Skip to main content

Govern

Models and the AI gateway

The AI gateway is the single endpoint your apps call instead of calling model providers directly. This page covers the model catalogue, fallback and the API.

The AI gateway is the single endpoint your apps and developers call instead of calling model providers directly. It checks each request against its key, workload, policy and budget, then sends it to the model and records the cost.

Providers and the model catalogue

An administrator connects each model provider once for your whole organisation, so your developers never handle provider credentials. Supported providers include OpenAI, Anthropic (with an option to keep inference in the US), DeepSeek and OpenRouter. Each provider's models are added to the organisation catalogue with their prices; only catalogue models can ever be called.

Model groups, defaults and fallback

Group models by purpose (for example coding, chat, low-cost) and give each group a default. Policies can then refer to groups instead of individual models.

Fallback keeps work moving when a provider is down: if the upstream fails before a response starts, the request can run on the next model in the fallback order. Fallback never bypasses a policy denial, a permission error, invalid input or an empty budget, and Novie never replays a request once streaming has started. Every attempt, the model that actually answered, and its cost are recorded once.

Before you remove a model, Novie shows which workloads and policies depend on it, and blocks changes that would leave a workload with nothing it can run.

Calling the gateway

The gateway speaks the APIs your code already uses. There are two entry points, one for each kind of key.

Run workloads: `/v1` with an Application Key

Base URL https://gateway.api.novamindlabs.ai/v1. The Application Key alone identifies the workload, so any app that lets you set a base URL and API key works unchanged.

EndpointAPI
GET /v1/modelsModels this workload may use
POST /v1/chat/completionsOpenAI Chat Completions
POST /v1/responsesOpenAI Responses
POST /v1/embeddingsOpenAI Embeddings
POST /v1/messagesAnthropic Messages

Build workloads: personal Novie Keys

A Novie Key belongs to a person and charges the Build workload they have selected. Tools that can send extra headers can name the workload on every request instead, through /project/v1:

HTTP
POST https://gateway.api.novamindlabs.ai/project/v1/chat/completions
Authorization: Bearer <personal Novie Key>
X-Novie-Contract-Version: 1
X-Novie-Workload-Id: <workload UUID>

On /project/v1 the workload is chosen only by the header; a workload or project ID in the body or query string is rejected, so a call can never be charged to the wrong workload by accident. Novie strips both headers before forwarding the request to the provider. GET /project/v1/models lists the models that key may use for that workload, and GET /project/v1/capabilities returns the contract version without a key.

Errors

Errors use the OpenAI error shape plus request_id and retryable. Every response carries X-Request-Id.

StatusCode (examples)Meaning
400context_length_exceeded, provider_content_blockedFix the request: it is too long, or the provider flagged the input.
401The key is missing, invalid, revoked or expired.
402The workload has no budget left, or no price is available for the model.
403workload_access_denied, workload_unavailableThe workload is not approved or not active, or (Novie Keys) you are not a member.
404Unknown endpoint, or model not available to this workload.
426Unsupported X-Novie-Contract-Version.
429rate_limitedThe provider is rate-limiting; honour Retry-After.
502upstream_unavailable, provider_credential_rejected, provider_access_deniedThe provider failed or refused. Your key was accepted.

Novie does not automatically resend a generation after a failed or unknown outcome; your application decides whether to retry.