Add Azure MCP Tools

Add a New Tool to Azure MCP

Purpose

Step-by-step workflow for adding a new command to any Azure MCP toolset. Each phase has an explicit gate — do not proceed until the gate passes.

Decision: New Toolset or Existing?

Before starting, determine:

• Adding to existing toolset? → Skip to Phase 1.
• Creating a new toolset? → Complete Phase 0 first.

⚠️ CRITICAL: Does your command interact with Azure resources?

Security Requirements

Terminology note: In this repo, tool refers to the MCP-exposed capability and command refers to the underlying C# command class implementation. Both terms appear in the codebase and docs.

All tool inputs are untrusted. These requirements apply at every phase and are not optional.

Input Validation

• Validate inputs against the specific naming rules of the Azure resource being targeted (length, allowed characters, casing). Do not apply a generic blocklist — Azure resource naming rules vary significantly by service.
  • Example: Storage account names are 3–24 lowercase alphanumeric characters only.
  • Example: Resource group names allow letters, digits, underscores, hyphens, and periods up to 90 characters.
  • Reference: Azure naming rules and restrictions
• Use ValidateOptions for semantic constraints beyond nullability (name length, format, mutual exclusivity, and allowed value sets). Only reject characters that are provably invalid for the specific resource type. This applies to both the new two-generic SubscriptionCommand pattern and the legacy one-generic pattern — see the ValidateOptions override guidance in Phase 1d.
• Prefer SDK/runtime validators and deterministic checks first (Length, explicit allowed-value sets, character/category checks).

Secure Logging

• Never log raw option objects ({@Options}) — they may contain secrets, connection strings, or PII.
• Log only individually named, known-safe parameters. For example: options.Subscription, options.ResourceGroup, Name.
• Do not include sensitive field values in error messages returned to callers.
• Strip or redact secret values before surfacing exception details.

// ✅ Log only known-safe, individually named fields
_logger.LogError(ex, "Error in {Operation}. Subscription: {Subscription}, ResourceGroup: {ResourceGroup}",
    Name, options.Subscription, options.ResourceGroup);

// ❌ Never log the whole options object — it may contain keys, connection strings, or PII
_logger.LogError(ex, "Error in {Operation}. Options: {@Options}", Name, options);

// ❌ Never surface raw exception bodies to callers — they may contain tokens or account metadata
return $"Request failed: {requestFailedException.Message}"; // may include auth headers

Safe Downstream Interactions

• Never concatenate user input directly without prior validation into URLs, shell commands, resource identifiers, query strings, etc.
• Use EndpointValidator from Microsoft.Mcp.Core.Helpers to guard all endpoint usage — choose the method that matches your scenario:
  • Azure service data-plane endpoint (endpoint derived from a resource name, e.g. storage account, ACR, App Config): call EndpointValidator.ValidateAzureServiceEndpoint(endpoint, serviceType, AzureService.CloudConfiguration.ArmEnvironment) before constructing the client. This enforces the correct per-cloud domain suffix (e.g. .blob.core.windows.net / .blob.core.chinacloudapi.cn) and HTTPS.
  • User-supplied URL to a known external service (e.g. a GitHub URL the user provides): call EndpointValidator.ValidateExternalUrl(url, allowedHosts) with an explicit allowlist of permitted hosts.
  • User-supplied target URL with no known domain (e.g. a load-test target the user controls): call EndpointValidator.ValidatePublicTargetUrl(url), which enforces HTTPS/HTTP-only schemes, rejects private/reserved IP ranges, rejects reserved hostnames, and resolves DNS to catch hostnames that map to internal IPs.
• For services that construct the endpoint internally (not from user input), use the cloud-type switch pattern (see Phase 1c: Service Implementation) — EndpointValidator is not required in that case but ValidateAzureServiceEndpoint can be added as a defense-in-depth layer.
• Control-plane operations (ARM resource creation, RBAC assignments, policy etc.) do not need EndpointValidator because they go through the typed Azure SDK ARM client. Construct ARM resource IDs using ResourceIdentifier or collection helpers — never by string-interpolating subscription/resource-group/resource-name directly into a raw ARM path.
• For new commands and services, always pass CancellationToken as the final parameter to all async downstream calls and propagate it throughout — never substitute CancellationToken.None or default at call sites.
• Fail closed: if tenant, subscription, or resource context is ambiguous, return an explicit validation error and require the caller to specify the value. Do not silently pick a default.

