Cloud model providers

Overview

mouaif ships with four additional first-class API-key providers on top of the original six: Azure OpenAI, Mistral, Groq, and DeepSeek. All four are OpenAI-shaped at the wire level — the same /chat/completions SSE contract and the same /models catalog shape — so they reuse the existing openai-compatible request builder, SSE parser, tool loop, and usage accounting. Adding them was a localized change: one ENDPOINTS entry (with a listModels adapter) in src/ai-endpoints.js, a row in the Settings providers list, and a keyring-namespace mapping in src/auth.js.

Usage

Add a provider

1. Settings → Providers → + Add provider.

2. Pick the provider from the dropdown:

• Azure OpenAI — paste the resource key. The API base URL must be the full deployment URL, e.g. https://<resource>.openai.azure.com/openai/deployments/<deployment>. The deployment name is the model id on the chat side. The api-version query parameter is appended automatically (default 2024-10-21; override per project model via model.apiVersion).

• Mistral — default base URL https://api.mistral.ai/v1. Paste a key from console.mistral.ai.

• Groq — default base URL https://api.groq.com/openai/v1. Paste a key from console.groq.com.

• DeepSeek — default base URL https://api.deepseek.com. Paste a key from platform.deepseek.com.

1. Save provider. All four are API-key only — the Settings form shows no OAuth option for them.

Add a model

In the project's Models editor set provider to azure / mistral / groq / deepseek and id to the model id (for Azure, the deployment name):

// <projectDir>/.mouaif.json
{
  "models": [
    { "id": "gpt-4o", "provider": "azure", "label": "GPT-4o (my deployment)", "apiVersion": "2024-10-21" },
    { "id": "mistral-large-latest", "provider": "mistral", "label": "Mistral Large" },
    { "id": "llama-3.3-70b-versatile", "provider": "groq", "label": "Llama 3.3 70B" },
    { "id": "deepseek-chat", "provider": "deepseek", "label": "DeepSeek Chat" }
  ]
}

The live model catalog (chat head refresh ↻) pulls each provider's real /models list, so picking from the picker works without hand-editing JSON.

Chat

No special path. The chat composer posts to /api/chats/:id/messages/stream, which hydrates the project model with the provider connection and streams SSE exactly like the other OpenAI-shaped providers. Tool calls, the multi-turn loop, usage/cost lines, and prompt-size profiles all work unchanged.

Behavior

• OpenAI-shaped wire protocol. All four use the OpenAI chat-completions request/response shape, so they share buildOpenAIRequest and parseOpenAISSE. Streaming requests set stream_options.include_usage so the Context/Cost line has authoritative token counts (Azure needs api-version ≥ 2024-10-21 for this).

• Azure specifics. Auth is the api-key header (the OpenAI Authorization: Bearer header is not sent). The base URL is the deployment URL; api-version is appended after /chat/completions. The default api-version is 2024-10-21; a project model may set apiVersion to pin a different one (e.g. a preview API or an Entra-ID proxy).

• Mistral / Groq / DeepSeek specifics. Plain Authorization: Bearer <key> against their /chat/completions; model catalogs are fetched from their /models endpoints. Missing keys surface as ENO_APIKEY (the chat UI says "add API key in Settings → Providers") instead of a generic upstream 401.

• Thinking levels. Reasoning families on these OpenAI-shaped endpoints (DeepSeek deepseek-reasoner / deepseek-r1, Qwen qwq / -thinking variants, Groq's qwen-qwq-32b) are inferred by thinkingForOpenAIModel and advertised as effort levels, so the chat thinking dropdown works without a client update.

• Pricing. Built-in defaults were added for common Mistral, Groq, and DeepSeek model ids in src/usage.js; unknown ids fall back to -- on the cost line (or a per-model pricing override).

• No OAuth. All four are API-key-only. AI_TO_AUTH_PROVIDER maps each to its own keyring namespace (azure, mistral, groq, deepseek) so a key stored under one provider is never reused as a credential for another.

Related

• docs/features/ai-client.md — the proxy architecture all providers ride on.

• docs/features/settings-ui.md — the provider form that renders SETTINGS_PROVIDERS.

• docs/features/usage-metrics.md — cost line behavior and pricing resolution.