Maester MCP Server

An MCP (Model Context Protocol) server that exposes Maester 2.2 — the Microsoft 365 / Entra / Defender / Active Directory / GitHub security test framework — as tools for AI agents.

Connect Claude, GitHub Copilot, or any MCP-compatible agent to your tenant's security posture and ask, for example:

"Run Maester, summarise the critical failures, and audit our Copilot Studio agents and MCP servers while you're at it."

Version note: aligned to Maester 2.2.0 (July 2026). Maester's test count has grown past 600 across Entra, Exchange, Teams, SharePoint, Purview, Defender, Intune, Global Secure Access, Active Directory, GitHub, and Azure DevOps. This server tracks the 2.2 cmdlet surface (Get-MtTestInventory, Merge-MtMaesterResult, Get-MtHtmlReport, -OutputMarkdownFile, -DriftRoot, the -Service connection model, and the AIAgent tag).

Where this fits alongside Maester Cloud

Maester now has a companion commercial product, Maester Cloud (private preview Aug 2026, GA targeted Sep 2026): a portal that keeps 5+ years of tenant history, highlights drift between runs, and alerts on change. It is the durable evidence layer.

This MCP server is a different, complementary layer:

Use them together: Maester Cloud remembers and alerts; the MCP server lets an agent investigate, explain, and act on any given run in natural language. They are not substitutes.

Related MCP server: m365-audit-mcp

Prerequisites

Auth setup

Fastest path — let Maester create the app

Connect-Maester -Service Azure
New-MtMaesterApp -GitHubActions -SetGitHubSecrets   # zero-config for GitHub Actions

Manual app registration — Graph permissions (application, admin-consented)

Copilot Studio (audit_ai_agents) additionally needs the Dataverse service connection and access to the environment named in maester-config.json.

Managed Identity (Azure-hosted, no secrets)

Set AZURE_USE_MANAGED_IDENTITY=true. Grant the managed identity the same Graph permissions.

Installation

git clone https://github.com/iffystrayer/maester-mcp-server
cd maester-mcp-server
npm install

Configuration

* Not required when AZURE_USE_MANAGED_IDENTITY=true.

Claude Desktop config

~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "maester": {
      "command": "node",
      "args": ["/absolute/path/to/maester-mcp-server/src/index.js"],
      "env": {
        "AZURE_TENANT_ID": "your-tenant-id",
        "AZURE_CLIENT_ID": "your-client-id",
        "AZURE_CLIENT_SECRET": "your-client-secret",
        "MAESTER_TESTS_PATH": "/Users/you/maester-tests",
        "MAESTER_DRIFT_ROOT": "/Users/you/maester-baseline"
      }
    }
  }
}

Restart Claude Desktop after saving.

Available tools

The reflexive angle: agents auditing agents

The most valuable new capability in 2.2 for an agentic system is that Maester now audits AI agents and MCP servers as an attack surface. That means the very agent you connect this server to can police the rest of your org's agent sprawl:

You: We've had teams spinning up Copilot Studio agents and MCP servers all
     quarter. Which ones are risky?

Agent: [calls audit_ai_agents]
       → "8 agents checked. 3 findings: 'Sales Helper' is shared with the
          whole org and has no user authentication (MT.1113/MT.1114);
          'Invoice Bot' can send email with AI-controlled inputs — an
          exfiltration path (MT.1116); one connected MCP server exposes
          tools flagged for review (MT.1120). Want the remediation steps?"

This closes a real governance gap that most orgs don't have tooling for yet.

Operational runbook

1. Establish a baseline (once). Run a full pass and snapshot the JSON as your drift baseline:

run_maester_tests(services: ["Graph","Exchange","Teams"]) → save resultsFile into MAESTER_DRIFT_ROOT

2. Daily scheduled run (CI/CD). Keep this in GitHub Actions / Azure DevOps, not the MCP server — the MCP server is for interactive investigation. Use New-MtMaesterApp -GitHubActions -SetGitHubSecrets for a secretless OIDC setup. Emit markdown for the PR/summary and JSON for retention.

3. Interactive triage (MCP server). When the scheduled run flags something, or ad hoc, have the agent call run_maester_tests with useDrift: true, then get_security_posture_summary, then drill in with get_test_detail.

4. AI-surface review (monthly). audit_ai_agents on a schedule as Copilot Studio and MCP adoption grows.

5. Evidence retention. For history/drift/alerting across runs, feed results into Maester Cloud rather than reinventing storage here.

Security notes

• Read-only by default. Maester tests only read Graph/Azure/AD data.

• Least privilege. Grant only the scopes listed; don't run as Global Admin.

• Secrets in env or a vault, never in code. Prefer Managed Identity where possible.

• Result files can contain tenant configuration detail — store and retain them securely.

• The run_ca_whatif and remediation-adjacent flows are advisory; keep a human in the loop before any tenant write.

Extending

Add an entry to TOOLS and a matching case in the request handler. Custom Maester tests in tests/Custom/ are picked up automatically by run_maester_tests — no server change needed.

Roadmap ideas

• HTTP+SSE / Streamable HTTP transport for remote deployment

• Streaming progress during long runs

• Optional Graph write tools for one-click remediation (behind an explicit approval gate)

• Auto-ticketing (Jira / Azure DevOps Boards / ServiceNow) on new failures

• Native Maester Cloud push once its ingest cmdlet ships