msgraphmcp

MCP Server for the Microsoft Graph API — runs as a container, exposes 140+ tools across all major Microsoft 365 workloads to any MCP-compatible client such as Claude Code, and supports four authentication modes: authorization code + PKCE (delegated, recommended for Kubernetes/HTTP), client secret (app-only), client certificate (app-only, recommended for production), and device code (interactive, local use).

---

Table of Contents

• Architecture
• Prerequisites
• Azure App Registration
• Quick Start
  • Local (Node.js)
  • Docker
  • Claude Code Integration
• Environment Variables
• Authentication Flow
• Tool Reference
  • Users
  • Mail
  • Calendar
  • OneDrive / Files
  • Groups
  • Microsoft Teams
  • Contacts
  • To Do / Tasks
  • SharePoint Sites
  • Intune — Apps
  • Intune — Device Configurations
  • Intune — Settings Catalog
  • Intune — Compliance Policies
  • Intune — Managed Devices
  • Intune — Device Diagnostics
  • Intune — Notification Templates
  • Auth
• Development
• CI/CD and Docker Registry
• Security Notes

---

Architecture

Claude Code (MCP client)           Another MCP client
       │  stdio / HTTP                    │  HTTP
       ▼                                  ▼
  msgraphmcp (Node.js / TypeScript)
  ├── auth/TokenManager     Four auth modes — per-session in HTTP mode
  ├── graph/GraphClient     Axios wrapper, user-stamped logs, auto-pagination, single retry on 401
  └── tools/
      ├── users · mail · calendar · files · groups
      ├── teams · contacts · tasks · sites
      └── intune (apps · device config · settings catalog · compliance · managed devices)
       │  HTTPS
       ▼
  Microsoft Graph API  (https://graph.microsoft.com/v1.0)

HTTP mode auth flow (authorization code, per-session):
  1. MCP client connects  →  server creates session with isolated TokenManager
  2. Tool call without auth  →  401 + loginUrl (/auth/login?token=<one-time-token>)
  3. Browser → GET /auth/login?token=<one-time-token>  →  Microsoft Login
  4. GET /auth/callback  →  token stored in that session's TokenManager only
  5. Subsequent tool calls from the same session use that user's token exclusively

The login token is a single-use 256-bit value that maps server-side to the
MCP session — the session ID itself never appears in any URL, browser
history, or proxy log.

In HTTP mode every MCP session gets its own isolated TokenManager — tokens are never shared between sessions. Each user authenticates independently. Tokens are kept in-memory for the lifetime of the session; no cross-user token bleed is possible even when multiple clients are connected simultaneously.

---

Prerequisites

Requirement	Version
Node.js	≥ 20
npm	≥ 10
Docker	≥ 24 (optional)
Microsoft 365 / Azure AD tenant	—
Azure App Registration	see below

---

Azure App Registration

1. Open Azure Portal → App registrations → New registration.

2. Name: msgraphmcp (or any name)

3. Supported account types: Accounts in this organizational directory only (or multitenant if needed)

4. Redirect URI — depends on the auth mode you plan to use:

   • Authorization code flow (Mode A): select Web → enter your callback URL, e.g. https://msgraph.example.com/auth/callback
   • Device code flow (Mode D): select Mobile and desktop applications → https://login.microsoftonline.com/common/oauth2/nativeclient
   • App-only (Modes B/C): no redirect URI needed

5. Click Register — copy the Application (client) ID and Directory (tenant) ID.

6. Add permissions — the type depends on the auth mode you choose:

   • Authorization code / device code (delegated): Go to API Permissions → Add a permission → Microsoft Graph → Delegated permissions and add the permissions below, then click Grant admin consent.
   • Client secret / certificate (app-only): Go to API Permissions → Add a permission → Microsoft Graph → Application permissions and add the same permissions, then click Grant admin consent.

     With app-only auth, userId: "me" does not resolve — use explicit UPNs or object IDs in all tool calls.

Permission	Purpose
User.ReadWrite.All	Manage users
Group.ReadWrite.All	Manage groups
GroupMember.ReadWrite.All	Manage group membership
Mail.ReadWrite	Read/write mailboxes
Mail.Send	Send email
Calendars.ReadWrite	Manage calendars & events
Files.ReadWrite.All	OneDrive CRUD
Sites.ReadWrite.All	SharePoint CRUD
Tasks.ReadWrite	Microsoft To Do
Contacts.ReadWrite	Contacts
Team.ReadWrite.All	Teams management
Channel.ReadWrite.All	Teams channels
ChannelMessage.Send	Send Teams messages
Directory.ReadWrite.All	Directory objects
DeviceManagementApps.ReadWrite.All	Intune apps
DeviceManagementConfiguration.ReadWrite.All	Intune device configs
DeviceManagementManagedDevices.ReadWrite.All	Managed devices
DeviceManagementServiceConfig.ReadWrite.All	Intune service config

7. Click Grant admin consent for <your tenant>.

Tip: You can restrict the scope by setting the GRAPH_SCOPES environment variable to only the permissions you actually need.

