@metaharness/example-azure

v0.2.0

Published

a month ago

MetaHarness example — Microsoft Azure scaffold: resource management, Blob storage, and Azure OpenAI via DefaultAzureCredential (npx @metaharness/example-azure my-bot)

0High
0Medium
0Low

ruvnet

metaharness azure microsoft-azure azure-sdk azure-openai arm-resources blob-storage default-azure-credential ai-agent mcp npx scaffold

@metaharness/example-azure

MetaHarness scaffold for Microsoft Azure — resource management, Blob storage, and Azure OpenAI via DefaultAzureCredential

Illustrative output. Code samples and example transcripts shown in this README are representative — actual output depends on your Azure subscription, resource configuration, deployed models, and environment. Run the commands to see real results.

Intro

@metaharness/example-azure scaffolds a ready-to-run multi-agent harness pre-wired to the Microsoft Azure JavaScript SDK. One command gives you a project directory containing specialized agents, a scoped MCP policy, a /azure slash command, and safe defaults that prevent accidental mutations to your Azure environment.

What it is: a development scaffold and learning tool. It demonstrates how to structure an agent harness that drives Azure resource management (@azure/arm-resources), Blob storage (@azure/storage-blob), and Azure OpenAI chat completions (openai + @azure/openai) using DefaultAzureCredential from @azure/identity.

What it is not: a production-hardened, compliance-certified, or security-audited deployment. It is not HIPAA-compliant, PCI-DSS-certified, or FedRAMP-authorized out of the box. Conduct your own security review before using agent-driven Azure automation in any regulated or production context. This scaffold does NOT replace a qualified security architect or cloud compliance program.

Features

| MetaHarness capability | How this example shows it | |---|---| | Tiered model routing | Haiku for fan-out planning and verification re-checks; Sonnet for executor reasoning; Opus only for mutation approval | | MCP default-deny | .harness/mcp-policy.json grants only 6 read tools + 2 write tools (write tools require allow_mutations flag); all calls logged to .harness/mcp-audit.jsonl | | Slash command | /azure <goal> routes to planner → executor → verifier | | Specialized agents | planner (Haiku), executor (Sonnet), verifier (Haiku) | | Verification gate | Verifier re-lists resource groups and blob containers after executor runs, cross-checks counts and names before reporting done | | Read-only by default | Mutation tools (create resource group, upload blob) require --allow-mutations flag AND explicit confirmation |

Azure-specific capabilities demonstrated

List all resource groups in a subscription with location, tags, and provisioning state
List blob containers in a storage account with access tier and public-access settings
Send chat completions through an Azure OpenAI deployment using Entra ID keyless auth
Dry-run mode: executor logs what it would do for any write operation when mutations are disabled
Local Azurite integration: blob operations route to the local emulator when AZURE_STORAGE_CONNECTION_STRING=UseDevelopmentStorage=true

Quickstart

npx @metaharness/example-azure@latest my-bot
cd my-bot && npm install && npm run doctor

npm run doctor checks that DefaultAzureCredential can acquire a token, that AZURE_SUBSCRIPTION_ID is set, and that the MCP policy file is present and valid.

Scaffold on a specific host

# Default: Claude Code
npx @metaharness/example-azure@latest my-bot

# GitHub Actions workflow
npx @metaharness/example-azure@latest my-bot --host github-actions

# All hosts at once
npx @metaharness/example-azure@latest my-bot --host all

Supported hosts: claude-code, codex, copilot, github-actions, hermes, openclaw, opencode, pi-dev, rvm. Host wiring delegates to @metaharness/host-<id>.

Configuration

Required environment variables

Copy .env.example to .env (already in .gitignore) and fill in your values:

# Azure identity — service principal (used by EnvironmentCredential)
AZURE_TENANT_ID=your-entra-tenant-id
AZURE_CLIENT_ID=your-app-registration-client-id
AZURE_CLIENT_SECRET=your-client-secret

# Target subscription
AZURE_SUBSCRIPTION_ID=your-subscription-id

# Blob storage
AZURE_STORAGE_ACCOUNT_NAME=yourstorageaccount
# For local Azurite development (overrides real storage):
# AZURE_STORAGE_CONNECTION_STRING=UseDevelopmentStorage=true

# Azure OpenAI
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_DEPLOYMENT=gpt-4o
# API version (defaults to 2024-10-21 if unset)
AZURE_OPENAI_API_VERSION=2024-10-21

How to get credentials

Azure subscription: Sign up at https://azure.microsoft.com/free/ (free account includes $200 credit).
Service principal: az ad sp create-for-rbac --name my-harness-sp --role Reader --scopes /subscriptions/<id> — copy appId (AZURE_CLIENT_ID), tenant (AZURE_TENANT_ID), password (AZURE_CLIENT_SECRET).
Storage Blob Data Reader role: az role assignment create --assignee <clientId> --role "Storage Blob Data Reader" --scope /subscriptions/<id>/resourceGroups/<rg>/providers/Microsoft.Storage/storageAccounts/<account>.
Cognitive Services OpenAI User role: az role assignment create --assignee <clientId> --role "Cognitive Services OpenAI User" --scope /subscriptions/<id>/resourceGroups/<rg>/providers/Microsoft.CognitiveServices/accounts/<resource>.
Azure OpenAI endpoint: Azure portal > your OpenAI resource > Keys and Endpoint. The endpoint format is https://<resource-name>.openai.azure.com/.

