Cardano turns HTTP's reserved 402 code into a per-call payment prompt for AI agents

Per-call payment is a cost question first. The settlement charge has to be small against the price of the call, and a weather lookup or one inference request is a small ticket. The cryptobriefing account does not give fee levels or settlement times, and it does not name an API operator already answering requests with a 402 ^[16]. So what landed around September 9 is distribution ^[1].

Distribution still matters. The code now sits in the repository the x402 Foundation governs under the Linux Foundation umbrella ^[9], and a developer who picks up the SDK gets Cardano client, server and facilitator components along with working examples to copy as templates ^[1][10]. Similar work existed before the merge: Masumi Network and other third-party projects built x402-style capability on Cardano as independent implementations, outside the canonical codebase ^[8].

The design borrows HTTP status code 402, "Payment Required", which was reserved for future use when the protocol was written ^[3]. A server answers a request with 402, the client constructs a signed Cardano transaction and sends it along with the request, and the server verifies the payment before delivering ^[4]. Under the extended UTxO model the transaction is user-signed and settles directly on chain, so the payment and the proof of payment are the same cryptographic step ^[5]. For the seller that removes the billing portal and the monthly invoice ^[12]. For the agent it removes the OAuth token that expires mid-task and the webhook callback confirming the charge went through ^[11].

Four assets can settle a call: ADA plus the Cardano-native stablecoins USDM, DJED and iUSD ^[14][7]. The stablecoin support is what makes tiny tickets usable, since an agent buying data thousands of times should not carry token price exposure on each purchase ^[7]. It also bounds what this does for ADA. If the agent pays in USDM and the transaction settles on Cardano, the chain gets the activity, the buyer avoids the price risk ^[13], and the unit the agent spends is USDM.

I think the binding constraint sits with sellers. An API operator that answers with 402 gives up predictable subscription billing for per-call revenue from counterparties that never open an account with it ^[2][4], and the merge does nothing to make that trade attractive. Two other readings are available. One is that agent traffic never fits a monthly plan, so the first operators to price per request capture demand the subscription shops cannot serve at all ^[11]. The other is that agents end up paying through custodial intermediaries and this path stays a developer convenience. Three networks are wired for it now, mainnet plus the preprod and preview testnets ^[6][15].