---

Quick Start

Local (Node.js)

git clone https://github.com/DustHoff/msgraphmcp.git
cd msgraphmcp
npm install
npm run build

export AZURE_CLIENT_ID="your-client-id"
export AZURE_TENANT_ID="your-tenant-id"
export TOKEN_CACHE_PATH="$HOME/.msgraphmcp/tokens.json"

node dist/index.js

On first run the device code authentication prompt appears on stderr:

============================================================
To sign in, use a web browser to open the page
https://microsoft.com/devicelogin and enter the code XXXXXXXX to authenticate.
============================================================

After successful authentication the MCP server is ready and keeps the token refreshed automatically.

Docker

cp .env.example .env
# Edit .env with your AZURE_CLIENT_ID and AZURE_TENANT_ID

docker-compose up

docker-compose.yml mounts a named volume (token-cache) at /data so the token cache survives container restarts.

First run (one-time device code):

docker-compose run --rm msgraphmcp
# Follow the authentication prompt on screen, then Ctrl+C
# Subsequent runs will use the cached refresh token
docker-compose up -d

Claude Code Integration

Add the server to your project's .claude/settings.json:

{
  "mcpServers": {
    "msgraphmcp": {
      "command": "node",
      "args": ["dist/index.js"],
      "env": {
        "AZURE_CLIENT_ID": "your-client-id",
        "AZURE_TENANT_ID": "your-tenant-id",
        "TOKEN_CACHE_PATH": "/path/to/tokens.json"
      }
    }
  }
}

Or using the pre-built Docker image from GHCR:

{
  "mcpServers": {
    "msgraphmcp": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-v", "msgraphmcp-tokens:/data",
        "-e", "AZURE_CLIENT_ID=your-client-id",
        "-e", "AZURE_TENANT_ID=your-tenant-id",
        "ghcr.io/DustHoff/msgraphmcp:latest"
      ]
    }
  }
}

Restart Claude Code after editing the config — the server appears in the MCP tools panel.

---

Environment Variables

Variable	Required	Default	Description
AZURE_CLIENT_ID	Yes	—	App Registration client ID
AZURE_TENANT_ID	No	common	Tenant ID or common for multi-tenant
AZURE_CLIENT_SECRET	No	—	Client secret — required for Mode A (auth code) and Mode B (app-only)
AZURE_REDIRECT_URI	No	—	Auth mode A — full callback URL, e.g. https://msgraph.example.com/auth/callback. When set together with AZURE_CLIENT_SECRET, activates the authorization code flow
AZURE_CLIENT_CERTIFICATE_PATH	No	—	Auth mode C — path to a PEM private key file; must be set together with THUMBPRINT
AZURE_CLIENT_CERTIFICATE_THUMBPRINT	No	—	Auth mode C — SHA-256 certificate thumbprint (64 hex chars) from the App Registration
GRAPH_SCOPES	No	all scopes	Space-separated delegated scopes (used in auth code and device code modes)
TOKEN_CACHE_PATH	No	/data/tokens.json	Path to the MSAL token cache file
PORT	No	—	When set, the server listens on HTTP (Kubernetes mode); otherwise uses stdio
LOG_LEVEL	No	info	Log verbosity: debug, info, warn, error
MAX_SESSIONS	No	50	Maximum concurrent MCP sessions in HTTP mode. New connections beyond this limit receive HTTP 503.
SESSION_IDLE_TIMEOUT_MINUTES	No	60	Minutes of inactivity before an MCP session is automatically closed and removed.

---

Authentication Flow

Four modes are selected automatically by which environment variables are set:

Mode	Env vars set	Type	User context
A — Authorization Code	AZURE_CLIENT_SECRET + AZURE_REDIRECT_URI	Delegated	Yes — full userId: "me" support
B — Client Secret	AZURE_CLIENT_SECRET (no redirect URI)	App-only	No — use explicit UPNs/object IDs
C — Client Certificate	AZURE_CLIENT_CERTIFICATE_PATH + THUMBPRINT	App-only	No
D — Device Code	none of the above	Delegated	Yes

Mode A — Authorization Code + PKCE (delegated, recommended for HTTP/Kubernetes)

Set AZURE_CLIENT_SECRET and AZURE_REDIRECT_URI. Each MCP session authenticates independently — tokens are isolated per session and kept in-memory. Suitable for containers with Entra ID Conditional Access — CA compliance is evaluated against the user's browser device, not the container. Multiple users can be authenticated simultaneously.

