Tools, APIs, data, pages, actions — if it wants paying, your agent can reach it through one call. It never has to learn a payment protocol, hold a wallet, or stop and ask a human for a card.
A page that wants money says so in its own response. We read that, pick the right way to pay it, settle, and return what you asked for.
One call. No protocol names, no chain, no wallet setup.
Its response identifies the network. Nothing has to be known in advance, which is why coverage grows on its own.
A per-request payment, a card, an account, a licence — whichever that network takes. How we settle is our problem, not yours.
What it cost, which network it came through, and whether anything was charged at all.
// ask POST /v1/fetch { "url": "https://api.example.com/search", "maxPrice": 0.05 } // get { "content": "{ results: [...] }", "cost_usd": 0.012, "via": "per-request", "paid": true } // or, if the door stays shut { "blocked": true, "cost_usd": 0 }
An agent that already holds funds needs nothing from us. A developer who would rather not think about wallets gets an account.
If your agent can already sign a payment, it just calls the endpoint and settles inline. Nothing to sign up for.
Top up once and hand your agent a key. We keep the ledger and show you what every request cost.
Ordered by what matters, labelled by what is true today. Each is a different kind of door, and each is reached however that network wants to be paid.
Search, scrapers, inference, market data — anything that answers with a price per call. We can already quote them; paying them is next.
Ask for something to be done on a site — filled in, booked, submitted — not only fetched from it.
Nobody wanted paying, so nobody was paid. You are not charged for a page that was free.
Ask what a URL costs — and whether it can be bought at all — before committing to pay for it. Most things that quote a price will not actually sell to you.
Publisher material bought properly rather than scraped, through a registered buyer identity.
Sites that refuse anonymous traffic and want a known buyer at the edge.
Services that will never take crypto, reached with a card behind the scenes.
Add it to the assistant you already use and every agent you run inherits it.
You pay for what came back. Nothing else.
One entry in a config file and every agent you run in Claude can pay for things on the open web. Nothing to install — the server runs here, not on your machine.
Create a key and set Max per call — a hard ceiling on any single purchase, whatever the balance says. A virtual card limit.
No repository, no Python, no paths. Restart, and the tools appear. Claude checks a price before it buys and tells you first.
// claude_desktop_config.json { "mcpServers": { "agentgateway": { "url": "https://safeconduct.vercel.app/mcp/", "headers": { "Authorization": "Bearer ag_..." } } } }
Then just ask: “find market data on Polymarket, tell me what it costs, and buy it if it is under a cent.”
Sign in with GitHub, get $5 of credit, and make a call. No card to start.
# ask for something that wants paying curl -X POST https://agentgateway.com/v1/fetch \ -H "Authorization: Bearer $AGENTGATEWAY_KEY" \ -H "Content-Type: application/json" \ -d '{"url": "https://example.com/article", "maxPrice": 0.05}'