ARM MCP server moves from reading to writing in Azure – but remains in public preview

Microsoft has expanded the Azure Resource Manager MCP server at Ignite 2026, so that AI agents can no longer just run read queries and deploy templates, but can create, update, and take actions on Azure resources directly.

Illustration: an industrial control panel in gray aluminum with one red switch flipped on, symbolizing AI agents moving from reading to writing in Azure.
Illustration

ARM MCP server moves from reading to writing in Azure – but remains in public preview

Microsoft has expanded the Azure Resource Manager MCP server at Ignite 2026, so that AI agents can no longer just run read queries and deploy templates, but can create, update, and take actions on Azure resources directly. Every write operation runs under the signed-in user's full Azure permissions — and the server is still officially in public preview.

What's new

When Microsoft first launched the Azure Resource Manager (ARM) MCP server in May 2026, it was limited to two things: read access across more than 3,500 resource types, and the ability to kick off deployments of ARM templates. At Ignite 2026, which runs through Friday at Moscone Center in San Francisco, Microsoft has significantly expanded the toolbox, according to TechTimes' coverage of the announcement.

A November 13 post on Microsoft Tech Community documents three new write tools. The first creates or updates individual resources. The second handles resource groups. Both are built on the ARM PUT pattern — the standard mechanism Azure Resource Manager uses to create and modify resources. The third tool runs POST-style actions against resources, with start, restart, key retrieval, and validation listed as examples.

The deployment tools have also been expanded: they now add what-if analysis previewing, start, status checks, and cancellation — that is, more than just firing off a template and hoping for the best.

The permissions model: the agent inherits the user

The central point for IT teams is how permissions work. The server runs tool calls under the signed-in user's Azure permissions. That means an IT department that connects an agent through the server is, in practice, giving the agent every change authority that user has. A developer with contributor rights on a production subscription hands the agent the same rights — without anyone configuring them explicitly for the agent.

This is not a separate agent identity with narrowed permissions, but full inheritance from the user's profile. For organizations piloting agentic Azure access — exactly what many are doing during Ignite week — that means the risk of a mistake or a manipulated agent decision is, in principle, as large as the connected user's own power.

Still in preview — with a labeling mystery

Microsoft Learn's documentation index still groups the server under preview headings, and the tool reference states that tool schemas can change during the public preview. The server's initial announcement in May 2026 was also a preview, and the November 13 post says tool names and behavior may evolve as the preview progresses.

Interestingly, however, the main Ignite post about agents gives the server no status label at all. This is a reported observation, not a resolved contradiction — Microsoft has not commented on why the Ignite posts omit the preview label while Microsoft Learn retains it.

Experimental skills and what's missing

The November 13 post lists twelve skills that are experimental and off by default, including one for authoring Azure Policy. Notably, the Microsoft Learn tool reference mentions no delete tool — since deletion is the most destructive write operation, it's worth noting that it is not documented as an available tool.

But Microsoft has also declined to provide two key numbers. First: how many resource types accept writes? The 3,500+ figure covers read access only, and the November 13 post gives no resource count for write operations. Second: the tool reference describes the action tool as running any action a specific resource supports, without listing which actions are permitted. The full scope of agent-initiated actions is therefore unknown.

The context

The announcement comes mid-Ignite, where a hands-on workshop on governing estates with Azure Arc and the ARM MCP server started at 1:45 PM ET today, with a final run at 12:00 PM ET on Friday. It also comes just days after a Microsoft 365 outage on November 12 that Microsoft attributed to a network upgrade problem — not to AI agents. The timing is worth noting: Microsoft is expanding agentic write access to production-adjacent tooling in a week when questions about infrastructure governance are highly current.

Open questions

For teams considering connecting agents to Azure, several questions remain unanswered. How large is the actual write scope, measured in resource types? Which POST actions are explicitly allowed, and is there any blocklist? How should the permissions model be handled in practice when the agent's power is a direct function of the user's? And why did the preview label disappear from the Ignite posts while Microsoft Learn keeps it? Microsoft has not answered any of this, and as documented: the tools are still in preview, the schemas can change, and every write runs with the user's full authority.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.