This project was forked from nquandt/pi-azure-foundry.

pi-azure-foundry

A pi extension that connects to your Azure AI Foundry project, auto-discovers your chat deployments, and registers them as models in pi.

Supports both API key and Azure identity (Managed Identity, Azure CLI, service principal, etc.) authentication.

Requirements

• A pi installation (npm install -g @earendil-works/pi-coding-agent)
• An Azure AI Foundry project with one or more chat-capable deployments

Installation

Global (works in any project)

pi install npm:@nquandt/pi-azure-foundry

Try without installing

pi -e npm:@nquandt/pi-azure-foundry

Configuration

Create an azure-foundry.config.json file. The extension looks in two places, in order:

1. <current working directory>/azure-foundry.config.json — project-specific
2. ~/.pi/azure-foundry.config.json — global fallback

The global location is recommended for most users since it works across all projects.

Finding your resource and project IDs

Both values come from your Azure AI Foundry project URL:

https://ai.azure.com/build/overview?wsid=/subscriptions/.../resourceGroups/.../providers/Microsoft.MachineLearningServices/workspaces/YOUR-PROJECT

Or from the Azure portal — your resource name is the Azure AI Services resource name, and your project name is the Foundry project name. They are often the same value.

API key auth

{
  "resourceId": "my-resource-eastus2",
  "projectId": "my-project-eastus2",
  "auth": {
    "type": "api-key",
    "apiKey": "your-api-key-here"
  }
}

Get your API key from the Azure AI Foundry portal under Settings → API keys.

Azure identity auth

{
  "resourceId": "my-resource-eastus2",
  "projectId": "my-project-eastus2",
  "auth": {
    "type": "azure-identity"
  }
}

No key needed. Uses DefaultAzureCredential which automatically tries, in order:

| Method | How to set up | |---|---| | Environment variables | Set AZURE_CLIENT_ID, AZURE_CLIENT_SECRET, AZURE_TENANT_ID | | Workload identity | Configured automatically in AKS | | Managed identity | Configured automatically on Azure VMs / App Service / Container Apps | | Azure CLI | Run az login | | Azure Developer CLI | Run azd auth login | | Visual Studio Code | Sign in via the Azure extension |

For local development, az login is the easiest option.

Usage

Once installed and configured, start pi normally. On startup you'll see:

[Azure Foundry] Loading config from: /Users/you/.pi/azure-foundry.config.json
[Azure Foundry] Auth: api-key
[Azure Foundry] Fetching deployments from: https://my-resource.services.ai.azure.com/...
[Azure Foundry] Found 3 deployment(s): claude-sonnet (Anthropic), gpt-4o (Microsoft), ...
[Azure Foundry] ✓ Registered 3 model(s)

Your deployments will appear in the pi model picker under the Azure Foundry provider.

How it works

• Deployment discovery — on startup the extension calls the Foundry deployments API and filters to chat-capable deployments. No model list to maintain manually.
• Metadata resolution — model details (context window, max output tokens, reasoning support, vision support, and per-token pricing) are resolved by matching the Azure catalog model name against pi-ai's built-in model providers. The match is case-insensitive, so Kimi-K2.7-Code resolves to pi-ai's kimi-k2.7-code.
• Config overrides — you can pin or override details for any catalog model via the optional models property in azure-foundry.config.json. This takes precedence over the pi-ai catalog lookup and is useful for custom deployments, negotiated pricing, or models not yet in pi-ai. See the example below.
• Routing — Anthropic deployments are routed to /anthropic/v1/messages (native Messages API with tool use and extended thinking). All other deployments use /openai/deployments/{id}/chat/completions (OpenAI-compatible). Newer GPT-5/o-series models use max_completion_tokens instead of max_tokens; this is inferred from model name or set explicitly in models config overrides.
• Auth headers — API key auth sends api-key: <key> on the OpenAI route and Authorization: Bearer <key> on the Anthropic route. Azure identity sends Authorization: Bearer <entra-token> on both. Tokens are cached and refreshed automatically 5 minutes before expiry.

Supported models

The extension auto-discovers whatever is deployed in your Foundry project. Known models are resolved through pi-ai's built-in provider catalogs (Anthropic, OpenAI, MoonshotAI, Mistral, DeepSeek, xAI, and their regional variants).

Any deployment whose catalog model name can't be matched falls back to conservative defaults (128K context / 4K output / text-only / no reasoning / no cost). To override or plug a gap, add an entry under the models key in azure-foundry.config.json.

Overriding model details

The models section lets you override any subset of the resolved metadata for a specific Azure catalog model. The key must match the Azure model name exactly (e.g. Kimi-K2.7-Code). Any omitted fields are kept from the pi-ai catalog or fallback defaults.

{
  "resourceId": "my-resource-eastus2",
  "projectId": "my-project-eastus2",
  "auth": {
    "type": "api-key",
    "apiKey": "your-api-key-here"
  },
  "models": {
    "Kimi-K2.7-Code": {
      "contextWindow": 262144,
      "maxTokens": 32768,
      "reasoning": true,
      "input": ["text", "image"],
      "cost": {
        "input": 0.95,
        "output": 4.0,
        "cacheRead": 0.19,
        "cacheWrite": 0
      },
      "openaiTokenLimit": "max_tokens"
    }
  }
}

Supported override fields:

| Field | Type | Purpose | |---|---|---| | contextWindow | number | Model context window in tokens | | maxTokens | number | Maximum output tokens per request | | reasoning | boolean | Whether the model emits reasoning/thinking content | | input | ["text"], ["text", "image"], etc. | Supported input modalities | | cost | { input, output, cacheRead?, cacheWrite? } | Per-1M-token pricing in USD | | openaiTokenLimit | "max_tokens" or "max_completion_tokens" | Which field pi sends for the output token limit |

Common use-cases include fixing stale data in the pi-ai catalog, setting custom parameters configured in Foundry (e.g. maxTokens), or ensuring cost estimates reflect special pricing from a negotiated arrangement with Microsoft Azure.

"models": {
  "Kimi-K2.7-Code": {
    "maxTokens": 32768
  }
}

Development

git clone https://github.com/nquandt/pi-azure-foundry
cd pi-azure-foundry
npm install
npm run build

# Test against your own config
cp azure-foundry.config.example.json azure-foundry.config.json
# edit azure-foundry.config.json with your real values
pi -e .

npm run dev       # watch mode
npm run type-check  # type check without building

License

MIT