m365-evidence-mcp

An MCP server that turns the Microsoft Graph reads an IT auditor keeps repeating into tools any MCP-capable AI client can call. Point it at a tenant with a read-only token and ask for the evidence in plain language instead of clicking through four admin portals.

The four tools cover ITGC classics: security posture, MFA coverage, dormant accounts and privileged access. The model fetches, the auditor concludes.

Tools

Setup

python3 -m venv .venv
.venv/bin/pip install mcp

Get a Graph token. The Azure CLI is the quickest way:

export GRAPH_TOKEN=$(az account get-access-token --resource https://graph.microsoft.com --query accessToken -o tsv)

The Access token tab in Graph Explorer works too. Scopes: SecurityEvents.Read.All for the score, Reports.Read.All for MFA coverage, User.Read.All plus AuditLog.Read.All for stale accounts, Directory.Read.All for roles. All of them are read scopes and most need admin consent.

Wire it into Claude Code:

claude mcp add m365evidence -- /path/to/m365-evidence-mcp/.venv/bin/python /path/to/m365-evidence-mcp/m365_evidence_mcp.py

Demo

What it looks like from Claude Code:

> pull the access review evidence for the quarter

⏺ m365evidence · privileged_roles()

  Global Administrator (2):
    - amir.admin@demo.example
    - breakglass@demo.example
  Helpdesk Administrator (1):
    - sara.support@demo.example

⏺ Two global admins and one of them is the break glass account.
  Helpdesk Administrator has a single holder. MFA coverage next?

The accounts above are demo data. The format is exactly what the server returns.

Design notes

• Read-only by construction. Every tool is a single Graph GET, the annotations declare it, and a token with read scopes could not write anyway. Two layers, both structural.
• Evidence, not judgement. Tools return counts with capped name lists, the kind of thing that goes straight into a workpaper. The conclusion stays with whoever reads it.
• Nothing touches disk. The token lives in an environment variable and reports are never cached or logged.

Honest notes

• Sign-in activity needs Entra ID P1 or better, so stale_accounts comes back empty on a bare Business Standard tenant.
• Pagination is capped at ten pages of 999 objects. Fine for a small or midsize tenant, raise the cap in _get_all for bigger ones.
• The logic is tested against canned Graph responses in test_server.py. Run python test_server.py.

About

Al Amin Bashir Afara, Dubai · github.com/aminafara123 · linkedin.com/in/aminafara