MCP-Specific Threat Patterns

Using AI to Generate Tool Code

When using an AI assistant (such as GitHub Copilot) to scaffold or generate command, service, or test code, include the following in every prompt:

Requirements:
- Validate user-controlled inputs in `ValidateOptions` using resource-specific rules (naming rules, allowed values, length caps) where applicable; do not use one generic rule for all options.
- Prefer SDK/runtime validators and deterministic checks for user input validation.
- Log only individually named, known-safe parameters; never log option objects, credentials, keys, connection strings, or other secret-bearing fields.
- For endpoint/URL inputs, use `EndpointValidator` methods appropriate to the scenario (`ValidateAzureServiceEndpoint`, `ValidateExternalUrl`, or `ValidatePublicTargetUrl`). Avoid direct interpolation of unvalidated input into URLs or downstream queries.
- Add negative tests for relevant security cases introduced by the command (for example malformed names, invalid endpoint hosts, or unsafe query text) rather than a one-size-fits-all set of tests.
- Keep error messages actionable but non-revealing: avoid exposing stack traces, raw backend payloads, or sensitive values to callers.

Review every AI-generated snippet for these properties before committing. Generated code that omits them must be corrected before the security gate in Phase 7 can be met.

Phase 0: New Toolset Setup (skip if adding to existing toolset)

Create the toolset directory structure:

tools/Azure.Mcp.Tools.{Toolset}/
├── src/
│   ├── Azure.Mcp.Tools.{Toolset}.csproj
│   ├── {Toolset}Setup.cs
│   ├── Commands/
│   │   ├── {Resource}/
│   │   │   └── {Resource}{Operation}Command.cs
│   │   └── {Toolset}JsonContext.cs
│   ├── Options/
│   │   └── {Resource}/
│   │       └── {Resource}{Operation}Options.cs
│   ├── Services/
│   │   ├── I{Toolset}Service.cs
│   │   └── {Toolset}Service.cs
│   └── Models/
└── tests/
    ├── Azure.Mcp.Tools.{Toolset}.Tests/
    │   └── Azure.Mcp.Tools.{Toolset}.Tests.csproj
    ├── test-resources.bicep          (Azure service commands only)
    └── test-resources-post.ps1       (Azure service commands only)

Required setup steps:

1. Add package version to Directory.Packages.props (if Azure SDK needed)

2. Register the project in solution files by running: pwsh eng/scripts/Update-Solutions.ps1 -All

3. Register the new toolset in servers/Azure.Mcp.Server/src/Program.cs RegisterAreas() (alphabetical order)

4. Choose the appropriate base class:

   • Commands that need an Azure subscription (most Azure service tools) → inherit from SubscriptionCommand<TOptions, TResult> and inject ISubscriptionResolver.
   • Commands that do NOT need a subscription (CLI wrappers, documentation tools, best-practice advisors) → inherit from BaseCommand<TOptions, TResult> directly.

   Only add a shared intermediate base command if you have real cross-command logic shared by multiple commands in the same toolset.

5. Register both the service and the command as singletons in {Toolset}Setup.cs ConfigureServices:

   public void ConfigureServices(IServiceCollection services)
   {
       services.AddSingleton<I{Toolset}Service, {Toolset}Service>();
       services.AddSingleton<{Resource}{Operation}Command>();
   }
   

GATE: dotnet build servers/Azure.Mcp.Server/src must pass.

Phase 1: Implement Command

Create these files in order:

1a. Options Class (with [Option] attributes)

File: src/Options/{Resource}/{Resource}{Operation}Options.cs

using Azure.Mcp.Core.Options;
using Microsoft.Mcp.Core.Models;
using Microsoft.Mcp.Core.Options;

namespace Azure.Mcp.Tools.{Toolset}.Options.{Resource};

public class {Resource}{Operation}Options : ISubscriptionOption
{
    [Option("Description of what this option does (e.g., 'The name of the resource').")]
    public string? MyOption { get; set; }

    [Option(OptionDescriptions.ResourceGroup)]
    public string? ResourceGroup { get; set; }

    [Option(OptionDescriptions.Subscription)]
    public string? Subscription { get; set; }

    [Option(OptionDescriptions.Tenant)]
    public string? Tenant { get; set; }
}

Rules:

