Evanesca's models in your Microsoft Foundry agents.

Before you start

• An Azure subscription and a Microsoft Foundry project.
• The Foundry User role (or higher) on the project, and Contributor on the resource group: the connection is deployed as a resource. Microsoft renamed these roles; you may still see Azure AI User.
• An Evanesca API key and credit above zero: an agent whose calls get 402 stops answering.
• The IDs of the models you want to offer, from GET /models — for example mistral-large-4 or deepseek-v4.1-flash (model table).

# The catalogue needs no key: the names you declare in Foundry come from here
curl https://evanesca.app/api/v1/models

# The chat endpoint needs the key Foundry will carry
curl https://evanesca.app/api/v1/chat/completions \
  -H "Authorization: Bearer $EVANESCA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gemma-4-26b", "messages": [{"role": "user", "content": "ping"}]}'

Step by step

1. 01
   Open the connection wizard

   In the Foundry portal, select Manage, then Resource details. On the Admin-connected models tab, select Add.

2. 02
   Choose Other source

   On the Connection type page, pick Other source — the option for a self-hosted or non-Azure gateway. Give the connection a short name, such as evanesca: it will prefix every model reference. Put https://evanesca.app/api/v1 in Base URL; Foundry appends /chat/completions itself.

3. 03
   Enter the key

   On the Authentication page, choose API key. Paste your Evanesca key in API key, put Authorization in API key header name and Bearer {api_key} in API key header value.

4. 04
   Declare the models

   On the Model configuration page, select + Add model for each model your teams may use. Name is the Evanesca ID exactly as GET /models gives it (gemma-4-26b, for example) — it is what goes in the call; Display name is up to you. Leave Version empty.

5. 05
   Leave the advanced settings as they are

   On the Advanced page, leave API version empty and Include deployment name in URL path off: Evanesca uses OpenAI's /chat/completions path, with no deployment name. No extra header is needed. Select Add.

6. 06
   Use it in an agent

   Each declared model becomes a deployment. In your agent, the model reads <connection-name>/<model-name> — here evanesca/gemma-4-26b. Only prompt agents in the Agent SDK support it.

Provisioning as code? Microsoft's foundry-samples repository has Bicep templates for ModelGateway connections, deployed with az deployment group create.

import os
from azure.ai.projects import AIProjectClient
from azure.ai.projects.models import PromptAgentDefinition
from azure.identity import DefaultAzureCredential

project = AIProjectClient(
    endpoint=os.environ["FOUNDRY_PROJECT_ENDPOINT"],
    credential=DefaultAzureCredential(),
)

agent = project.agents.create_version(
    agent_name="eu-assistant",
    definition=PromptAgentDefinition(
        model="evanesca/gemma-4-26b",  # <connection-name>/<model-name>
        instructions="You are a helpful assistant.",
    ),
)

openai = project.get_openai_client()
conversation = openai.conversations.create()
reply = openai.responses.create(
    conversation=conversation.id,
    input="Hello",
    extra_body={"agent_reference": {"name": agent.name, "type": "agent_reference"}},
)
print(reply.output_text)

What changes, and what does not

What changes: Microsoft no longer serves the model. The inference runs at the operator behind the model you picked — in France by Scaleway and OVHcloud (as a fallback), in Mistral AI's own EU region, in Germany by AKI.IO and in Switzerland by Infomaniak — and the weights, the GPUs and the writing of the answer stay with that operator.

What does not: the agent still runs in Foundry. Prompts, tool calls and conversation history go through Microsoft's orchestration before they reach Evanesca. Microsoft says so from its side: with a bring-your-own model, you are responsible for reviewing the data you share with it and for your own safety measures, such as content filters (Microsoft Learn).

This setup therefore answers “our agents must not run on a US-hosted model”, not “no US company may process our data”. For the second, call Evanesca's API directly — same key, same models, no detour through Azure.

Troubleshooting

Frequently asked questions