Ask Your AI Assistant About Your Microsoft 365 CIS Posture (MCP)

A CIS Microsoft 365 assessment produces a lot of structured output: a posture score, a result for each control, and evidence down to individual user accounts. The hard part is rarely generating it. It is getting the answer to a specific question out of it at the moment someone asks — in a change meeting, an audit call, or a Monday morning triage.

ConfigCobra includes a native MCP (Model Context Protocol) server for exactly that. This post explains what MCP is in practical terms, what an assistant can and cannot answer once it is connected to your scan data, and why the data source matters more than the assistant.

What MCP is, without the hype

Model Context Protocol is an open protocol that lets an AI assistant call tools exposed by an external server, instead of answering only from whatever it learned in training or whatever you paste into the chat. The assistant discovers what the server offers, decides when a question needs it, and uses the returned data in its answer.

For compliance work the consequence is simple. Ask a general-purpose assistant “are my global admins configured safely?” and the best it can do is describe what good looks like in general. Connect it to a server that holds your tenant’s scan results and the same question is answered about your tenant.

The ConfigCobra MCP server connects ConfigCobra to any MCP-compatible AI assistant, including Claude and Cursor.

What the assistant draws on

The assistant’s answers come from live ConfigCobra scan data, not from a generic description of the benchmark. Three layers of that data are available to ask about:

• Posture score — the overall result of the most recent assessment against the CIS Microsoft 365 Foundations Benchmark v7.0.0.
• Per-control results — the outcome for each of the 160 controls, across the benchmark’s 9 sections, including the severity tagged on each control by the benchmark itself.
• Per-user evidence — the account-level detail behind a finding, which is what turns “a control failed” into “these specific accounts are the reason.”

That third layer is where the interface change is most noticeable. In a dashboard, getting from a failed identity control to the affected accounts means navigating, filtering, and often exporting. In a conversation it is one follow-up question.

Questions worth asking

The useful questions are the ones that cut across the data rather than ones a single dashboard page already answers. A few that map directly onto what ConfigCobra stores:

• “Who are my global admins, and is there anything wrong with how they are set up?” — privileged accounts are the first thing an auditor asks about, and the answer rests on per-user evidence.
• “What is my current posture score, and which findings are dragging it down?” — a briefing for someone who will not log in to a dashboard.
• “Which controls are failing in the identity section?” — Section 5 of the benchmark is Microsoft Entra, and at 63 controls it is the largest of the nine, so it is where most triage questions end up.
• “Which of my failing controls are the highest severity?” — severity comes from the benchmark itself, so the ranking is not something the assistant invents.

None of these require a prompt-engineering trick. They are ordinary questions put to a source that actually holds the answer.

Why the data source matters more than the assistant

There is a failure mode worth naming. An assistant with no grounding will answer a compliance question fluently and confidently regardless of whether it knows the answer. In security work, a confident wrong answer is worse than no answer, because it ends the investigation.

Grounding the assistant in scan results changes what it is doing: it is reading and summarising results that the assessment already produced, rather than estimating them. The assessment is the source of truth; the assistant is the interface to it. If you want to know how a given result was determined, the evidence is in the PDF report and the audit-ready evidence export, which carry the findings, the evidence, and a timestamped tenant snapshot.

The answer is only as fresh as the last scan

“Live scan data” means the data from your most recent scan. A full assessment completes in 20–25 minutes, and the scheduler can run scans on a daily, weekly, or monthly cadence, so how current the assistant’s answers are depends on how often you scan.

That is worth deciding deliberately. If the assistant is going to be used for week-to-week triage, a weekly or daily cadence keeps its answers close to the tenant’s real state. The same data feeds continuous scanning and drift detection, which alerts you when a control regresses — so the two work together: monitoring tells you something changed, the conversation is where you ask what.

What it does not replace

Of the 160 controls, 143 are Automated and 17 are Manual in the benchmark’s own classification. ConfigCobra assesses the vast majority automatically; a small number need a person to verify something the APIs do not expose. An assistant reading scan data does not change that — it summarises results, it does not perform the manual verification.

Nor does a conversation replace the audit artefact. When an auditor needs evidence, the deliverable is still the report and the evidence export. The assistant gets you to the right finding quickly; the report is what you hand over.

Getting connected

Setup for each supported client — the server address, the sign-in flow, and a set of example prompts — is in the AI integration documentation. The MCP server product page covers the use cases by role: IT admin, CISO, MSP, and auditor.

You need scan data for the assistant to have anything to read, so the first step is an assessment. The free trial runs for 14 days with full access, which is enough to see how the data looks in conversation.