• Implement ISubscriptionOption for commands that need subscription resolution
• Use [Option("description")] for the description — property name auto-converts to --kebab-case
• Use [Option(Name = "custom")] only when the default kebab-case conversion is wrong (e.g., when property is named FooBar and has [Option(Name = "foobar")] you get --foobar instead of --foo-bar)
• Use [Option(OptionDescriptions.X)] for shared descriptions (Subscription, Tenant, ResourceGroup, AuthMethod)
• Use [OptionContainer<TContainer>(Prefix = "prefix")] for model types which contain nested parameters. "prefix" will be prepended to the [Option]s in the model type (e.g., when [OptionContainer<TContainer>(Prefix = "foo")]'s model contains [Option(Name = "bar")] the parameter name is --foo-bar).
• Use subscription (never subscriptionId) — supports both IDs and names
• Use resourceGroup (never resourceGroupName)
• Use singular nouns for resources (server not serverName)
• Remove unnecessary name suffixes (Account / --account not AccountName / --account-name)
• Use required on required options; use nullable types (?) for optional options.
• Non-nullable value types (e.g., public int Count { get; set; }) are always valid without required — they default to 0. Use required if the caller must explicitly provide a value, or use int? if the parameter should be truly optional.
• Order: command-specific options first, then ResourceGroup, Subscription, Tenant, AuthMethod
• Keep parameter names consistent with Azure SDK parameters when possible

Note: Options are defined entirely via [Option] attributes. A static {Toolset}OptionDefinitions class is not needed

1c. Service Interface and Implementation

File: src/Services/I{Toolset}Service.cs

Return type depends on operation type:

• Resource Graph queries → Task<ResourceQueryResults<MyModel>> (includes AreResultsTruncated flag)
• Data plane operations → Task<List<MyModel>>
• Write operations → Task<MyResultModel>

public interface I{Toolset}Service
{
    // Resource Graph read operation
    Task<ResourceQueryResults<MyModel>> GetResourcesAsync(
        string? myOption,
        string subscription,
        string? resourceGroup = null,
        string? tenant = null,
        CancellationToken cancellationToken = default);

    // Data plane operation (returns simple List)
    Task<List<MyDetail>> GetDetailsAsync(
        string resourceName,
        string subscription,
        string? tenant = null,
        CancellationToken cancellationToken = default);
}

File: src/Services/{Toolset}Service.cs

Choose base class:

• Operations that need Resource Graph (ARG) queries: inherit BaseAzureResourceService
• All other operations (ARM, data plane): inherit BaseAzureService

BaseAzureResourceService extends BaseAzureService — neither is inherently read-only or write-only. The distinction is whether you need ARG querying functionality.

public class {Toolset}Service(IAzureService azureService)
    : BaseAzureResourceService(azureService), I{Toolset}Service
{
    public async Task<ResourceQueryResults<MyModel>> GetResourcesAsync(
        string? myOption,
        string subscription,
        string? resourceGroup = null,
        string? tenant = null,
        CancellationToken cancellationToken = default)
    {
        return await ExecuteResourceQueryAsync(
            "Microsoft.{Provider}/{resourceType}",
            resourceGroup,
            subscription,
            null,
            ConvertToModel,
            tenant: tenant,
            cancellationToken: cancellationToken);
    }

    private static MyModel ConvertToModel(JsonElement item)
    {
        var data = MyModelData.FromJson(item);
        return new MyModel(
            Name: data.ResourceName,
            Id: data.ResourceId,
            Location: data.Location.ToString(),
            Tags: data.Tags as IReadOnlyDictionary<string, string>
        );
    }
}

For write operations (using direct ARM clients):

public class {Toolset}Service(IAzureService azureService)
    : BaseAzureService(azureService), I{Toolset}Service
{
    public async Task<MyResource> CreateResourceAsync(
        string resourceName,
        string resourceGroup,
        string subscription,
        string? tenant = null,
        CancellationToken cancellationToken = default)
    {
        var subscriptionResource = await AzureService.GetSubscription(subscription, tenant, cancellationToken: cancellationToken);

        // CRITICAL: Use GetResourceGroupAsync with await
        var rgResource = await subscriptionResource.GetResourceGroupAsync(resourceGroup, cancellationToken);
        var resource = await rgResource.Value
            .GetMyResources()
            .GetAsync(resourceName, cancellationToken: cancellationToken);
        return resource.Value;
    }
}

Sovereign cloud rules:

• ARM/Resource Graph operations: cloud-aware automatically, no extra work
• Data plane endpoints: use AzureService.CloudConfiguration.CloudType switch — never hardcode URLs

