Terms

What this is

A pay-per-call front for Cloudflare Workers AI models run on the operator's own account: text-to-image, captioning, speech, transcription, translation, summaries, embeddings, chat completions, classification, reading text in images and a language heuristic. Outputs are model outputs, returned as produced; they are not checked for accuracy and may be wrong, and you are responsible for how you use them. Models change: when Cloudflare retires a model the route moves to the nearest replacement and the model field in every response says what ran.

Payments

Payments are made with the x402 protocol in USDC on Base to 0xc91cE6291eDC0713ec753BAFBA002506ffb2b95c and settled by a third-party facilitator (facilitator.payai.network); there are no accounts, subscriptions or refunds, and settlement is final. Prices are the amounts in the 402 challenges; pricing and the manifest are generated from the same table. A request that carries a payment header but is malformed or over a limit is rejected with 400 or 413 before anything is settled.

Failed calls: one free replay

Payment is settled before the model runs, so a model failure happens after your USDC has moved. In that case the call answers 502 with code: "model_failed", nothing further is charged, and the same payment is honoured once more: resend the identical request with the identical X-PAYMENT header to the same route at least 10 seconds later (the record has to propagate) and within 10 minutes, and it runs without a new payment. A payment can be replayed once; a second failure is final. A payment that already bought a successful result cannot be reused (409). Partial results are results: a /batch or /image-set that returns some successful items has succeeded, except that an /image-set with 10 or more failed slots and a /batch with every item failed answer 502 and can be replayed.

Content

Three things are refused with a 400 and a plain reason, before payment: sexual content involving minors, non-consensual intimate imagery, and deepfakes or sexual imagery of identifiable real people. Nothing else is judged; the models keep their own behaviour and may decline on their own.

Limits

Request bodies up to 6 MB; media by URL up to 5 MB (images) and 10 MB (audio), fetched from public https hosts within 8 seconds; per-route text limits are listed on pricing. Unpaid requests to paid routes are limited to 20 per minute per address (429 with Retry-After); paid requests are not limited. Long jobs (/image-set, /batch) can take a minute; keep the connection open.

Data

Prompts, images, audio and outputs are processed in memory and not stored. Stored data: the payment nonce with the route, amount and outcome for 10 minutes (to honour replays and block double use), the nonce for one day (double-spend guard), anonymous per-route daily counters, and a hashed rate-limit counter for 60 seconds. No cookies, no IP addresses, no accounts. Inference runs on Cloudflare Workers AI under Cloudflare's data-usage terms.

Operator

Agent Exchange, one operator, one wallet. All properties are listed at network.agentexchange.work; the payment address above is the only one used. Health: /health; counters: /live.json.