Prerequisites:

• Register AZURE_REDIRECT_URI (e.g. https://msgraph.example.com/auth/callback) as a Web redirect URI in the Entra ID app registration.
• Grant Delegated permissions (not Application) + admin consent.

Authentication flow

Step 1 — Connect your MCP client (Claude Code or other)
  POST /mcp  (initialize)
  └─► Server creates a session with its own isolated TokenManager
  └─► Session ID returned in response header:  mcp-session-id: <uuid>

Step 2 — Authenticate (once per session / pod restart)
  Tool call without auth → 401 response containing a fresh loginUrl
  GET /auth/login?token=<one-time-token>
  └─► Server resolves the one-time token → session, invalidates the token
  └─► Server generates PKCE code_verifier + S256 challenge
  └─► Redirects browser → Microsoft login page
  └─► User authenticates + consents
  └─► Microsoft redirects → GET /auth/callback?code=...&state=...
  └─► Server exchanges code and stores tokens in that session's TokenManager only
  └─► Green success page shown — browser can be closed

Step 3 — Tool calls
  └─► acquireTokenSilent() — uses the session's in-memory refresh token
  └─► On 401 from Graph: single retry with fresh token, then error

How to start the login flow

A fresh loginUrl is embedded in every 401 response to an unauthenticated tool call, and is also returned by the get_login_url MCP tool:

{
  "error": "Unauthorized",
  "loginUrl": "https://msgraph.example.com/auth/login?token=3941f8c7dca11388..."
}

Open the loginUrl directly in a browser to authenticate. The token is single-use and expires after 15 minutes — each tool call / get_login_url invocation mints a new one bound to the current MCP session. The session ID itself is never exposed in URLs.

Health check

The /health endpoint intentionally returns only aggregate counts — no session IDs, no UPNs — so that anyone who can reach the endpoint cannot enumerate active users or sessions.

GET /health
→ {
    "status": "ok",
    "service": "msgraphmcp",
    "authMode": "authorization-code",
    "sessions": 2,
    "authenticatedSessions": 1
  }

Token lifetime

Tokens are kept in-memory for the session's lifetime. They survive within a running pod but are lost on pod restart — each session must re-authenticate after a restart. The loginUrl in the 401 response makes re-authentication a single browser visit.

Kubernetes deployment

Uncomment the Option A block in k8s/deployment.yaml and set AZURE_REDIRECT_URI to your ingress hostname:

- name: AZURE_CLIENT_SECRET
  valueFrom:
    secretKeyRef:
      name: msgraphmcp-azure
      key: AZURE_CLIENT_SECRET
- name: AZURE_REDIRECT_URI
  value: https://msgraph.example.com/auth/callback

Mode B — Client Secret (app-only)

Set AZURE_CLIENT_SECRET without AZURE_REDIRECT_URI. Requires Application permissions + admin consent. Recommended for fully automated deployments where no user context is needed.

Every request
  └─► acquireTokenByClientCredential({ scopes: ['.default'] })
  └─► Entra ID returns a short-lived access token (no refresh token stored)
  └─► Device compliance CA policies are NOT evaluated — safe in containers

Mode C — Client Certificate (app-only, recommended for production)

Set AZURE_CLIENT_CERTIFICATE_PATH + AZURE_CLIENT_CERTIFICATE_THUMBPRINT. Same flow as Mode B but uses a certificate assertion instead of a shared secret — no secret rotation required.

Every request
  └─► acquireTokenByClientCredential with cert assertion
  └─► Entra ID returns a short-lived access token
  └─► Device compliance CA policies are NOT evaluated

Kubernetes setup:

# 1. Generate a self-signed key + cert (or use your PKI)
openssl req -x509 -newkey rsa:2048 -keyout tls.key -out tls.crt -days 365 -nodes -subj "/CN=msgraphmcp"

# 2. Get the SHA-256 thumbprint (64 hex chars)
openssl x509 -in tls.crt -fingerprint -sha256 -noout | tr -d ':' | sed 's/.*=//'

# 3. Upload tls.crt to the App Registration → Certificates & secrets → Certificates

# 4. Store the private key as a Kubernetes Secret
kubectl create secret generic msgraphmcp-client-cert --from-file=tls.key -n msgraphmcp

Then uncomment the client-cert volume in k8s/deployment.yaml and set AZURE_CLIENT_CERTIFICATE_THUMBPRINT in k8s/secret.yaml.

Mode D — Device Code (delegated, local / interactive)

No secret or certificate configured. Suitable for local use and Claude Code stdio integration. Not recommended in containers under Entra ID Conditional Access device-compliance policies — token refresh from a non-enrolled host will be rejected.