Data plane endpoint pattern (required for services like Storage, Cosmos, Search):

public class MyService(IAzureService azureService)
    : BaseAzureResourceService(azureService), IMyService
{

    private async Task<MyDataPlaneClient> CreateDataPlaneClientAsync(
        string resourceName,
        string? tenant = null,
        CancellationToken cancellationToken = default)
    {
        var endpoint = GetResourceEndpoint(resourceName);
        var options = AddDefaultPolicies(new MyClientOptions());
        options.Transport = new HttpClientTransport(AzureService.GetClient());
        return new MyDataPlaneClient(
            new Uri(endpoint),
            await GetCredential(tenant, cancellationToken),
            options);
    }

    private string GetResourceEndpoint(string resourceName)
    {
        return AzureService.CloudConfiguration.CloudType switch
        {
            AzureCloudConfiguration.AzureCloud.AzurePublicCloud =>
                $"https://{resourceName}.service.core.windows.net",
            AzureCloudConfiguration.AzureCloud.AzureChinaCloud =>
                $"https://{resourceName}.service.core.chinacloudapi.cn",
            AzureCloudConfiguration.AzureCloud.AzureUSGovernmentCloud =>
                $"https://{resourceName}.service.core.usgovcloudapi.net",
            _ => $"https://{resourceName}.service.core.windows.net"
        };
    }
}

Patterns and anti-patterns:

// ❌ Hardcoded public-cloud endpoint
var client = new BlobServiceClient(new($"https://{account}.blob.core.windows.net"), credential, options);

// ❌ Hardcoded connection string
var connectionString = $"AccountEndpoint=https://{server}.documents.azure.com:443/;...";

// ✅ Cloud-aware endpoint via switch expression
var endpoint = GetBlobEndpoint(account);
var client = new BlobServiceClient(new(endpoint), credential, options);

Reference implementations: StorageService, CosmosService, SearchService, ConfidentialLedgerService.

1d. Long-running operations

Long-running operations (at this time) don't offer the ability to configure polling intervals, and even if they were able to there is a limit on how small of a polling interval can be used. Due to this, to prevent long-running operations with a significant number of polls from wasting CPU time waiting during testing, all long-running operations should use a two call pattern. The first call is the service method starting the polling operation, that should pass WaitUntil.Started to simply begin the operation. Then waiting for completion should call BaseAzureService.WaitForLroCompletionAsync to wait for completion in a way that testing can ignore the polling interval to prevent CPU wait loops that aren't necessary when playback testing.

var lroOperation = Service.LroAsync(WaitUntil.Started, cancellationToken);
await WaitForLroCompletionAsync(lroOperation, cancellationToken);

1e. Command Class

File: src/Commands/{Resource}/{Resource}{Operation}Command.cs

Required using statements:

using Azure.Mcp.Core.Commands.Subscription;
using Azure.Mcp.Core.Services.Azure.Subscription;
using Azure.Mcp.Tools.{Toolset}.Models;
using Azure.Mcp.Tools.{Toolset}.Options.{Resource};
using Azure.Mcp.Tools.{Toolset}.Services;
using Microsoft.Extensions.Logging;
using Microsoft.Mcp.Core.Commands;
using Microsoft.Mcp.Core.Models.Command;

[CommandMetadata(
    Id = "<generate-new-guid>",
    Name = "operation",
    Title = "Human Readable Title",
    Description = """
        What this command does. Include required options and return format.
        """,
    Destructive = false,
    Idempotent = true,
    OpenWorld = false,
    ReadOnly = true,
    Secret = false,
    LocalRequired = false)]
public sealed class {Resource}{Operation}Command(
    ILogger<{Resource}{Operation}Command> logger,
    I{Toolset}Service service,
    ISubscriptionResolver subscriptionResolver)
    : SubscriptionCommand<{Resource}{Operation}Options, {Resource}{Operation}Command.{Resource}{Operation}CommandResult>(subscriptionResolver)
{
    private readonly ILogger<{Resource}{Operation}Command> _logger = logger;
    private readonly I{Toolset}Service _service = service;

    public override async Task<CommandResponse> ExecuteAsync(
        CommandContext context, {Resource}{Operation}Options options, CancellationToken cancellationToken)
    {
        try
        {
            var results = await _service.GetResourcesAsync(
                options.MyOption,
                options.Subscription!,
                options.ResourceGroup,
                options.Tenant,
                cancellationToken);