Local development without real Azure (Azurite)

npm run azurite          # starts Azurite on localhost:10000/10001/10002
# Then set in .env:
# AZURE_STORAGE_CONNECTION_STRING=UseDevelopmentStorage=true

The harness detects UseDevelopmentStorage=true and routes all blob operations to Azurite automatically. You still need a real Azure subscription for ARM resource listing and Azure OpenAI calls.

Developer auth (no service principal needed)

az login                 # or: azd auth login
# Then unset AZURE_CLIENT_ID/AZURE_TENANT_ID/AZURE_CLIENT_SECRET —
# DefaultAzureCredential falls through to AzureCliCredential automatically.

Usage

Slash command

/azure list all resource groups in eastus
/azure audit storage account mystorage and show public-access containers
/azure summarise Azure OpenAI usage and estimate token spend

The command routes: planner (Haiku, decomposes goal) → executor (Sonnet, drives SDK) → verifier (Haiku, re-checks against live platform) → result.

Natural-language prompts

"List every resource group in my subscription with its location and tags."

"Check my storage account 'mystorage' for containers with public blob access
 enabled and report them as a potential misconfiguration."

"Ask the gpt-4o deployment: what are the top 3 Azure security best practices
 for a startup? Use keyless Entra ID auth."

Enabling mutations (opt-in)

The harness defaults to read-only. To allow writes, set the flag at scaffold time:

npx @metaharness/example-azure@latest my-bot --allow-mutations

Or set HARNESS_ALLOW_MUTATIONS=true in .env after scaffolding. With mutations enabled, the executor will:

Generate a dry-run plan (logged to .harness/mcp-audit.jsonl)
Present the plan and ask for explicit confirmation
Only then call the write SDK method (e.g. resourceGroups.createOrUpdate, BlockBlobClient.upload)
Verifier re-checks the result via a read-back call

Safety

Secrets via ENV only: no credentials are ever written to scaffolded files. .env.example contains only placeholder values. .gitignore covers .env, *.pem, *.pfx, and .harness/mcp-audit.jsonl.
Read-only by default: mutation MCP tools (azure.resourceGroups.createOrUpdate, azure.blobs.uploadBlob) are granted in the policy but carry requires_flag: allow_mutations. The harness sets HARNESS_ALLOW_MUTATIONS=false by default.
No Azure sandbox: Azure has no universal test-key mode. Read-only RBAC roles (Reader, Storage Blob Data Reader) prevent mutations at the platform level even if code attempts them. For local blob development, use Azurite.
MCP audit log: every tool call is appended to .harness/mcp-audit.jsonl with timestamp, tool name, calling agent, flag state, and whether it was allowed or denied.
Verification gate: the verifier agent always re-checks output against the live API before the harness reports done. Discrepancies are flagged, not silently dropped.
Not for production / not certified: this scaffold is a developer starting point. It is not a security-hardened deployment and carries no compliance certifications (HIPAA, PCI-DSS, FedRAMP, ISO 27001, SOC 2). Azure platform certifications apply to Azure itself, not to code you build on top of it.

How it works

Agents

planner (Haiku)
  Receives: natural-language goal
  Emits: structured task list (which subscription, resource groups,
         storage account, OpenAI deployment, what to check)

executor (Sonnet)
  Receives: task list from planner
  Drives: ResourceManagementClient.resourceGroups.list()
          BlobServiceClient.listContainers()
          AzureOpenAI.chat.completions.create()
  Applies: mutation gate (dry-run if allow_mutations=false)
  Emits: raw SDK results + plan confirmation for writes

verifier (Haiku)
  Receives: executor's reported results
  Re-calls: same list APIs independently
  Cross-checks: counts, names, container settings
  Emits: verified result or discrepancy report

Routing tiers

| Tier | Model | Task | |---|---|---| | 1 | WASM booster | Template fills — no LLM | | 2 | Haiku | Planner decomposition; verifier re-check | | 3 | Sonnet | Executor: SDK calls, policy reasoning, error handling | | 3+ | Opus | Mutation approval (escalated only when allow_mutations=true) |

MCP policy (granted tools)

Defined in .harness/mcp-policy.json (default-deny):

| Tool | Tier | Mutation flag required | |---|---|---| | azure.resourceGroups.list | 2 | No | | azure.resourceGroups.get | 2 | No | | azure.subscriptions.list | 2 | No | | azure.blobs.listContainers | 2 | No | | azure.blobs.listBlobs | 2 | No | | azure.openai.chatComplete | 3 | No | | azure.resourceGroups.createOrUpdate | 3 | Yes (allow_mutations) | | azure.blobs.uploadBlob | 3 | Yes (allow_mutations) |

All other tools: denied. All calls audited to .harness/mcp-audit.jsonl.

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

@metaharness/example-azure

Intro

Features

Azure-specific capabilities demonstrated

Quickstart

Scaffold on a specific host

Configuration

Required environment variables

How to get credentials

Local development without real Azure (Azurite)

Developer auth (no service principal needed)

Usage

Slash command

Natural-language prompts

Enabling mutations (opt-in)

Safety

How it works

Agents

Routing tiers

MCP policy (granted tools)

Links