First run
  └─► Device code prompt (stderr) → user visits URL, enters code
  └─► MSAL receives tokens → writes cache to TOKEN_CACHE_PATH

Subsequent runs / token expiry
  └─► acquireTokenSilent() uses cached refresh token
  └─► CA compliance evaluated against the container → may fail

401 from Graph API
  └─► Interceptor retries once with fresh token, then throws error

The refresh token typically lasts 90 days. If it expires, the device code prompt appears again on next start.

---

Tool Reference

All tools accept userId: string parameters that default to "me" (the signed-in user) unless noted.
Parameters marked optional can be omitted.

All ids used in examples below (00000000-0000-0000-0000-..., grp-id, …) are placeholders — never substitute real tenant, user, group or app ids into the documentation.

Users

Tool	Description	Key Parameters
list_users	List directory users	filter, select, top, search
get_user	Get a single user	userId (required), select
create_user	Create a new user	displayName, userPrincipalName, mailNickname, password
update_user	Update user properties	userId, displayName, jobTitle, department, …
delete_user	Delete a user	userId
get_user_member_of	Get groups/roles the user belongs to	userId
reset_user_password	Reset a user's password	userId, newPassword, forceChangePasswordNextSignIn

Example — create user:

create_user displayName="Alice Müller" userPrincipalName="alice@contoso.com" mailNickname="alice" password="P@ssw0rd!"

---

Mail

Tool	Description	Key Parameters
list_messages	List messages in a folder	userId, folderId (default inbox), filter, top, search
get_message	Get a specific message	userId, messageId
send_mail	Send an email	subject, body, toRecipients[], ccRecipients[], bccRecipients[]
reply_to_message	Reply to a message	userId, messageId, comment
forward_message	Forward a message	userId, messageId, toRecipients[], comment
delete_message	Delete a message	userId, messageId
move_message	Move to another folder	userId, messageId, destinationFolderId
list_mail_folders	List mail folders	userId, includeHiddenFolders
create_mail_folder	Create a new folder	userId, displayName, parentFolderId

Well-known folder IDs: inbox, sentitems, drafts, deleteditems, archive, junkemail

Example — send mail:

send_mail subject="Meeting tomorrow" body="Hi Bob,\nSee you at 10:00." toRecipients=[{address:"bob@contoso.com"}]

---

Calendar

Tool	Description	Key Parameters
list_calendars	List all calendars	userId
create_calendar	Create a calendar	userId, name, color
list_events	List events (or calendar view)	userId, calendarId, startDateTime, endDateTime, filter, top
get_event	Get a specific event	userId, eventId
create_event	Create an event	subject, startDateTime, endDateTime, attendees[], location, isOnlineMeeting
update_event	Update an event	userId, eventId, + any field
delete_event	Delete an event	userId, eventId

Example — create Teams meeting:

create_event subject="Sprint Review" startDateTime="2024-06-14T14:00:00" endDateTime="2024-06-14T15:00:00" isOnlineMeeting=true attendees=[{address:"bob@contoso.com",type:"required"}]

---

OneDrive / Files

Tool	Description	Key Parameters
list_drive_items	List items in a folder	userId, itemPath (default /), top
get_drive_item	Get metadata	userId, itemPath or itemId
create_drive_folder	Create a folder	userId, parentPath, folderName, conflictBehavior
upload_drive_file	Upload text file (≤ 4 MB)	userId, filePath, content, conflictBehavior
delete_drive_item	Delete an item	userId, itemPath or itemId
copy_drive_item	Copy an item	userId, itemId, destinationParentId, newName
search_drive	Search OneDrive	userId, query, top
list_shared_with_me	List shared items	userId

---

Groups

Tool	Description	Key Parameters
list_groups	List groups	filter, select, top, search
get_group	Get a group	groupId, select
create_group	Create M365 or Security group	displayName, mailNickname, groupType (Microsoft365|Security)
update_group	Update group	groupId, displayName, description, visibility
delete_group	Delete a group	groupId
list_group_members	List members	groupId, select
add_group_member	Add a member	groupId, memberId
remove_group_member	Remove a member	groupId, memberId
list_group_owners	List owners	groupId
add_group_owner	Add an owner	groupId, ownerId

---

Microsoft Teams

Tool	Description	Key Parameters
list_joined_teams	List teams the user belongs to	userId
get_team	Get team details	teamId
create_team	Create a new team	displayName, description, visibility, template
list_channels	List channels	teamId
get_channel	Get a channel	teamId, channelId
create_channel	Create a channel	teamId, displayName, membershipType
delete_channel	Delete a channel	teamId, channelId
list_channel_messages	List messages	teamId, channelId, top
send_channel_message	Post a message	teamId, channelId, content, contentType
reply_to_channel_message	Reply to a message	teamId, channelId, messageId, content
list_team_members	List members	teamId
add_team_member	Add a member	teamId, userId, roles