            context.Response.Results = ResponseResult.Create(
                new {Resource}{Operation}CommandResult(results?.Results ?? [], results?.AreResultsTruncated ?? false),
                {Toolset}JsonContext.Default.{Resource}{Operation}CommandResult);
        }
        catch (Exception ex)
        {
            _logger.LogError(ex, "Error in {Operation}. Subscription: {Subscription}",
                Name, options.Subscription);
            HandleException(context, ex);
        }

        return context.Response;
    }

    public record {Resource}{Operation}CommandResult(List<MyModel> Items, bool AreResultsTruncated);
}

Key points (two-generic pattern from docs/option-conversion.md):

• Two generic parameters: SubscriptionCommand<TOptions, TResult> — TResult is the command's result record
• ISubscriptionResolver injected via primary constructor and passed to base
• ExecuteAsync receives pre-bound TOptions options — no ParseResult parameter
• No RegisterOptions()/BindOptions() overrides needed — OptionBinder handles binding via [Option] attributes
• No manual Validate() call — framework validates based on nullability and ValidateOptions() override
• Result record is public (for JSON serialization context visibility) and declared inside the command class
• DO NOT log {@Options} — may expose sensitive information

Custom validation (required for semantic and security constraints beyond nullability):

public override void ValidateOptions({Resource}{Operation}Options options, ValidationResult validationResult)
{
    base.ValidateOptions(options, validationResult);  // checks --subscription

    // Required-field check
    if (string.IsNullOrEmpty(options.MyRequiredField))
    {
        validationResult.Errors.Add("--my-required-field is required.");
    }

    // Security: validate against the specific Azure resource's naming rules.
    // Prefer deterministic checks first (length + character/category checks).
    // Look up exact constraints at:
    // https://learn.microsoft.com/azure/azure-resource-manager/management/resource-name-rules
    //
    // Example for a Storage account name (3–24 lowercase alphanumeric only):
    if (options.Account is not null &&
        !IsValidStorageAccountName(options.Account))
    {
        validationResult.Errors.Add("--account must be 3–24 lowercase alphanumeric characters (storage account naming rule).");
    }
}

private static bool IsValidStorageAccountName(string value)
{
    if (value.Length is < 3 or > 24)
        return false;

    foreach (var ch in value)
    {
        if (!((ch >= 'a' && ch <= 'z') || (ch >= '0' && ch <= '9')))
            return false;
    }

    return true;
}

Intermediate base commands (only if you have shared cross-command logic):

// Use interface constraints for type-safe access to shared options
public abstract class Base{Toolset}Command<
    [DynamicallyAccessedMembers(TrimAnnotations.CommandAnnotations)] TOptions, TResult>(
    ISubscriptionResolver subscriptionResolver)
    : SubscriptionCommand<TOptions, TResult>(subscriptionResolver)
    where TOptions : class, ISubscriptionOption, I{Toolset}Option
{
    public override void ValidateOptions(TOptions options, ValidationResult validationResult)
    {
        base.ValidateOptions(options, validationResult);
        // Shared validation using options.SharedProperty
    }
}

1f. JSON Serialization Context

File: src/Commands/{Toolset}JsonContext.cs

[JsonSerializable(typeof({Resource}{Operation}Command.{Resource}{Operation}CommandResult))]
[JsonSerializable(typeof(MyModel))]
[JsonSourceGenerationOptions(
    PropertyNamingPolicy = JsonKnownNamingPolicy.CamelCase,
    DefaultIgnoreCondition = JsonIgnoreCondition.WhenWritingNull)]
internal partial class {Toolset}JsonContext : JsonSerializerContext;

Guidelines:

• Only include types actually serialized as top-level result payloads
• Keep [JsonSerializable] attributes sorted by typeof model name
• Use one context per toolset always
• Filename must match class name ({Toolset}JsonContext.cs)
• Use {Toolset}JsonContext.Default.{CommandResult} when serializing — never JsonSerializer.Deserialize<T>() without a context

1g. Register Command

File: src/{Toolset}Setup.cs

public class {Toolset}Setup : IAreaSetup
{
    public string Name => "{toolset}";

    public string Title => "Manage Azure {Toolset}";

    public void ConfigureServices(IServiceCollection services)
    {
        services.AddSingleton<I{Toolset}Service, {Toolset}Service>();

