API Preview In Development

Status: the hosted InferaStack Gateway is in development and not yet available to customers. api.inferastack.ai does not accept public traffic today. This page previews the API we are building to — request early access and we will let you know when it opens.

The InferaStack Gateway is an OpenAI-compatible API by design. When it opens, code that already works with OpenAI will only need a new base URL — no SDK to install, no request format to learn.

Request early access

Keys will be provisioned per organisation with a daily budget cap. Contact us to join the early-access list.

The same API is designed to front InferaGrid nodes, recording the execution region and site-envelope state for each request.

What your first request will look like

curl https://api.inferastack.ai/v1/chat/completions \
  -H "Authorization: Bearer $INFERASTACK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "inferastack-chat-smart",
    "messages": [{"role": "user", "content": "Hello from Sydney"}]
  }'

Or with an OpenAI SDK

Set base_url to https://api.inferastack.ai/v1 and pass your InferaStack key:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.inferastack.ai/v1",
    api_key=os.environ["INFERASTACK_API_KEY"],
)

response = client.chat.completions.create(
    model="inferastack-chat-fast",
    messages=[{"role": "user", "content": "Summarise this in one line: ..."}],
)
print(response.choices[0].message.content)

Model aliases

You will call a stable alias and the gateway will route to the right backend. Swapping the underlying model is our config change, not your migration.

Alias	Use for	Routing
`inferastack-chat-fast`	High-volume, latency-sensitive tasks — drafts, classification, extraction	Amazon Bedrock (Sydney gateway)
`inferastack-chat-smart`	Complex reasoning, longer generations	Amazon Bedrock (Sydney gateway)

Region of execution (by design): inferastack-chat-fast/smart aliases use our Sydney (ap-southeast-2) gateway. Workload placement follows data sensitivity: identified data stays in a controlled environment, while residential InferaGrid nodes are designed for encrypted, non-identified work. Audit records contain metadata only.

Costs and budgets

Every response will be metered: the gateway computes the cost of each request and records it against your key, and daily budget caps halt further requests once the cap is reached. Cost reporting will be available at onboarding.

Errors

Chat endpoints return OpenAI-shaped errors, so existing retry and error-handling logic keeps working:

{
  "error": {
    "message": "Daily budget exceeded for this key.",
    "type": "insufficient_quota",
    "code": "budget_exceeded"
  }
}

Deployment options

The Gateway is designed for a hosted service, your VPC, on-premises infrastructure or a dedicated data-centre deployment. Talk to us about the appropriate model.