---

Contacts

Tool	Description	Key Parameters
list_contacts	List contacts	userId, filter, select, top
get_contact	Get a contact	userId, contactId
create_contact	Create a contact	userId, givenName, surname, emailAddresses[], businessPhones[]
update_contact	Update a contact	userId, contactId, + any field
delete_contact	Delete a contact	userId, contactId

---

To Do / Tasks

Tool	Description	Key Parameters
list_todo_lists	List task lists	userId
create_todo_list	Create a task list	userId, displayName
delete_todo_list	Delete a task list	userId, listId
list_tasks	List tasks in a list	userId, listId, filter, top
create_task	Create a task	userId, listId, title, dueDateTime, importance, reminderDateTime
update_task	Update a task	userId, listId, taskId, status, title, importance
complete_task	Mark as completed	userId, listId, taskId
delete_task	Delete a task	userId, listId, taskId

Task status values: notStarted, inProgress, completed, waitingOnOthers, deferred

---

SharePoint Sites

Tool	Description	Key Parameters
list_sites	List sites	filter, top
get_site	Get a site	siteId or (hostname + sitePath)
search_sites	Search by keyword	query
list_site_lists	List site lists/libraries	siteId
get_site_list	Get a list	siteId, listId
list_site_list_items	List list items	siteId, listId, filter, top, expand
get_site_list_item	Get an item (with fields)	siteId, listId, itemId
create_site_list_item	Create an item	siteId, listId, fields (key-value object)
update_site_list_item	Update item fields	siteId, listId, itemId, fields
delete_site_list_item	Delete an item	siteId, listId, itemId

---

Intune — Apps

Tool	Description	Key Parameters
list_intune_apps	List managed apps	filter, appType, select, top
get_intune_app	Get app details	appId, select
create_intune_web_app	Add a web shortcut app	displayName, publisher, appUrl
create_intune_store_app	Add a store app	displayName, publisher, storeType (windowsStore|iosStore|androidStore), appStoreUrl
update_intune_app	Update app metadata (common mobileApp fields + Win32-specific). Win32 fields auto-mark the body as #microsoft.graph.win32LobApp.	appId, displayName, publisher, description, isFeatured, privacyInformationUrl, informationUrl, notes, owner, developer, roleScopeTagIds[], largeIcon, installCommandLine, uninstallCommandLine, setupFilePath, applicableArchitectures, allowedArchitectures, minimumSupportedWindowsRelease, minimumFreeDiskSpaceInMB, minimumMemoryInMB, minimumNumberOfProcessors, minimumCpuSpeedInMHz, displayVersion, installExperience, returnCodes[], rules[]
delete_intune_app	Delete an app	appId
upload_win32_lob_app	Upload a .intunewin Win32 LOB package. Source is exactly one of: (a) filePath, (b) fileUrl, or (c) a OneDrive reference (oneDriveItemPath or oneDriveItemId, optionally with oneDriveUserId — defaults to "me").	filePath | fileUrl | (oneDriveItemPath | oneDriveItemId [+ oneDriveUserId]), displayName, publisher, description, installCommandLine, uninstallCommandLine, setupFilePath, applicableArchitectures, minimumSupportedWindowsRelease, runAsAccount, deviceRestartBehavior
update_windows_msix_app_content	Replace the package content of an existing #microsoft.graph.windowsUniversalAppX (MSIX) app. Creates a new contentVersion, uploads the .msix unencrypted to the Intune-allocated Azure Blob, commits it, and patches committedContentVersion. Verifies the target app's @odata.type first — Win32, web and store apps are rejected.	appId, filePath | fileUrl | (oneDriveItemPath | oneDriveItemId [+ oneDriveUserId]), fileName (optional)
list_intune_app_relationships	List supersedence / dependency links	appId
set_intune_app_relationships	Set supersedence / dependency links (replaces all)	appId, relationships[]
list_intune_app_assignments	List assignments	appId
assign_intune_app	Assign to groups (replaces all existing)	appId, assignments[] (groupId, intent, filterMode, filterId)
get_intune_app_install_status	Per-device install status (Reports API)	appId, top, includeUserStatuses
list_intune_app_protection_policies	List MAM / App Protection policies	platform (ios|android|all)

Assignment intent values: available, required, uninstall, availableWithoutEnrollment

Win32 LOB upload notes:

• Provide exactly one source: filePath (server-side absolute path), fileUrl (HTTP/HTTPS URL) or a OneDrive reference (oneDriveItemPath or oneDriveItemId, optionally with oneDriveUserId — defaults to "me").
• fileUrl is fetched server-side with an SSRF guard (http/https only; loopback, link-local, RFC1918 ranges, and cloud-metadata hostnames are blocked).
• OneDrive sources resolve the DriveItem via Graph ($select=id,name,size,file,folder,@microsoft.graph.downloadUrl), reject folders, and stream the short-lived pre-authenticated download URL through the same bounded writer.
• Downloads are capped at 2 GB (applies to fileUrl and OneDrive sources alike).
• After upload, set detection/requirement rules via update_intune_app (field rules[]) and assign to groups via assign_intune_app.

