# Set up your account
Source: https://documentation.deepmask.io/account-setup
Create your DeepMask account, select a plan, configure your team, and review the security and compliance settings that protect all your data.
DeepMask offers accounts for individual users and enterprise organizations, each with different setup paths and capabilities. This page walks you through creating an account, selecting the right plan, onboarding your team (for enterprise customers), and understanding the security and compliance guarantees that apply to all data you bring into the platform.
## Create your account
Go to [deepmask.io](https://deepmask.io) and click **Get Started** or **Try DeepMask**. Fill in your name, work email address, and a password to create your account.
Use your work email address if you plan to join or create an enterprise organization. This makes it easier for administrators to verify and manage team membership.
Check your inbox for a verification email from DeepMask and click the confirmation link. This activates your account and takes you to the plan selection step.
DeepMask offers plans for individuals and for enterprise teams.
The individual plan is for solo users who want access to the full model library, Projects, web search, and data visualization for personal or freelance work.
Select **Individual** during signup and you will be taken directly to the chat interface at [chat.deepmask.io](https://chat.deepmask.io) to start using the platform.
The enterprise plan adds team management, organization-wide usage analytics, centralized billing, and dedicated support. It is designed for companies that need AI across multiple departments at scale.
Select **Enterprise** to begin the organization setup flow, or contact the sales team to discuss your requirements before committing.
Enterprise pricing and feature scope depend on your team size and usage needs. The sales team can provide a tailored proposal. See the [contact sales](#contact-sales) section below.
## Set up an enterprise organization
If you are on an enterprise plan, complete these additional steps to configure your organization and onboard your team.
During the enterprise setup flow, enter your company name and configure your organization's display settings. This creates the shared workspace that all team members will join.
From the **Team Management** dashboard, enter the email addresses of the colleagues you want to invite. DeepMask sends each person an invitation email with a link to join the organization's workspace.
You can invite members in bulk or one at a time. Each member joins with access to shared Projects and the full model library.
Use the team management dashboard to assign administrator or member roles. Administrators can manage seats, view organization-wide usage analytics, configure MCP connectors, and adjust billing settings. Members have access to chat, Projects, and connectors within the permissions set by the administrator.
Connect DeepMask to the tools your team already uses — Microsoft Enterprise, OneDrive, Outlook, SharePoint, and more — through the MCP connectors panel. Connectors configured at the organization level are available to all team members.
See [MCP connectors](/features/mcp-connectors) for a full setup guide.
Once your team is active, use the **Usage** dashboard to track total token consumption, input and output token breakdown, and the distribution of usage across models. This data helps you understand which models your team relies on and optimize your plan accordingly.
## Contact sales
For enterprise pricing, custom contracts, or a guided demo, reach out to the DeepMask sales team:
Send your requirements to [contact@deepmask.io](mailto:contact@deepmask.io). The team responds within 42 hours.
Schedule a 30-minute call to walk through your use case and get a tailored proposal.
When you reach out, include your company name, team size (options: 0–50, 50–300, 300–1000, 1000+), and a brief description of what you expect from DeepMask. This helps the sales team prepare relevant information before your conversation.
## Security and compliance
All DeepMask accounts — individual and enterprise — are protected by the same security and compliance standards.
DeepMask is fully aligned with GDPR for strong privacy and data protection. Data is processed and stored within EU jurisdiction.
Hosted primarily on StackIT, the German sovereign cloud operated by the Schwarz Group. Select models use additional EU-region infrastructure for redundancy and data residency.
Your conversations, files, and project data are never used to train or fine-tune AI models. Your data is yours, and it stays that way.
DeepMask applies enterprise-grade encryption to all data in motion and at rest. ISO 27001 certification is currently in progress.
Never share your account credentials with other users. If you are an enterprise administrator, use the team invitation flow to grant access rather than sharing a single login.
# Get OneDrive Info
Source: https://documentation.deepmask.io/connectors/OneDrive/get-drive
Returns the user's OneDrive metadata — name, type, owner, and storage quota.
`onedrive_get_drive`
Returns the current user's OneDrive metadata, including the drive name, type, owner, and storage quota (used, remaining, and total).
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A drive object with `id`, `name`, `driveType`, `webUrl`, `owner.user.displayName`, `owner.user.email`, and `quota` (containing `total`, `used`, `remaining`, and `state`).
## Example use
> "How much storage do I have left in OneDrive?"
# Get File Content
Source: https://documentation.deepmask.io/connectors/OneDrive/get-file-content
Downloads and returns the text content of a OneDrive file.
`onedrive_get_file_content`
Downloads and returns the text content of a OneDrive file. Best suited for plain-text formats: `.txt`, `.csv`, `.json`, `.xml`, `.md`. Large files are automatically truncated. Get the `item_id` from `onedrive_list_items`, `onedrive_search_files`, or `onedrive_list_recent`.
