Microsoft 365 Admin

ms-365-admin-mcp-server

A Model Context Protocol (MCP) server for Microsoft 365 administration via Graph API application permissions (client credentials).

Built on the architecture and endpoint-driven design pioneered by Softeria/ms-365-mcp-server, and complementary to it: Softeria's server uses delegated permissions for end-user productivity scenarios, while this one uses application permissions for admin operations — security monitoring, identity audits, incident response, and service health. See Acknowledgments below.

Features

• 674 tools covering security, audit, identity, app credentials, guest users, Exchange, Intune (devices, apps, MAM, reports, Settings catalog & Endpoint security (BitLocker, Antivirus, Firewall, ASR), encryption report, Windows Update profiles, BitLocker and LAPS rotation, macOS Platform Scripts, macOS custom attribute scripts, assignment filters, Remediations, Windows PowerShell scripts, custom compliance scripts), governance (PIM, access reviews, entitlement, lifecycle), compliance, threat intelligence, advanced hunting, Defender for Identity (sensors, candidates, migration, identity accounts, audit policy), Microsoft 365 Copilot admin (usage reports, interaction history audit, AI users, meeting insights, agent registrations, policy settings), custom security attributes, LAPS, policies, reports, incident response, eDiscovery v3 (cases, custodians, noncustodial data sources, review sets, queries, exports, operations), Purview DSPM (protection scopes), event-based retention triggers, Teams online meeting attendance reports (app-only with Application Access Policy), deleted chats restore (admin recovery flow), Teams chat investigation reads (Chat.Read.All for triage; eDiscovery v3 for court-admissible production), Cloud PC, call records, Universal Print, information protection, SharePoint admin, and records management
• Application permissions (client credentials) — no user interaction required
• Read-only by default — write operations require explicit --allow-writes
• Risk classification on write tools (low/medium/high/critical)
• Presets to filter tools by domain (security, audit, identity, etc.)
• Two transports: stdio (default) and HTTP (StreamableHTTP)
• Multi-cloud: Microsoft global and China (21Vianet)
• Key Vault support for secrets management

Documentation

Prerequisites

• Node.js >= 18
• An Azure AD app registration with application permissions (not delegated)
• A specific tenant ID (not "common")

Installation

npm (recommended)

npm install -g @okapi-ca/ms-365-admin-mcp-server

ms-365-admin-mcp-server --help

Docker

docker pull ghcr.io/okapi-ca/ms-365-admin-mcp-server:latest

docker run --rm -i \ -e MS365_ADMIN_MCP_CLIENT_ID=... \ -e MS365_ADMIN_MCP_CLIENT_SECRET=... \ -e MS365_ADMIN_MCP_TENANT_ID=... \ ghcr.io/okapi-ca/ms-365-admin-mcp-server:latest

From source

git clone https://github.com/okapi-ca/ms-365-admin-mcp-server.git

cd ms-365-admin-mcp-server npm install npm run generate npm run build

Configuration

Environment variables

MCP client configuration (Claude Desktop, etc.)

{

"mcpServers": { "ms365-admin": { "command": "node", "args": ["/path/to/ms-365-admin-mcp-server/dist/index.js"], "env": { "MS365_ADMIN_MCP_CLIENT_ID": "your-client-id", "MS365_ADMIN_MCP_CLIENT_SECRET": "your-client-secret", "MS365_ADMIN_MCP_TENANT_ID": "your-tenant-id" } } } }

VS Code (1.102+)

VS Code consumes the same MCP protocol but uses a different config layout — servers instead of mcpServers, an explicit type field, and inputs for secret prompts. A ready-to-copy sample lives at .vscode/mcp.json.example; copy it to .vscode/mcp.json and VS Code will prompt for the tenant / client / secret on first start, then store them in its secret store (the real mcp.json is gitignored so resolved secrets never reach the repo).

Minimal stdio setup:

{

"inputs": [ { "type": "promptString", "id": "ms365-tenant-id", "description": "Tenant ID" }, { "type": "promptString", "id": "ms365-client-id", "description": "Client ID" }, { "type": "promptString", "id": "ms365-client-secret", "description": "Client secret", "password": true, }, ], "servers": { "ms365-admin": { "type": "stdio", "command": "ms-365-admin-mcp-server", "args": ["--preset", "security,audit,identity,health"], "env": { "MS365_ADMIN_MCP_TENANT_ID": "${input:ms365-tenant-id}", "MS365_ADMIN_MCP_CLIENT_ID": "${input:ms365-client-id}", "MS365_ADMIN_MCP_CLIENT_SECRET": "${input:ms365-client-secret}", }, }, }, }

For remote HTTP deployments, use "type": "http" with a url field (VS Code 1.103+ handles OAuth 2.0 Dynamic Client Registration natively) or fall back to the mcp-remote bridge when the native browser flow is unavailable — see the example file for both shapes.

Tools surface in Agent mode (GitHub Copilot Chat). VS Code asks for per-tool approval; the --preset flag above keeps the catalog manageable. Use Cmd/Ctrl+Shift+P → MCP: List Servers → Show Output to see logs.

Catalog size limit on Claude.ai surfaces

Claude.ai, Claude Desktop and cloud sessions serve each connector's tools under a fixed budget of 1 MiB of input schema (compact JSON; tool names and descriptions are not counted). Tools are admitted in alphabetical order and a tool that no longer fits is dropped silently. Claude Code CLI is not affected. The limit is not documented publicly.

In a cloud session linked to a Mac, the claude-device bridge relays every local MCP server under one connector (lcl- - ), so all local servers share that budget, and this server's tools sort after most of them.

Request bodies are therefore advertised shallow (src/advertised-schema.ts): the entity's own fields with their descriptions, nested objects as free-form objects, and no Read-only. fields. Graph validates the body; the server forwards it as sent. With writes enabled the catalog advertises about 595 KB of input schema, and test/advertised-schema.test.ts fails the build above 768 KiB.

If tools still go missing in a cloud session, shrink what this server advertises (read-only mode, --preset, --enabled-tools), disable unused local servers, or register this server as its own connector so it gets a budget of its own.

Remote HTTP server: device_code authentication (RFC 8628)

If the server runs in HTTP / OAuth mode on a remote host (e.g. Azure Container Apps) and the client connects via mcp-remote, the standard flow requires a browser to reach localhost:14543/oauth/callback. When that isn't possible — macOS Platform SSO hijacks the WebKit flow, Claude Code runs in a headless Docker container, the user is on a remote SSH dev env — use the ms-365-admin-mcp-auth bootstrap to pre-seed mcp-remote's token cache instead.

npx -p @okapi-ca/ms-365-admin-mcp-server@latest ms-365-admin-mcp-auth \

--server https://your-mcp-host.azurecontainerapps.io/mcp

The helper prints a URL and a user code; you sign in on any device you trust (phone, another laptop) and the tokens are written to ~/.mcp-auth/mcp-remote- /. Claude Desktop / Claude Code then launches mcp-remote normally and finds the cached tokens without ever opening a browser.

See docs/TROUBLESHOOTING.md for Docker / remote-dev patterns and exit code reference.

Usage

CLI options

--read-only              Read-only mode (default)

--allow-writes Enable write operations --enabled-tools Filter tools by regex pattern --preset Use preset categories (comma-separated) --partition Tool partition served on /mcp/ in HTTP mode (default: domains) --list-presets List available presets and exit --list-tools List available tools and exit --list-permissions List required Graph API permissions and exit --verify-login