Azure OpenAI Open-Core

The proxy serves Azure OpenAI and Azure AI Foundry deployments directly — OpenAI models through Azure's OpenAI v1 API, and Claude models on Foundry through the Messages API Foundry serves — and applies the same governance it applies to OpenAI — DLP and PII scanning in both directions, SOP gates, WASM and Rego rules, the response gate, cost metering and budgets, traces and the response cache.

Naming a deployment

Prefix the deployment name with azure/: a deployment called gpt4o-prod is azure/gpt4o-prod. The proxy sends it to <endpoint>/openai/v1/chat/completions or /openai/v1/responses with the deployment as model, so no api-version is involved. Foundry models deployed on the same resource (*.services.ai.azure.com) are named the same way.

To serve a model name your client already sends, alias it in config.yaml:

model_list:
  - model_name: gpt-4o
    litellm_params:
      model: azure/gpt4o-prod

Endpoint and credentials

A self-hosted proxy's own credentials. Set AZURE_OPENAI_ENDPOINT or intutic_settings.providers.azure.endpoint — any http(s) address, so an API gateway or private endpoint in front of Azure works — and authenticate with one of:

1. an API key: AZURE_OPENAI_API_KEY, or providers.azure.api_key: os.environ/<NAME>;
2. Microsoft Entra ID client credentials: AZURE_TENANT_ID, AZURE_CLIENT_ID and AZURE_CLIENT_SECRET (or the matching providers.azure settings);
3. the host's managed identity: providers.azure.managed_identity: true, with client_id for a user-assigned identity.

Entra ID tokens are requested for https://cognitiveservices.azure.com/.default (for Claude deployments, https://ai.azure.com/.default) and cached until five minutes before they expire; the identity needs the Cognitive Services OpenAI User role on the resource (Foundry User for Claude).

What you can send

Send OpenAI-model deployments through /v1/chat/completions or /v1/responses — the OpenAI formats Azure speaks. Requests and answers, streaming and tool calls included, pass through as OpenAI's; Azure's extra content-filter chunks are forwarded as they arrive.

Claude on Foundry

Claude models deployed on a Foundry resource (https://<resource>.services.ai.azure.com) are served through the Messages API Foundry exposes at <endpoint>/anthropic/v1/messages: the request goes unchanged, with the deployment name as model, the key in api-key and anthropic-version, and the answer and stream come back as Anthropic's. The proxy uses this API when:

• the request is in Anthropic's format (/v1/messages), whatever the deployment is called; or
• the deployment's name starts with claude (Foundry's default deployment names are the model ids, such as claude-opus-5-5), in which case OpenAI-format requests are translated to it as for any Claude model.

curl http://localhost:4000/v1/messages -H "authorization: Bearer $INTUTIC_KEY" \
  -d '{"model": "azure/claude-opus-5-5", "max_tokens": 1024, "messages": [{"role": "user", "content": "Hello"}]}'

An Azure OpenAI resource (*.openai.azure.com) cannot host Claude, so a Claude request for one is refused with 400 and the reason. Features Foundry does not offer for Claude (for example the Message Batches and Files APIs, or code execution on Azure-hosted deployments) are refused by Foundry itself.

Errors, retries and cost

Azure's errors are OpenAI's and reach the client unchanged. A 429 carries retry-after-ms, which the proxy's retries honour; a 400 content_filter refusal is final.

Cost is priced by the deployment name, so azure/gpt-4o costs what gpt-4o costs and azure/claude-opus-5-5 what claude-opus-5-5 costs. A deployment named after nothing in the price list is charged at the conservative estimate; name deployments after their model, or alias them as above, to price them exactly.