Example — upload from OneDrive:

upload_win32_lob_app
  oneDriveItemPath="/Apps/myapp.intunewin"
  displayName="My App"
  publisher="Contoso"
  installCommandLine="setup.exe /S"
  uninstallCommandLine="setup.exe /U"
  setupFilePath="setup.exe"

MSIX content update notes:

• update_windows_msix_app_content requires the target app to be a #microsoft.graph.windowsUniversalAppX; other types are rejected up-front so you don't create an orphaned contentVersion.
• MSIX payloads are uploaded unencrypted — the commit body is {}, no fileEncryptionInfo is sent.
• OneDrive and fileUrl sources reuse the same 2 GB cap and SSRF guard as the Win32 flow.
• Assignments on the app are preserved — only the committed content version changes.

Example — update an existing MSIX app from OneDrive:

update_windows_msix_app_content
  appId="00000000-0000-0000-0000-000000000000"
  oneDriveItemPath="/intune/myapp/myapp.msix"

Example — assign app as required:

assign_intune_app appId="00000000-0000-0000-0000-000000000001" assignments=[{groupId:"grp-id",intent:"required"}]

---

Intune — Device Configurations

Tool	Description	Key Parameters
list_device_configurations	List config profiles	filter, select, top
get_device_configuration	Get a profile	configId
create_device_configuration	Create a profile	displayName, odataType, settings (key-value)
update_device_configuration	Update a profile	configId, displayName, settings
delete_device_configuration	Delete a profile	configId
assign_device_configuration	Assign to groups	configId, assignments[] (groupId, intent: include|exclude)
get_device_configuration_assignments	List assignments	configId
get_device_configuration_device_status	Per-device status	configId, top

Common odataType values:

Platform	@odata.type
Windows 10 (general)	#microsoft.graph.windows10GeneralConfiguration
Windows 10 (endpoint protection)	#microsoft.graph.windows10EndpointProtectionConfiguration
iOS	#microsoft.graph.iosGeneralDeviceConfiguration
Android	#microsoft.graph.androidGeneralDeviceConfiguration
macOS	#microsoft.graph.macOSGeneralDeviceConfiguration

Example — create Windows 10 password policy:

create_device_configuration
  displayName="Windows Password Policy"
  odataType="#microsoft.graph.windows10GeneralConfiguration"
  settings={"passwordRequired":true,"passwordMinimumLength":12,"passwordRequiredType":"alphanumeric"}

---

Intune — Settings Catalog

The Settings Catalog is the modern replacement for device configuration profiles and supports a broader set of settings.

Tool	Description	Key Parameters
list_configuration_policies	List catalog policies	filter, top
get_configuration_policy	Get policy + settings	policyId
create_configuration_policy	Create a policy	name, platforms, technologies, settings[]
update_configuration_policy	Update name/description	policyId, name, description
delete_configuration_policy	Delete a policy	policyId
assign_configuration_policy	Assign to groups	policyId, assignments[]
search_settings_catalog	Find available settings	keyword, platform, top

Workflow — create a BitLocker policy:

1. search_settings_catalog keyword="BitLocker" platform="windows10"
   → note the settingDefinitionId values

2. create_configuration_policy
     name="BitLocker Encryption"
     platforms="windows10"
     technologies="mdm"
     settings=[{
       id:"0",
       settingInstance:{
         "@odata.type":"#microsoft.graph.deviceManagementConfigurationChoiceSettingInstance",
         "settingDefinitionId":"device_vendor_msft_bitlocker_requiredeviceencryption",
         "choiceSettingValue":{"value":"device_vendor_msft_bitlocker_requiredeviceencryption_1","children":[]}
       }
     }]

---

Intune — Compliance Policies

Tool	Description	Key Parameters
list_compliance_policies	List compliance policies	filter, top
get_compliance_policy	Get a policy	policyId
create_compliance_policy	Create a policy	displayName, odataType, settings, scheduledActionsForRule
update_compliance_policy	Update a policy	policyId, displayName, settings
delete_compliance_policy	Delete a policy	policyId
assign_compliance_policy	Assign to groups	policyId, assignments[]
get_compliance_policy_device_status	Per-device status	policyId, top

Common odataType values:

Platform	@odata.type
Windows 10	#microsoft.graph.windows10CompliancePolicy
iOS	#microsoft.graph.iosCompliancePolicy
Android	#microsoft.graph.androidCompliancePolicy
macOS	#microsoft.graph.macOSCompliancePolicy