        // Register all commands as singletons
        services.AddSingleton<{Resource}{Operation}Command>();
    }

    public CommandGroup RegisterCommands(IServiceProvider serviceProvider)
    {
        var root = new CommandGroup(Name,
            """
            {Toolset} operations - description of what this toolset covers.
            """,
            Title);

        var resource = new CommandGroup("{resource}", "{Resource} operations description");
        root.AddSubGroup(resource);
        resource.AddCommand<{Resource}{Operation}Command>(serviceProvider);

        return root;
    }
}

Also register the toolset in servers/Azure.Mcp.Server/src/Program.cs:

private static IAreaSetup[] RegisterAreas()
{
    return [
        // ... existing toolsets (alphabetical order) ...
        new Azure.Mcp.Tools.{Toolset}.{Toolset}Setup(),
        // ... more toolsets ...
    ];
}

The RegisterAreas() list must remain alphabetically sorted (excluding the #if !BUILD_NATIVE block).

Command group naming: concatenated lowercase or dash-separated. Never underscores.

• ✅ Good: "entraadmin", "resourcegroup", "storageaccount", "entra-admin"
• ❌ Bad: "entra_admin", "resource_group", "storage_account"

Command hierarchy patterns and anti-patterns:

• ✅ Good: azmcp postgres server param set (command groups: server → param, operation: set)
• ❌ Bad: azmcp postgres server setparam (mixed operation setparam at same level)
• ✅ Good: azmcp storage blob upload permission set
• ❌ Bad: azmcp storage blobupload

This pattern improves discoverability and allows grouping related operations.

GATE: dotnet build tools/Azure.Mcp.Tools.{Toolset}/src must pass with 0 errors.

Phase 2: Unit Tests

File: tests/Azure.Mcp.Tools.{Toolset}.Tests/{Resource}/{Resource}{Operation}CommandTests.cs

Required test methods

using System.Net;
using Azure.Mcp.Core.Services.Azure;
using Azure.Mcp.Tests.Commands;
using Azure.Mcp.Tools.{Toolset}.Commands;
using Azure.Mcp.Tools.{Toolset}.Commands.{Resource};
using Azure.Mcp.Tools.{Toolset}.Models;
using Azure.Mcp.Tools.{Toolset}.Services;
using Microsoft.Mcp.Core.Options;
using NSubstitute;
using NSubstitute.ExceptionExtensions;
using Xunit;

namespace Azure.Mcp.Tools.{Toolset}.Tests.{Resource};

public class {Resource}{Operation}CommandTests
    : SubscriptionCommandUnitTestsBase<{Resource}{Operation}Command, I{Toolset}Service>
{
    [Fact]
    public void Constructor_InitializesCommandCorrectly()
    {
        var command = Command.GetCommand();
        Assert.Equal("operation", command.Name);
        Assert.NotNull(command.Description);
        Assert.NotEmpty(command.Description);
    }

    [Theory]
    [InlineData("--my-option val --subscription sub123", true)]
    [InlineData("--subscription sub123", true)]  // my-option is optional
    [InlineData("", false)]  // missing args
    public async Task ExecuteAsync_ValidatesInputCorrectly(string args, bool shouldSucceed)
    {
        if (shouldSucceed)
        {
            Service.GetResourcesAsync(
                Arg.Any<string?>(),
                Arg.Any<string>(),
                Arg.Any<string?>(),
                Arg.Any<string?>(),
                Arg.Any<CancellationToken>())
                .Returns(new ResourceQueryResults<MyModel>([], false));
        }

        var response = await ExecuteCommandAsync(args);

        Assert.Equal(shouldSucceed ? HttpStatusCode.OK : HttpStatusCode.BadRequest, response.Status);
        if (!shouldSucceed)
            Assert.Contains("required", response.Message.ToLower());
    }

    [Fact]
    public async Task ExecuteAsync_DeserializationValidation()
    {
        Service.GetResourcesAsync(
            Arg.Any<string?>(),
            Arg.Any<string>(),
            Arg.Any<string?>(),
            Arg.Any<string?>(),
            Arg.Any<CancellationToken>())
            .Returns(new ResourceQueryResults<MyModel>([], false));

        var response = await ExecuteCommandAsync("--subscription", "sub123");

        var result = ValidateAndDeserializeResponse(
            response, {Toolset}JsonContext.Default.{Resource}{Operation}CommandResult);
        Assert.Empty(result.Items);
    }

