A metered endpoint the agent buys its tokens through

Credits stopped being a number on a page: the agent now reaches a model through an endpoint of ours that measures and can refuse, rather than beside one that can only observe.

POST /v1/chat/completions, scoped by a book key. Not an agent — the loop runs in the customer’s own Hermes, and each turn arrives as its own request. Sitting on the inference path rather than beside it is the whole point: a meter beside it is advisory, and a meter on it can refuse.

Three settings decide whether the credit ladder means anything, and every one of them is silent when it is wrong, so each is asserted on the wire rather than trusted. Reasoning effort defaults to a level that bills at the output rate — measured at 92 completion tokens pinned against 160 unpinned on the same short question. Provider price is pinned, because the same model was available across six endpoints spanning 7.2x on input price; unpinned, the margin is somebody else’s routing decision. And usage reporting is requested explicitly, because relying on it arriving by default is how a silent zero reaches the ledger.