An agent can run this end to end without a person.

MCP, not scraping your own docs

Point any MCP client at the streamable HTTP endpoint and it gets typed tools: create a batch, read a batch, check the balance, list sample captures, and create a checkout session. Input schemas describe every render option, so a model does not have to guess what full_page means.

The server card, agent card and agent skill are all published under /.well-known, so a client that discovers the origin can wire itself up without being told where anything is.

POST https://url2image.com/mcp
{"jsonrpc":"2.0","id":1,"method":"tools/call",
 "params":{"name":"create_screenshot_batch",
           "arguments":{"urls":["https://stripe.com"],"format":"png"}}}

Identity without a signup form

An agent needs a stable identity so its credits and its batches survive a restart. There are two ways to get one and neither involves an email confirmation loop: register anonymous OAuth client credentials at /oauth/register, or mint an API key at /api/v1/keys. Both produce a bearer token that scopes every batch to one account.

Paying without a card

Two rails, and the choice is about who is in the loop. When a person is, the checkout tool returns a hosted Stripe URL and the agent hands it over. Card details never reach the agent, this service, or the model context.

When no person is, x402 handles it. Call a priced endpoint, read the HTTP 402 body, retry with a signed X-PAYMENT header, and a facilitator verifies and settles the transfer. The terms are published at /.well-known/x402 and priced to match the entry credit rate, so paying per call is never worse than buying a pack.

What the agent should tell the user

The agent skill spells this out: ask before creating a checkout session, never request card details, and report how many screenshots a batch will consume before it runs. Those are instructions the model reads, not just guidance in a README.