    [Fact]
    public async Task ExecuteAsync_HandlesServiceErrors()
    {
        Service.GetResourcesAsync(
            Arg.Any<string?>(),
            Arg.Any<string>(),
            Arg.Any<string?>(),
            Arg.Any<string?>(),
            Arg.Any<CancellationToken>())
            .ThrowsAsync(new Exception("Test error"));

        var response = await ExecuteCommandAsync("--subscription", "sub123", "--my-option", "val");

        Assert.Equal(HttpStatusCode.InternalServerError, response.Status);
        Assert.Contains("Test error", response.Message);
        Assert.Contains("troubleshooting", response.Message);
    }

    [Fact]
    public async Task ExecuteAsync_HandlesNotFound()
    {
        Service.GetResourcesAsync(
            Arg.Any<string?>(),
            Arg.Any<string>(),
            Arg.Any<string?>(),
            Arg.Any<string?>(),
            Arg.Any<CancellationToken>())
            .ThrowsAsync(new RequestFailedException((int)HttpStatusCode.NotFound, "Resource not found"));

        var response = await ExecuteCommandAsync("--subscription", "sub123", "--my-option", "val");

        Assert.Equal(HttpStatusCode.NotFound, response.Status);
        Assert.Contains("Resource not found", response.Message);
    }
}

Critical: Choose the correct test base class:

• Commands extending SubscriptionCommand → use SubscriptionCommandUnitTestsBase<TCommand, TService>
• Commands extending BaseCommand directly (no subscription) → use CommandUnitTestsBase<TCommand, TService>

Using the wrong base class will cause DI failures.

Prefer string args over constructing options directly. Using ExecuteCommandAsync("--account", ...) tests the full pipeline: [Option] attribute registration, OptionBinder parsing, and SubscriptionResolver post-processing.

Mock rules:

• Use Arg.Any<CancellationToken>() for CancellationToken in mocks
• Use TestContext.Current.CancellationToken when invoking real code
• Use Arg.Is(value) or the value directly for specific match assertions
• Never pass CancellationToken.None or default in test code

Deserialization rules:

• Use {Toolset}JsonContext.Default.{Operation}CommandResult for deserialization — never define custom test models
  • ✅ ValidateAndDeserializeResponse(response, {Toolset}JsonContext.Default.{Operation}CommandResult)
  • ❌ JsonSerializer.Deserialize<TestModel>(json)

GATE: dotnet test tools/Azure.Mcp.Tools.{Toolset}/tests --filter "FullyQualifiedName~{Resource}{Operation}CommandTests" must pass.

Phase 3: Live Tests (Azure service commands only)

Skip this phase for non-Azure commands (CLI wrappers, best practices, documentation tools).

3a. Test Infrastructure

File: tests/test-resources.bicep

targetScope = 'resourceGroup'

@minLength(3)
@maxLength(17)
param baseName string = resourceGroup().name

param testApplicationOid string = deployer().objectId
param location string = resourceGroup().location

resource myResource 'Microsoft.{Provider}/{type}@{api-version}' = {
  name: baseName
  location: location
  properties: { /* minimal config */ }
}

resource roleAssignment 'Microsoft.Authorization/roleAssignments@2022-04-01' = {
  name: guid(roleDefinition.id, testApplicationOid, myResource.id)
  scope: myResource
  properties: {
    principalId: testApplicationOid
    roleDefinitionId: roleDefinition.id
  }
}

output resourceName string = myResource.name

File: tests/test-resources-post.ps1 (required even if empty logic)

[CmdletBinding()]
param (
    [Parameter(Mandatory)] [hashtable] $DeploymentOutputs,
    [Parameter(Mandatory)] [hashtable] $AdditionalParameters
)
Write-Host "{Toolset} post-deployment setup completed."

