How to set Microsoft Foundry as a provider in OpenCode and Copilot CLI

Be your own AI provider

After Microsoft Copilot’s move from subscription requests to credit-based usage on 1 June 2026, the company has tried to lighten the bill shock by providing access to cheaper open weight models, such as Kimi K2.7 Code (available just over a one month later). Meanwhile, there have been plenty of even cheaper model options on another Microsoft enterprise platform hiding in plain sight: Microsoft Foundry. Read on to learn how to consume models you deploy yourself on MS Foundry alongside your regular options in both OpenCode and Copilot CLI.

Prerequisites

This is not a tutorial on Microsoft Foundry. The new Foundry view is pretty intuitive, documentation from Microsoft is pretty good for Foundry (in my opinion), and the barrier to entry to do the basics is pretty low (just don’t try and do evaluations straight away - they are a nightmare).

To follow along, you will need to have:

• A Foundry azure resource group (called a “project”)
• A base model deployment in that project, with an endpoint and API key
• The Cognitive Services OpenAI User or Foundry User role assigned to your user ID (at minimum)

When you look at the model endpoint, note how it conforms to the common openAI API endpoint format. We used this endpoint format previously to connect OpenCode Go to openwebUI and we’re using it again to connect Foundry to OpenCode/Copilot. Foundry does add a small quirk though. If your project is called “foundry-project”:

• The provider will be called “foundry-project”,
• The base URL will be https://foundry-project -resource .services.ai.azure.com/ openai/v1

I’ve separated the parts of the URL to show the first part (the project name) and the last part (the openAI compatible url ending) are what we expect, but the middle (appending “-resource” to the project name and the top level Azure domain) are something you couldn’t guess.

OpenCode

We’ll start with OpenCode first as the configuration is a lot easier. OpenCode’s main USP is that it connects to everything, and Foundry is configured as an openAI compatible custom provider rather than the Azure OpenAI or Azure Cognitive Services provider documented in the help page.

Below is a minimal opencode.json configuration file that does the following:

• Restricts the providers to Copilot and a custom provider I call ‘foundry-project’
• Define the custom foundry-project url
• Define two models from the provider (Deepseek V4 Flash and Kimi K2.6)
• Set the default model to be Deepseek V4 Flash

{
  "$schema": "https://opencode.ai/config.json",
  "enabled_providers": [
    "github-copilot",
    "foundry-project"
  ],
  "model": "foundry-project/DeepSeek-V4-Flash",
  "provider": {
    "foundry-project": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Solution Delivery Foundry",
      "options": {
        "baseURL": "https://foundry-project-resource.services.ai.azure.com/openai/v1"
      },
      "models": {
        "DeepSeek-V4-Flash": {
          "name": "DeepSeek V4 Flash"
        },
            "Kimi-K2.6": {
          "name": "Kimi-K2.6"
        }
      }
    }
  }
}

Copilot

Where OpenCode is open, agnostic to provider and the issue is locking down providers you don’t need (looking at you, privacy sucking free Zen models), Copilot is the inverse. Copilot has the feel of an enterprise tool: providers are often centrally managed along with the models they serve. To add a custom provider you may need to supply environment variables at runtime.

Assuming you have installed Copilot CLI in Windows Subsystem for Linux (WSL) or in Linux, edit or create your ~/.bashrc file with the following.

alias copilot-foundry='export COPILOT_PROVIDER_BASE_URL=https://foundry-project-resource.services.ai.azure.com/openai/v1 && \
export COPILOT_PROVIDER_API_KEY=<KEY HERE> && \
export COPILOT_MODEL=DeepSeek-V4-Flash && \
export COPILOT_PROVIDER_MAX_PROMPT_TOKENS=1000000 && \
export COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=128000 && \
copilot'

After restarting your terminal or typing source ~/.bashrc, you can type the alias copilot-foundry and launch your Foundry model.

Foundry VSCode chat provider

Copilot Chat in VS Code is a little easier to work with than Copilot CLI, insofar as you can specify custom endpoints rather than having to resort to workarounds with aliases. However, you do need to specify each model in turn (like OpenCode) along with their individual parameters (unlike OpenCode).

Below is the same example Foundry custom provider as for Copilot, but for the same two models as we had for OpenCode. Note the API key is supplied on the first invocation of the model in the chat window and stored as a secret.

      {
            "name": "foundry-project",
            "vendor": "customendpoint",
            "apiKey": "${input:chat.lm.secret.-28e8f8fd}",
            "apiType": "chat-completions",
            "models": [
                  {
                        "id": "DeepSeek-V4-Flash",
                        "name": "DeepSeek-V4-Flash",
                        "url": "https://foundry-project-resource.services.ai.azure.com/openai/v1",
                        "toolCalling": true,
                        "vision": true,
                        "maxInputTokens": 1000000,
                        "maxOutputTokens": 128000
                  },
                  {
                        "id": "Kimi-K2.6",
                        "name": "Kimi-K2.6",
                        "url": "https://foundry-project-resource.services.ai.azure.com/openai/v1",
                        "toolCalling": true,
                        "vision": true,
                        "maxInputTokens": 128000,
                        "maxOutputTokens": 16000
                  }
            ]
      }
]

Summary

When I first used this technique, I estimated I had saved around £72 in a week across 4 people and just over 15 million tokens, comparing Foundry Deepseek against Sonnet 4.6 from Copilot. Now there are cheaper models available directly from Copilot, the savings per token may not be as great, but there is still nothing as cheap as Deepseek V4 Flash or MiMo 2.5 available on Copilot as of now.

Maybe things will change again, but at least you know how to connect terminal user interfaces to Microsoft Foundry. And since the base model deployments are all openAI API compatible, it should serve to demystify any other such model deployment process you may stumble across in future.