice cannot quietly go through twice.
“Order the replacement filter before we run out.”
Works todayYes, end to end today — no Amazon account and no saved cart involved.
Purch bills per search and per order inside the HTTP request over x402 and settles in USDC on Solana, so the payment happens in the request rather than at a checkout page. If your wallet is funded on another chain, the shortfall bridges itself rather than failing — which is the part that makes this feel like one step instead of three.
“Use a maps API just for this one trip, without a subscription.”
Works todayYes — per-call billing is the native case for a wallet that can answer a 402.
The agent pays for each call as it makes it, so a one-off lookup costs a fraction of a cent instead of a monthly plan per provider. This is the rail that makes “try an API for one task” a reasonable thing for an agent to do on its own.
“Send my flatmate $40 on Venmo.”
Different routeVenmo is not the route — its Developer and Payouts APIs are closed to new integrations — but person-to-person is very much doable.
Businesses granted Venmo Developer or Payouts access years ago keep it, but there is no path to it for a new integration, so an agent-to-person Venmo send is not something you can build today. Venmo as a way to pay a business is a different product, and that one is available through PayPal and Braintree.
For programmatic person-to-person: Dwolla covers domestic US bank rails including Same Day ACH, RTP and FedNow; Wise is the cleanest cross-border option; PayPal's Hyperwallet handles global marketplace payouts. Or use the rail already connected here — a USDC transfer from your Sumday wallet settles to an address in seconds, if the person on the other end has a wallet.
Checked against Venmo developer docs· 3 August 2026“Check this is the right part for my model before you buy it.”
Works, with one piece addedThe buying half is solved; this is the half we are wiring up next.
The step before a purchase is almost always a question — is this the right part, is this seller real, has the price moved. Retrieval that hands back the content of a page rather than ten links is what closes it, which is why both of these are on the bench.