## Parameters
| Parameter | Type | Required | Description |
| ----------- | ------ | -------- | ---------------------------------------------------------- |
| `item_id` | string | Yes | Item ID of the file to download. |
| `max_chars` | number | No | Maximum characters to return (100–25,000, default 10,000). |
## Returns
An object with:
* `content` — the extracted text
* `truncated` — `true` if the file was cut off at `max_chars`
* `length` — the number of characters returned
If `truncated` is `true`, only a portion of the file was returned. Increase `max_chars` (up to 25,000) or use `onedrive_search_files` to locate a more targeted excerpt.
## Example use
> "Read the contents of the project plan CSV in my OneDrive."
# Get Item
Source: https://documentation.deepmask.io/connectors/OneDrive/get-item
`onedrive_get_item`
Returns metadata for a specific OneDrive file or folder by its item ID — name, size, type, path, created and modified dates, and sharing info. Use `onedrive_list_items`, `onedrive_search_files`, or `onedrive_list_recent` to find item IDs.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ------------------------------------------------------------------------------------------------------------------------ |
| `item_id` | string | Yes | Item ID from `onedrive_list_items`, `onedrive_search_files`, or `onedrive_list_recent`. Use `"root"` for the drive root. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A `DriveItem` object with `id`, `name`, `webUrl`, `size`, `createdDateTime`, `lastModifiedDateTime`, `createdBy.user.displayName`, `lastModifiedBy.user.displayName`, `file.mimeType` (for files), `folder.childCount` (for folders), `parentReference.path`, and `shared.scope`.
## Example use
> "What is the size and last-modified date of item ID `01ABC123`?"
# List Folder Contents
Source: https://documentation.deepmask.io/connectors/OneDrive/list-items
List files and sub-folders in a OneDrive folder. Use folder_id='root' to start at the drive root.
`onedrive_list_items`
Lists files and sub-folders in a OneDrive folder. Pass `folder_id="root"` to start at the top of the drive, then drill into sub-folders by passing their `id` as `folder_id`. Use `onedrive_get_file_content` to download the content of a listed file.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | --------------------------------------------------------------------- |
| `folder_id` | string | No | Item ID of the folder to list. Defaults to `"root"` (the drive root). |
| `limit` | number | No | Max items to return (1–100, default 20). |
| `offset` | number | No | Items to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An object with an `items` array, `folder_id`, `total_count`, `count`, `offset`, `has_more`, and `next_offset`. Each item includes `id`, `name`, `webUrl`, `size`, `createdDateTime`, `lastModifiedDateTime`, `file.mimeType` (for files), `folder.childCount` (for folders), and `parentReference.path`.
## Example use
> "Show me what's in the root of my OneDrive."
> "List the contents of folder ID `01ABC123`."
# List Recent
Source: https://documentation.deepmask.io/connectors/OneDrive/list-recent
`onedrive_list_recent`
Lists files recently accessed or modified in OneDrive, ordered most-recent first. Useful for quickly finding files the user was working on without needing to know their folder path.
This action does not support offset paging. It returns up to `limit` most-recently accessed files in a single call.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `limit` | number | No | Max files to return (1–100, default 20). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An object with an `items` array and a `count`. Each item includes `itemId`, `name`, `webUrl`, `mimeType`, `size`, `lastModifiedDateTime`, `parentPath`, and `isFolder`.
## Example use
> "What files have I recently worked on in OneDrive?"
# List Files Shared With Me
Source: https://documentation.deepmask.io/connectors/OneDrive/list-shared-with-me
Lists OneDrive files and folders that others have shared with the signed-in user.
`onedrive_list_shared_with_me`
Lists OneDrive files and folders that others have shared with the current user. Use `onedrive_get_file_content` with the returned item IDs to read the content of a shared file.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `limit` | number | No | Max items to return (1–100, default 20). |
| `offset` | number | No | Items to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An object with an `items` array, `total_count`, `count`, `offset`, `has_more`, and `next_offset`. Each item includes `id`, `name`, `webUrl`, `size`, `lastModifiedDateTime`, and sharing info.
## Example use
> "What files have been shared with me in OneDrive?"
# OneDrive
Source: https://documentation.deepmask.io/connectors/OneDrive/overview
A reference for every OneDrive action available in DeepMask — what each one does, when to use it, and what parameters it accepts.
## Overview
DeepMask exposes a set of read-only OneDrive actions that let users browse their drive, retrieve file metadata, view recent and shared files, search for content, and download file text — all scoped to what the signed-in user already has access to.