Non-compliance actions: block, retire, wipe, notification, pushNotification

---

Intune — Managed Devices

Tool	Description	Key Parameters
list_managed_devices	List enrolled devices	filter, select, top
get_managed_device	Get device details	deviceId
sync_managed_device	Trigger policy sync	deviceId
restart_managed_device	Reboot a device	deviceId
shutdown_managed_device	Shut down a Windows device	deviceId
lock_managed_device	Remote lock (PIN/password required to unlock)	deviceId
set_managed_device_name	Rename a Windows device (≤ 15 chars)	deviceId, deviceName
retire_managed_device	Unenroll (keep personal data)	deviceId
wipe_managed_device	Factory reset (destructive)	deviceId, keepEnrollmentData, keepUserData
disable_managed_device	Disable a device — keeps enrollment, blocks access	deviceId
reenable_managed_device	Re-enable a previously disabled device	deviceId
send_device_notification	Push a Company Portal notification	deviceId, notificationTitle, notificationBody
windows_defender_scan	Trigger a Defender antivirus scan	deviceId, quickScan
windows_defender_update_signatures	Force Defender signature update	deviceId
rotate_bitlocker_keys	Rotate BitLocker recovery key (escrowed to Entra)	deviceId
rotate_local_admin_password	Rotate Windows LAPS local admin password	deviceId
trigger_proactive_remediation	Run a Proactive Remediation script on demand	deviceId, scriptId
get_device_compliance_overview	Tenant-wide compliance stats	—

Warning: wipe_managed_device erases all data on the device. Use with extreme caution.

Useful filter examples for list_managed_devices:

filter="operatingSystem eq 'Windows'"
filter="complianceState eq 'noncompliant'"
filter="contains(deviceName,'DESKTOP')"

---

Intune — Device Diagnostics

Tool	Description	Key Parameters
collect_device_diagnostics	Start a "Collect diagnostics" remote action	deviceId
list_device_diagnostics	List / get status of diagnostic collection requests	deviceId, requestId (optional)
download_device_diagnostics	Generate a SAS download URL for a completed ZIP	deviceId, requestId

Workflow:

1. collect_device_diagnostics → returns a requestId.
2. Poll list_device_diagnostics until status: "completed".
3. download_device_diagnostics → returns { value: "<sas-url>" } pointing at the ZIP archive.

---

Intune — Notification Templates

Tool	Description	Key Parameters
list_notification_templates	List templates	top
get_notification_template	Get a template + localized messages	templateId
create_notification_template	Create a template (via beta)	displayName, defaultLocale, brandingOptions, roleScopeTagIds
update_notification_template	Update template metadata	templateId, displayName, brandingOptions, …
delete_notification_template	Delete a template	templateId
add_notification_template_message	Add / update a localized message	templateId, locale, subject, messageTemplate, isDefault
send_notification_template_test	Send a test email using the default locale	templateId

create_notification_template is routed through the beta API because the v1.0 endpoint returns 400 for all write operations on this resource.

---

Auth

Tool	Description	Key Parameters
get_login_url	Returns authentication status and a fresh one-time login URL when sign-in is required. Call this first whenever another tool returns an auth error.	—

---

Development

# Install dependencies
npm install

# Start in watch mode (ts-node, no compile step)
npm run dev

# Build production bundle (esbuild, ~10 ms)
npm run build

# Run unit tests
npm test

# Run tests with coverage report
npm run test:coverage

# TypeScript type check only (no emit)
npm run typecheck

Project Structure

msgraphmcp/
├── src/
│   ├── index.ts                  Entry point — creates MCP server, connects stdio transport
│   ├── auth/
│   │   └── TokenManager.ts       MSAL public client, device code flow, silent refresh
│   ├── graph/
│   │   └── GraphClient.ts        Axios wrapper: auth header injection, pagination, error handling
│   └── tools/
│       ├── shared.ts             URL / OData / HTML escaping helpers
│       ├── auth.ts               1 tool  (get_login_url)
│       ├── users.ts              7 tools
│       ├── mail.ts               9 tools
│       ├── calendar.ts           7 tools
│       ├── files.ts              8 tools
│       ├── groups.ts             10 tools
│       ├── teams.ts              12 tools
│       ├── contacts.ts           5 tools
│       ├── tasks.ts              8 tools
│       ├── sites.ts              10 tools
│       └── intune.ts             64 tools (apps · win32 LOB · MSIX · config · settings catalog · compliance · devices · diagnostics · notification templates)
├── tests/
│   ├── helpers/
│   │   ├── MockMcpServer.ts      Captures tool registrations for unit testing
│   │   └── mockGraphClient.ts    Jest mock for GraphClient
│   ├── auth/
│   │   └── TokenManager.test.ts
│   ├── graph/
│   │   └── GraphClient.test.ts
│   └── tools/
│       ├── users.test.ts … intune.test.ts
├── .github/
│   └── workflows/
│       ├── ci.yml                Test + build on push/PR
│       └── docker.yml            Build multi-arch image → ghcr.io
├── Dockerfile
├── docker-compose.yml
└── .env.example

