How to Turn an AI Assistant into a Cloud Engineer for Azure
When you ask a language model to help with cloud deployment, you usually get a polite but useless response. The model cites generic documentation from five years ago, mixes up service names, and cheerfully recommends commands that fail with authorization errors. AI agents can write function code, but when it comes to infrastructure, access rights, and real resource management, they're blind.
Microsoft took a systematic approach to this problem and released the azure-skills repository. It's a plugin for AI agents that connects specific task context with live command execution through the MCP protocol.
What's Inside the Box
The project team divided the solution into two levels: decision-making logic and practical actions.
The first level is called skills. These are ready-made scenarios, decision trees, and constraints. They explain to the assistant how the Azure workflow is structured, what checks need to be performed before deployment, how to set up monitoring through Application Insights, or how to calculate resource costs.
The second level is execution tools. The plugin connects to Azure MCP Server, which contains over 200 tools for 40 services. If an agent decides to check the list of resource groups or read logs from a crashed container, it doesn't make up an answer—it makes a real API call through the MCP protocol.
There's also separate integration with Foundry MCP. It's needed for those who deploy machine learning models and build pipelines based on Microsoft Foundry.
To get started, you'll need Node.js 18 or higher, the Azure CLI installed (
), and Azure Developer CLI (
) if you're deploying infrastructure from templates.
If you use the APM package manager, installation across all supported environments is done with a single line:
First, add the source:
Then install the plugin itself:
In Claude Code, the plugin is available directly from the official catalog. Just run the command:
Or find it through the plugins interface:
In Cursor, installation goes through the extensions catalog in the editor settings:
In Codex, the plugin is added similarly through the marketplace:
After that, open the plugins list with the command
and select the one you need:
In Codex, you can also selectively disable individual skills through the
menu:
In the JetBrains environment, the process consists of a couple of steps. First, make sure the GitHub Copilot plugin is updated to the latest version, then enable experimental skills support in the settings: Tools > GitHub Copilot > Chat > Enable Skills.
After restarting the IDE, you can install the official Azure Toolkit for IntelliJ, which will itself offer to install the skills:
Or install skills globally via npx in the terminal:
You can verify the result directly in the Copilot chat by typing
:
Supported Environments and Installation
Developers made the plugin cross-platform. It can be connected to almost any popular agent.
You'll need Node.js 18+, the Azure CLI installed (
), and if you're deploying infrastructure from templates, also Azure Developer CLI (
).
If you use the APM package manager, installation across all supported environments is done with a single line:
First, add the source:
Then install the plugin itself:
GitHub Copilot CLI
First, add the source:
Then install the plugin itself:
Claude Code
In Claude Code, the plugin is available directly from the official catalog. Just run the command:
Or find it through the plugins interface:
Cursor
In Cursor, installation goes through the extensions catalog in the editor settings:
Codex CLI
In Codex, the plugin is added similarly through the marketplace:
After that, open the plugins list with the command
and select the one you need:
In Codex, you can also selectively disable individual skills through the
menu:
IntelliJ IDEA
In the JetBrains environment, the process consists of a couple of steps. First, make sure the GitHub Copilot plugin is updated to the latest version, then enable experimental skills support in the settings: Tools > GitHub Copilot > Chat > Enable Skills.
After restarting the IDE, you can install the official Azure Toolkit for IntelliJ, which will itself offer to install the skills:
Or install skills globally via npx in the terminal:
You can verify the result directly in the Copilot chat by typing
:
Working with Isolated Clouds
By default, the MCP server connects to Azure Public Cloud. If your company operates in isolated clouds like Azure China or Azure US Government, you need to adjust the configuration manually.
In Copilot CLI, the list of active servers is opened with the command
:
Then edit the parameters by calling
and add the argument
or
:
Before launching, make sure that local utilities are also authorized in the required environment. For example, for Azure CLI, the switch looks like this:
Real-World Scenarios
Once everything is configured, the assistant stops responding with abstract advice. You can give direct practical tasks right in the context of an open project:
— the agent analyzes the project, selects a suitable service (for example, Azure Container Apps), and generates configuration files.
— checks Bicep or Terraform templates for errors before sending to the cloud.
— makes a real request through MCP and outputs the list of storage accounts from the active subscription.
— runs cost analysis and finds forgotten unused disks or redundant virtual machines.
— requests fresh application logs, checks probe status, and explains the cause of the failure.
— selects the minimum necessary RBAC roles based on the principle of least privilege.
The Bottom Line
The
repository solves the main problem of cloud development with AI: the gap between code generation and real infrastructure context. Instead of manually transferring logs from the portal to the chat window and back, you give the agent access to managed tools.
Who will the project benefit most:
- Developers who regularly deploy services to Azure and are tired of switching between the terminal, portal, and documentation.
- Teams implementing AI agents in everyday workflows based on VS Code, Cursor, or Claude Code.
The repository is open under the MIT license, is actively updated, and is synchronized with the Azure team's developments. It's definitely worth trying, especially since basic setup takes just a few minutes.
Related projects