Actions are grouped into three categories:
| Category | Actions | Purpose |
| ---------- | ------- | ---------------------------------------------------------------- |
| **Drive** | 5 | Browse the drive, folders, items, recent files, and shared files |
| **Search** | 1 | Search files and folders by name or content |
| **Files** | 1 | Download the text content of a file |
All OneDrive actions are read-only. DeepMask cannot create, modify, or delete files or folders. Every action runs on behalf of the signed-in user — they can only access content they already have permission to see.
### Drive
| Action | Description |
| ------------------------------ | --------------------------------------------------------------------------------------------- |
| `onedrive_get_drive` | Returns the user's OneDrive metadata — name, type, owner, and storage quota. |
| `onedrive_list_items` | Lists files and sub-folders in a OneDrive folder. Use `folder_id="root"` to start at the top. |
| `onedrive_get_item` | Returns metadata for a specific file or folder by its item ID. |
| `onedrive_list_recent` | Lists files recently accessed or modified, ordered most-recent first. |
| `onedrive_list_shared_with_me` | Lists files and folders that others have shared with the signed-in user. |
### Search
| Action | Description |
| ----------------------- | -------------------------------------------------------------------------------------------------- |
| `onedrive_search_files` | Searches across files and folders by name or content. Returns item IDs for use with other actions. |
### Files
| Action | Description |
| --------------------------- | -------------------------------------------------------------------------------------------------- |
| `onedrive_get_file_content` | Downloads and returns the text content of a file. Best for `.txt`, `.csv`, `.json`, `.xml`, `.md`. |
***
## Prerequisites
Before using OneDrive actions in DeepMask, confirm the following are in place.
### Required Access
* **Microsoft 365 account (Business Basic or higher)** — Your organization must have an active Microsoft 365 or Office 365 subscription with OneDrive enabled.
* **User signed in via Microsoft** — Each user must sign in with their own Microsoft account through DeepMask's connector. Actions run under that user's identity and are limited to content they already have access to.
### What You Do Not Need
* No developer tools, code, or command-line experience
* No changes to existing OneDrive permissions
* No service account or shared credentials
DeepMask uses OAuth 2.0 delegated authentication. When a user connects their account, they sign in with their own Microsoft credentials. DeepMask never stores passwords or receives broader access than the user already has in your tenant.
***
## Common Parameters
Several parameters appear across multiple actions.
| Parameter | Type | Description |
| ----------------- | -------------------------- | ----------------------------------------------------------------------------------------------------------------- |
| `response_format` | `"markdown"` \| `"json"` | Controls the output format. `"markdown"` (default) returns human-readable text; `"json"` returns structured data. |
| `limit` | number (1–100, default 20) | Maximum number of results to return in a single call. |
| `offset` | number (default 0) | Number of results to skip. Use with `limit` to page through large result sets. |
***
## Typical Workflows
### Browsing your OneDrive
Call `onedrive_get_drive` to confirm the drive is accessible and check storage quota.
Call `onedrive_list_items` with `folder_id="root"` to see top-level files and folders.
Pass a folder's `id` from the listing as `folder_id` to browse its contents.
Call `onedrive_get_file_content` with the file's `id` as `item_id` to download its text content.
### Finding a file by keyword
Call `onedrive_search_files` with a keyword. Results include `itemId`, name, path, and size.
Pass the `itemId` to `onedrive_get_file_content` (for text content) or `onedrive_get_item` (for metadata).
### Checking recent or shared files
Call `onedrive_list_recent` to see files the user has recently accessed or modified.
Call `onedrive_list_shared_with_me` to see files others have shared with the signed-in user.
***
## Pagination
Actions that return lists (`onedrive_list_items`, `onedrive_list_shared_with_me`, `onedrive_search_files`) support pagination via `limit` and `offset`.
| Field in response | Description |
| ----------------- | ----------------------------------------------------------- |
| `count` | Number of items in this page |
| `total_count` | Total matching items (where available) |
| `has_more` | `true` if there are more pages |
| `next_offset` | Pass this as `offset` in the next call to get the next page |
`onedrive_list_recent` does not support offset paging. It returns up to `limit` most-recently accessed files in a single call.
***
## Security & Privacy
DeepMask is designed so that you retain full control of your data and your users' access. This section explains the key security properties of the OneDrive integration.
### Delegated Authentication
DeepMask uses OAuth 2.0 delegated permissions, not application-level (app-only) permissions. This means:
* Every action performed by DeepMask is done on behalf of the signed-in user.
* A user can only access OneDrive content they already have permission to see.
* DeepMask cannot bypass OneDrive's existing access controls.
* Removing a user's OneDrive access in Microsoft 365 immediately removes their access in DeepMask.
### No Stored Credentials
DeepMask does not store your Microsoft password or raw file content. Authentication is handled entirely through short-lived OAuth access tokens and refresh tokens, which are encrypted at rest.
