criogenai

All posts

CriogenAI 0.1.6: your models through Azure, Bedrock and Vertex AI

The CriogenAI team2 min read

  • release
  • providers
  • azure
  • bedrock
  • vertex
  • openai

Most companies do not buy models from the model makers directly. They buy them through their cloud: Azure, AWS or Google Cloud. The reasons are data residency, the agreements they already have, and one bill. CriogenAI 0.1.6 reaches models that way, in Python, TypeScript and Rust:

  • Azure OpenAI and Microsoft Foundry;
  • Amazon Bedrock;
  • Google Vertex AI.

It also moves OpenAI's own models to OpenAI's newer API, so they can reason while they call tools.

pip install -U criogenai
npm install criogenai@latest

Azure OpenAI and Foundry#

The model name is your deployment's name. The endpoint is your resource.

from criogenai import Agent, Azure

agent = Agent(Azure("gpt-5-mini", endpoint="my-resource", managed_identity=True))

You can sign in four ways:

  • an API key;
  • a service principal (an app registered in Entra ID);
  • the identity of the Azure machine or app the code runs on, so there is no key to store at all;
  • a token from somewhere else.

CriogenAI gets the sign-in tokens and renews them before they run out. Foundry's models from other makers (DeepSeek, Llama, Mistral, Grok) work too, with api="chat".

Amazon Bedrock#

from criogenai import Agent, Bedrock

agent = Agent(Bedrock("us.anthropic.claude-sonnet-5", region="us-east-1"))

Claude goes over Anthropic's own message format, and other makers' models over the OpenAI-compatible one. You don't have to choose: Bedrock's model names say who made the model. The key comes from AWS_BEARER_TOKEN_BEDROCK.

There is no default region. A request never goes to a region you did not name, because the region decides where your data is processed.

Google Vertex AI#

from criogenai import Agent, Vertex

agent = Agent(Vertex("claude-sonnet-5", project="my-project", location="europe-west4"))

Gemini and Google's other hosted models work, and so does Claude. CriogenAI signs in with:

  • your gcloud auth application-default login credentials;
  • the service account of the Google Cloud machine it runs on;
  • or a token you pass.

As with Bedrock, project and location have no defaults.

OpenAI's newer API#

openai: models now use OpenAI's Responses API. This matters for newer models: OpenAI's older Chat Completions API does not let them reason while they use tools.

CriogenAI also asks OpenAI to store nothing. The model's reasoning comes back encrypted and is sent along with the conversation, so it keeps its train of thought across tool calls. We checked all of this against OpenAI's live API.

What to know#

  • Tested: Azure, Bedrock and Vertex AI were tested against local test servers, not live accounts yet. The status page says so, and will say when that changes.
  • Not yet supported: signing in with AWS IAM roles, and Google service-account key files. Both need request signing, which comes next.
  • Platforms: 0.1.6 was built on Windows, like 0.1.5. Windows and the browser build have everything. Linux and macOS get it when their builds can run again, after 15 November. Until then those platforms say plainly which engine a feature needs.

Every provider, with its environment variables, is in the models guide.