---

CI/CD and Docker Registry

GitHub Actions

Workflow	Trigger	Jobs
ci.yml	push to main/develop, PR to main	typecheck → test (with coverage) → build → smoke test
docker.yml	push to main, version tags v*.*.*, manual	test (gate) → build multi-arch image → push to GHCR → provenance attestation

Docker image tags

Tag	When created
latest	Every push to main
1.2.3	When a v1.2.3 tag is pushed
1.2	When a v1.2.x tag is pushed
sha-abc1234	Every push (traceability)

Pulling from GHCR

docker pull ghcr.io/DustHoff/msgraphmcp:latest

Setup in your repository

1. Replace DustHoff in the badge URLs above with your GitHub username or organisation.
2. Push to a GitHub repository — Actions will run automatically.
3. For private packages: go to Settings → Packages → msgraphmcp → Package settings → Change visibility if needed.

---

Security Notes

• Per-session token isolation (Mode A): each MCP session in HTTP mode holds its own in-memory token. Tokens are never shared between sessions — a compromised session cannot access another user's data. Tokens are lost on pod restart; re-authentication is a single browser visit via the loginUrl in the 401 response.
• One-time login tokens: the loginUrl carries a single-use 256-bit token (not the MCP session ID). It is consumed on first visit, expires after 15 minutes, and binds server-side to exactly one session — so even if a login URL is disclosed via browser history, a proxy log, or an email it cannot be replayed.
• HTML escaping on the auth pages: the /auth/callback error page escapes all five HTML-significant characters (&, <, >, ", ') — an attacker cannot reflect a crafted error_description (including HTML numeric entities like <script>) into executable markup.
• URL-path safety: all user-supplied opaque ids passed to Graph (groupId, teamId, channelId, appId, configId, policyId, templateId, deviceId, siteId, listId, itemId, memberId, ownerId, userId) are percent-encoded before they are embedded in Graph URLs, so a tool argument cannot smuggle extra path segments, query strings, or fragments into a request. SharePoint composite site ids (hostname,guid,guid) keep their commas intact.
• SSRF guard on Win32 LOB upload: upload_win32_lob_app with a fileUrl argument rejects non-http(s) schemes and blocks loopback, link-local, RFC1918, cloud-metadata and multicast ranges before issuing the download. OneDrive sources reuse the same guard against the short-lived pre-authenticated download URL returned by Graph. The download size is capped at 2 GB in all cases.
• Request-body limit: the HTTP server caps incoming request bodies at 4 MB and destroys the socket immediately on overflow to prevent OOM via a large JSON-RPC payload.
• Session limits: MAX_SESSIONS (default 50) caps concurrent sessions; SESSION_IDLE_TIMEOUT_MINUTES (default 60) reaps idle sessions so abandoned connections do not leak memory.
• Sensitive-value redaction in debug logs: values of keys matching /password|secret|token|credential|private[-_]?key|apikey/i are replaced with ***REDACTED*** before any request/response body is logged.
• Token cache file (tokens.json) is only written in device-code (Mode D) and stdio mode. Written with mode 0o600; mount as a restricted Docker volume; do not bake into images.
• The image runs as the non-root node user (see Dockerfile).
• Secrets (AZURE_CLIENT_ID, AZURE_CLIENT_SECRET, certificate thumbprint) — use Kubernetes Secrets or an external vault; never commit real values to the repository. secrets.txt, tokens.json, and data/ are excluded via both .gitignore and .dockerignore.
• Prefer authorization code (Mode A) over device code (Mode D) for HTTP deployments — tokens are tied to the user's browser device so CA compliance is evaluated correctly. Device code refresh from a non-enrolled container is rejected by device-compliance CA policies.
• Prefer client certificate (Mode C) over client secret (Mode B) for app-only deployments — certificates are not transmitted over the wire and can be rotated without application downtime.
• Authorization code flow: /auth/login?token=<one-time-token> and /auth/callback must be reachable by the authenticating browser but do not need to be internet-facing — internal DNS is sufficient. Requests without a valid, unexpired token are rejected with 400.
• Graph API logging includes the authenticated user's UPN ("user": "alice@contoso.com") in every log entry, making all MS Graph calls attributable to a specific user identity. Set LOG_LEVEL=warn in production to suppress URL logging if UPNs in logs are a concern.
• Scope down GRAPH_SCOPES (delegated modes) or grant only the required permissions (app-only modes) for your use case.
• wipe_managed_device is irreversible — consider requiring explicit confirmation in your workflows.
• See SECURITY-NOTICE.md for the full security assessment including dependency risk analysis.