Questions and answers
Short answers to what people trip over most, each with the page where it is worked through in full.
Connecting
Which base URL do I use?
For an OpenAI-compatible client, https://api.kumorouter.com/v1. For a client written against Anthropic's native API, https://api.kumorouter.com — the bare origin, with no path: such a client composes the path itself, and a base URL carrying a prefix sends it somewhere nothing is served.
How the header is written → Your first call →
Which models can I name?
The ones the gateway itself lists at https://api.kumorouter.com/v1/models, which are the ones in the price list. This documentation names no model on purpose: the examples carry a placeholder, and the current list belongs to the catalog rather than to a text.
The catalog and how to read it →
Why is my request refused without `max_tokens`?
An output ceiling is required. On the Chat Completions surface that is max_tokens or max_completion_tokens; on the Messages surface it is max_tokens, and it has no default. The reason is money: a billable dimension with no finite ceiling cannot be reserved for before the upstream call.
Do I have to change my code for Kumo?
No: the surfaces answer in their own protocols rather than in a translation of somebody else's. What changes is the base URL and the key.
Keys
I lost a key's secret. How do I see it?
You do not. The secret is shown once, at creation; after that the platform holds only its hash, and nothing can read it back — not a screen, not an operation, not support. The cure is a new key.
Can a key be renamed?
No. The name is fixed at creation, and the key has no rename operation. What does change is the expiry, the project and the switch.
How do I check a key is alive without spending a call?
The identity echo answers about the key itself rather than with a model's output: its name, its display fragments, what it spends from, and its expiry. Right here on this site, from a browser.
Why does `401` not say what exactly is wrong?
Because missing, malformed, unknown, revoked, switched-off and expired keys are answered identically — on purpose. Telling them apart would make the gateway an oracle on key material: with a list of guesses you could learn which of them exist.
Refusals
`429` or `503` — what is the difference?
429 is a ceiling you ended up above, and time cures it. 503 means nobody upstream can serve this call right now. Neither is cured by an immediate retry.
My stream ended early — is the answer complete?
No. Treat a stream that ended without its terminal frame as a failed call rather than a short one: a failure after the stream started simply ends it, and nothing in the bytes you received will say the answer was truncated.
How to read a cut stream → How a refusal behaves →
How do I raise my limits?
By talking to support. No ceiling is editable in the console, and no request member asks for a higher one.
What a key is issued with → How to write →
Money
Wallet or package — which do I pick?
The wallet counts in dollars and is topped up by any amount; a package is a prepaid token limit, one per account, bought from 30 million tokens up and only topped up after that. Each key chooses what it spends from — the package or the balance — and the choice can be changed any time; a key on the package can carry on from the balance once the limit runs out.
Compared side by side → How the wallet counts →
Is there a minimum top-up?
Yes: $2, or the rouble equivalent at the checkout's own exchange rate the moment you pay. A package, meanwhile, is bought by a token volume — from 30 million up.
Can I try it without money?
No. A wallet with nothing on it refuses every call of the keys spending from the balance with a 402: issuing a key is free, making a call is not.
When there is nothing to pay with →
Where is the price list?
On the pricing page. There are no prices in the text of this documentation on purpose: the price list publishes them, and a page that repeated a figure would go stale in silence.
Why is spend shown as two numbers rather than one?
Because they are two different units: the wallet's dollars and the package's tokens. The platform publishes no rate between them, so their sum would be a figure with no unit.
Data and tools
Does Kumo store my prompts?
A call is recorded by its shape and not by its content: no prompt, no answer, no headers, no client address is in the history — they are not stored at all. The other side of that is the same: you cannot reconstruct from the log what a request was about either.
How long is the money history kept?
Money movements for five years. One request reads a window of at most 366 days, so an older period is taken by moving the window back.
How do I connect Claude Code, Cursor or another tool?
With a ready-made recipe: the platform publishes them itself, together with a prompt that will configure the tool for you.
Every tool → Claude Code → Cursor →
I changed my password and was signed out everywhere. Is that right?
Yes. Changing the password ends every session, including the one it was changed from: if a password is changed because the old one leaked, somebody else's open tab has to stop working at once.