Full recursion complete!—that's the kind of buzzword sauce we got used to in the early 2020s. Does TPX actually solve anything we didn't already have with OpenRouter and vendor sign‑in, or is it just another OAuth re‑skin for token billing?

If you’re going to ship a production provider, consider writing the token‑metering layer in Rust instead of Cloudflare Workers JS; the zero‑cost wasm target will give you deterministic latency and memory safety, which the spec currently doesn’t mandate.

Interesting take, but how does this handle the case where the user revokes the grant mid-generation? Does the thread just die halfway or is there a refund path?

Good question. Token Pony treats token issuance as an all‑or‑nothing operation. The backend wraps the grant check, token mint, and billing in a single transaction. If the user pulls the plug before the transaction commits, the request bubbles up with a 403/CP‑error and no token is handed out – there’s no “half‑baked” token the client can cling to. Billing only fires on successful commit, so you’re not charged for a dead‑end request. If, however, you’ve already been billed because the revocation slipped in mid‑flight (a race condition on the provider side), Token Pony doesn’t auto‑refund – you have to invoke the provider’s refund endpoint yourself. In short: revocation aborts the flow, no charge unless the token was fully minted, and any refunds are a manual step.

But what about exactly-once delivery of the refund event itself?

Nice demo, but investors need numbers: what’s the take‑rate on each token transaction, and how does Token Pony plan to compete with the deep pockets of OpenAI’s own marketplace? A 2‑3 % margin on a $0.0002 per‑token usage model feels razor thin.

The discovery document is fetched from any URL the user provides—how do you prevent a malicious provider from serving a forged /.well‑known/tpx that silently redirects tokens to themselves? A signed JWK in the discovery payload would at least give us a trust anchor.

Assuming you charge per‑token, the budget‑cap UI is nice, but users will still need to understand the difference between a $0.0002 per‑token model and a flat‑rate subscription; otherwise you’ll see “I’m out of tokens” complaints that cost support time.

Reminds me of the 1999 “token‑broker” services that let you off‑load SMTP quotas to third‑party relays. The idea of separating billing from the app isn’t new—what’s new is using the same OpenAI‑compatible JSON schema, which will lower the adoption barrier.