Validate: az bicep build --file tools/Azure.Mcp.Tools.{Toolset}/tests/test-resources.bicep

3b. Live Test Class

File: tests/Azure.Mcp.Tools.{Toolset}.Tests/{Toolset}CommandTests.cs

public class {Toolset}CommandTests(ITestOutputHelper output, TestProxyFixture fixture, LiveServerFixture liveServerFixture)
    : RecordedCommandTestsBase(output, fixture, liveServerFixture)
{
    [Fact]
    public async Task {Resource}{Operation}_ReturnsExpectedResult()
    {
        var result = await CallToolAsync(
            "{toolset}_{resource}_{operation}",
            new()
            {
                ["subscription"] = SubscriptionId,
                ["resource-group"] = ResourceGroupName,
            });

        Assert.NotNull(result);
        var items = result.Value.AssertProperty("items");
        Assert.Equal(JsonValueKind.Array, items.ValueKind);
    }
}

LocalRequired tools

Remote HTTP mode intentionally excludes tools marked LocalRequired = true. Every recorded test for such a tool must verify that exclusion and return before exercising local-only behavior:

if (await AssertLocalToolIsUnavailableInHttpMode("{toolset}_{resource}_{operation}"))
{
    return;
}

Use the inherited helper in every applicable test in a class extending RecordedCommandTestsBase; do not duplicate the transport check or unavailable-tool assertions.

3c. Record and Verify

Create assets.json

Create assets.json if it doesn't exist:

{
  "AssetsRepo": "Azure/azure-sdk-assets",
  "AssetsRepoPrefixPath": "",
  "TagPrefix": "Azure.Mcp.Tools.{Toolset}.Tests",
  "Tag": ""
}

Deploy

eng/common/TestResources/New-TestResources.ps1 `
  -TestResourcesDirectory tools/Azure.Mcp.Tools.{Toolset}

Record tests

dotnet test tools\Azure.Mcp.Tools.{Toolset}\tests\Azure.Mcp.Tools.{Toolset}.Tests `
  --filter "FullyQualifiedName~{Resource}{Operation}"

Push recordings

.proxy\Azure.Sdk.Tools.TestProxy push `
 -a tools\Azure.Mcp.Tools.{Toolset}\tests\Azure.Mcp.Tools.{Toolset}.Tests\assets.json

Verify playback

Change TestMode to "Playback" in .testsettings.json, then re-run tests

3c-1. Recorded Test Pitfalls

These are common causes of recorded test failures. Always verify playback passes after recording.

Always pass Settings.TenantId in live test calls

If the test subscription lives in a non-default tenant, the command will fail with InvalidAuthenticationTokenTenant. Include tenant when your subscription requires it:

var result = await CallToolAsync(
    "{toolset}_{resource}_{operation}",
    new()
    {
        { "subscription", Settings.SubscriptionId },
        { "resource-group", Settings.ResourceGroupName },
        { "tenant", Settings.TenantId }  // Always include
    });

Use RegisterOrRetrieveVariable for all dynamic values

Any non-deterministic value (Guid.NewGuid(), DateTime.Now) must be wrapped so the same value is used in both Record and Playback runs:

// ✅ Value is recorded and replayed deterministically
var topicName = RegisterOrRetrieveVariable("create_topic_name", $"topic-{Guid.NewGuid():N}"[..24]);

// ❌ Different GUID each run — breaks playback request matching
var topicName = $"topic-{Guid.NewGuid():N}"[..24];

Assertions must survive sanitization

Recording sanitizers replace sensitive values (resource names, IDs, endpoints) with placeholders like "Sanitized". Your assertion strategy depends on your test class sanitizer configuration:

How to check: After recording, inspect the session recording JSON (use .proxy/Azure.Sdk.Tools.TestProxy.exe config locate -a <assets.json>). If the "name" field shows "Sanitized", you cannot use exact name asserts without the SanitizeAndRecord pattern.

// Safe assertions that survive any sanitizer configuration:
topic.AssertProperty("name");  // Checks existence only
Assert.Equal("Succeeded", topic.GetProperty("provisioningState").GetString());  // Enum values aren't sanitized
Assert.Equal(JsonValueKind.Object, topic.ValueKind);  // Type checks

Credential type in .testsettings.json

Deploy-TestResources.ps1 sets AZURE_TOKEN_CREDENTIALS=AzurePowerShellCredential. If the MCP server subprocess cannot access the PowerShell credential cache (common on some machines), switch to AzureCliCredential:

"EnvironmentVariables": {
    "AZURE_TOKEN_CREDENTIALS": "AzureCliCredential"
}

Ensure az login --tenant <tenant-id> is active. If recording fails with credential errors from the subprocess, this is the likely fix.

Redeploy if resource group is missing

Test resource groups are auto-deleted after 12 hours. If tests fail with ResourceGroupNotFound, redeploy:

./eng/scripts/Deploy-TestResources.ps1 -Paths {Toolset}

3d. Test Project Configuration (Critical)

The test

…(truncated)