Get started
How it works
How a request travels through Roteo, where the inference comes from, and what Roteo never does.
The request path
- You send a request to one base URL with your
sk_key. - Roteo routes it. The gateway sits in front of multiple model providers and picks an efficient path. Where an eligible alternative is available, it may retry or reroute a failed request; failover depends on model availability, request type and limits, and is not guaranteed.
- The model runs and its output comes back to you.
- You pay for that request. The cost settles onchain from your own wallet over x402, capped by your balance. You only hold stablecoins; Roteo covers the network gas.
Where the inference comes from
Every request is served by the same frontier models you would reach going direct. Roteo may source inference capacity at volume-based provider rates and aims to price requests competitively, and onchain settlement leaves out card processors and subscription margins. Listed prices may change by model, provider and network conditions.
What Roteo never does
- Store your content. Prompts and responses are never kept, whether they come through the API, the web chat or the Station. The web chat stays in your browser, and Station images go straight to it; the Station keeps your runs and their files in your browser too. Only videos, and API images you ask for as a link, stay on our servers, and only so you can download them, then they are deleted within about 48 hours, a Station video as soon as your browser has saved it. Files you upload as a reference (videos, and images and audio through the API) are deleted as soon as the video they were used for is made, and never kept more than about an hour. We keep only what it takes to run your account and bill each call. The one opt-in exception is Seedance 2.5 draft mode: you approve it, and our video provider then holds that draft's prompt and inputs so it can be finished in 1080p, and is instructed to delete them after 48 hours. See Privacy.
- Pass on who you are. Every model is anonymized: each request reaches the model provider from Roteo, never with your name, email, wallet, IP address or account. The provider still reads the prompt and any files you send, because it needs them to run the model, so anything in them that identifies you goes with them.
- Hold your funds. Your wallet is self-custodied and only you control it. See Pay-per-call billing.
- Rewrite the output. Model-generated content is never editorially rewritten. Unsupported request parameters may be normalized before inference, and the OpenAI-compatible endpoints translate the response format; provider-specific processing and API behavior still apply.