One door for every model your app calls.

What it is

The AskShiva AI Gateway is the single place where AskShiva’s apps reach AI models. It holds the provider keys, so an app needs only its own gateway key. It speaks the OpenAI HTTP format, so any OpenAI SDK works by changing the base URL and the key. It records the spend of every call against the key that made it.

It was built for our own products first. AskShiva’s apps now send their model calls through it, and each keeps its own direct path to a provider as a fallback. Outside teams can use it by invitation only.

What changes in your app

Three values. Everything else in your OpenAI-style code stays as it is.

- base_url = "https://api.provider.example/v1"
+ base_url = os.environ["ASKSHIVA_BASE_URL"]   # from your invitation
- api_key  = os.environ["PROVIDER_KEY"]
+ api_key  = os.environ["ASKSHIVA_API_KEY"]    # your app's virtual key
- model    = "some-provider-model-id"
+ model    = "askshiva-smart"                  # a role, not a model

Then follow the quickstart for a first call in curl, Python and JavaScript.

Concepts

• Model aliases

  Ask for a role: fast, smart, reason, tool, vision, indic, embed, image. The model behind a role can change without an app release.

• Virtual keys

  One key per app, with a 30-day budget, rate limits and a list of the models it may call. Provider keys stay inside the gateway.

• Data classes

  Public, personal and children’s data each have their own aliases, and a key can only call the aliases of its class.

• Fallbacks

  Most aliases have an ordered chain of models. A timeout moves on at once; children’s aliases never fall back to anything.

• Endpoints and errors

  Chat, embeddings, images, speech, rerank and a model list, plus the error codes your app should handle.

• Policy layer

  In development Signed short-lived product tokens and named tasks with cost ceilings, in front of the gateway.

Honest notes

• Early access, by invitation. Not for sale. There is no public sign-up, no self-serve key and no service-level agreement.
• Where it runs. On rented servers in the USA. The model providers behind the aliases process requests under their own terms. We do not publish which provider sits behind each alias, because that changes.
• What is recorded. The spend ledger records the key, the model, token counts and cost of each call. It is configured not to store prompts or responses.
• Outputs can be wrong. A gateway routes requests; it does not check answers. Validate model output in your app before you rely on it.