azure-advisor is published by Microsoft from microsoft/azure-skills, pinned to 233c51fc11c73d459be46e099e417ebaf3d20a26. Review the source before installing.
Use Azure Advisor recommendations to understand optimization opportunities, identify resources affected by service retirements, and generate customer-ready remediation guidance.
This plugin includes:
azure-advisorfor recommendation discovery, summaries, affected resources, metadata, service retirements, and lifecycle updates.azure-advisor-actionabilityfor one-recommendation remediation, followed by resource-specific command placeholder resolution. When Advisor returns no usable remediation, it falls back to relevant public Microsoft Learn documentation. Responses identify whether guidance comes from Advisor or Microsoft Learn and cite documentation used for the fallback.azure-advisor-cost-optimizationfor scope resolution,Cost/IdleResourcesdiscovery, prioritized opportunities, full retail USD savings reports, historical/lifecycle reporting, confirmed state changes, and explicitly selected remediation handoffs. Known subscription or Service Group names/IDs pass directly to the existing Advisor tools. All-accessible subscriptions are enumerated and grouped by tenant; no separate scope skill or SG-discovery tool is needed. Capped resource lists are never treated as complete totals.
Generic Cost and other Advisor categories remain on azure-advisor. Cost reports explicitly show
the tool-returned pricing basis and measure currency, currently Retail USD, rather than hardcoding
the label or inventing contractual figures or availability lines. Advisor findings do not prove
deletion safety. Explicit safety requests use the connected Advisor read-only prevalidation
operation when available; passing automated checks never authorizes deletion.
Explicit prevalidation selections use full affected-resource ARM IDs only, reusing already
returned IDs or resolving names through list when needed. Type-wide prevalidation follows
discovery pages and processes bounded batches instead of stopping at 25 or 100 resources;
requested subsets and incomplete coverage remain visible.
New summary status overrides, historical dates and minimum-savings require both Cost and
IdleResources; generic list status and summary status grouping retain their existing behavior.
Lifecycle tables label supported lastUpdated as State last changed, separately from
resource deletion dates. Cost/IdleResources never substitutes legacy money for missing
structured savings. Generic Advisor does not inherit the specialized 15-row HTML layout.
Response parity checks
Cost/IdleResources uses the same shared response templates on every model. The existing eval
suite tags its supplied-result and selection cases with parity=response-shape, covering
opening discovery, unavailable pricing, selected resources, Completed details, period choices,
inline pagination, compact prevalidation evidence, postpone-date selection before the batch proposal,
and partial update results. These cases do not require live Azure recommendation queries.
Prevalidation uses compact tables for all three returned buckets and one shared
Check | Outcome | Minimal detail table, with failed/unverified checks and outstanding manual confirmations visible. Detailed
per-resource evidence is shown only on request. State changes use one compact proposal with exact IDs
separately available, followed by one server destructive approval covering the action and chosen
date/reason. Postpone asks the date after resource selection; it does not add a separate
assistant confirmation before the server approval. Batch updates reuse the existing tool's
recommendation-ids input and require one shared scope, tenant, state and date/reason. Completed batches
use a per-record result table that distinguishes success, failure and unverified status. Postpone/dismiss
proposals and results show both Yearly potential savings and accumulated Cost, savings first,
using the actual returned values without recalculation. Results identify carried-forward
amounts as previously returned estimates, not savings achieved by the state change.
The existing update response is authoritative for success and returned status; a later
stale or empty list must not downgrade a successful update.
From tests, run the same cases against both model families:
$env:VALLY_RUNNER_DISABLE_AZURE_MCP = "true"
$env:VALLY_RUNNER_EXACT_SKILL = "true"
npm run test:vally -- --plugin azure-advisor --skill azure-advisor-cost-optimization --tag parity=response-shape --model gpt-5.6-sol --runs 1 --workers 1 --threshold 1 --pass-rate 1 --require-pass
npm run test:vally -- --plugin azure-advisor --skill azure-advisor-cost-optimization --tag parity=response-shape --model claude-sonnet-5 --runs 1 --workers 1 --threshold 1 --pass-rate 1 --require-pass
Require both runs to pass; a shared failure is not parity success. The strict thresholds prevent missing required fields from being averaged away. Rerun after template, model or client changes; passing checks establish the tested structure/data contract, not identical wording or guaranteed rendering in every app. Local runs need the supported Node runtime, restored test dependencies and Copilot authentication; inspect executor errors separately from model-format failures.
The period-choice and postpone-date cases expose ask_user and stop at the pending question
without answering or querying Azure. Proposal-only cases never request an update, and
cancellation uses an ordinary follow-up turn:
npm run test:vally -- --plugin azure-advisor --skill azure-advisor-cost-optimization --tag area=batch-confirmation-regression --runs 1 --workers 1 --threshold 1 --pass-rate 1 --require-pass
The batch-postpone-confirmation cases additionally cover date selection and cancellation;
batch-approval-reuse is a no-write planning check of one selected batch call. These cases prohibit
updates; they do not verify post-confirmation batch execution. That flow requires separate
validation against dedicated test recommendations,
with approval before writes. Tool-call assertions use the default namespace routers,
checking the requested operation and its parameters rather than forcing individual-tool mode.
Security
[!WARNING] The
azure-advisorplugin usesnpxto download and run the Azure MCP Server, inheriting the local environment's.npmrcconfiguration. Install this plugin only on trusted devices. A compromised.npmrcconfiguration could causenpxto download and execute malicious code, potentially resulting in remote code execution.
Telemetry
The shared telemetry hook records skill and MCP tool usage. Set AZURE_MCP_COLLECT_TELEMETRY=false to opt out.
Setup
Authenticate with Azure CLI or another Azure Identity credential supported by Azure MCP, then install the plugin from the marketplace.