### No Service Account
DeepMask authenticates each user individually. This ensures audit logs in your Microsoft 365 tenant accurately reflect which user accessed which content.
### Read-Only Permissions
All OneDrive actions are strictly read-only. DeepMask requests no write or delete permissions — it cannot create, modify, or delete files or folders.
The delegated Microsoft Graph permissions used by these actions are:
| Permission | Actions that use it |
| ---------------- | --------------------------------------------- |
| `User.Read` | All actions (identity context) |
| `Files.Read.All` | All drive, search, and file actions |
| `offline_access` | Keeps the user's session active between calls |
### Revoking Access
To disconnect DeepMask from OneDrive at any time, go to DeepMask → **Connectors** → **OneDrive** and click **Disconnect**.
This immediately revokes all access tokens. No data is retained after disconnection.
Questions about data residency, compliance, or security? Contact DeepMask support at [support@deepmask.io](mailto:support@deepmask.io).
# Search Files
Source: https://documentation.deepmask.io/connectors/OneDrive/search-files-1
`onedrive_search_files`
Searches across the user's OneDrive files and folders by name or content. Returns item IDs, names, paths, sizes, and modified dates. Pass the returned `itemId` to `onedrive_get_file_content` to download a file or to `onedrive_get_item` for full metadata.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------------------------------------------- |
| `query` | string | Yes | Search query matching file names and content — e.g. `"budget 2024"`, `"project plan"`. |
| `limit` | number | No | Max results to return (1–100, default 20). |
| `offset` | number | No | Results to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An object with a `results` array, `count`, `offset`, `has_more`, and `next_offset`. Each result includes `itemId`, `name`, `webUrl`, `mimeType`, `size`, `lastModifiedDateTime`, `parentPath`, and `isFolder`.
## Example use
> "Search my OneDrive for any files related to the Q3 report."
# Catch Me Up
Source: https://documentation.deepmask.io/connectors/Outlook/catch-me-up
Returns a prioritized summary of recent inbox activity and upcoming calendar events.
# Get Attachment
Source: https://documentation.deepmask.io/connectors/Outlook/get-attachment
`outlook_get_attachment`
Downloads a specific attachment from an Outlook message. File attachments return their content as base64 in the structured output. Use `outlook_list_attachments` first to get the attachment ID.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `message_id` | string | Yes | Message ID that owns the attachment. |
| `attachment_id` | string | Yes | Attachment ID from `outlook_list_attachments`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
The attachment object with `id`, `name`, `contentType`, `size`, `contentBytes` (base64-encoded content for file attachments), and `sourceUrl` (for reference/linked attachments).
## Example use
> "Download the PDF attached to message ID `AAMkAGI2...`."
# Get Contact
Source: https://documentation.deepmask.io/connectors/Outlook/get-contact
Returns full details of an Outlook contact by ID — name, email addresses, phone numbers, company, and department.
`outlook_get_contact`
Returns full details of a contact by ID — name, email addresses, phone numbers, company, and department. Use `outlook_list_contacts` or `outlook_search_contacts` to find contact IDs.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | --------------------------------------------------------------------- |
| `contact_id` | string | Yes | Contact ID from `outlook_list_contacts` or `outlook_search_contacts`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
The complete contact object with all available fields including `displayName`, `emailAddresses`, `mobilePhone`, `businessPhones`, `jobTitle`, `companyName`, `department`, and `officeLocation`.
## Example use
> "Get the full contact details for contact ID `AAMkAGI2...`."
# Get Event
Source: https://documentation.deepmask.io/connectors/Outlook/get-event
`outlook_get_event`
Returns the full details of a calendar event by its ID, including body, all attendees, recurrence patterns, and response status. Use `outlook_list_events` or `outlook_search_events` to find event IDs.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | --------------------------------------------------------------- |
| `event_id` | string | Yes | Event ID from `outlook_list_events` or `outlook_search_events`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
The complete event object including body, full attendee list with response status, recurrence pattern, online meeting details, and all metadata.
## Example use
> "Get the full details of the board meeting event ID `AAMkAGI2...`."
# Get Message
Source: https://documentation.deepmask.io/connectors/Outlook/get-message
Returns the full content of an Outlook message by ID, including the complete body.
`outlook_get_message`
Returns the full content of an Outlook message by its ID, including the complete body. HTML bodies are returned as plain text. Use `outlook_list_messages` or `outlook_search_messages` to find message IDs.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | --------------------------------------------------------------------- |
| `message_id` | string | Yes | Message ID from `outlook_list_messages` or `outlook_search_messages`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
The complete message object with full body, all recipients, headers, and metadata.
## Example use
> "Read the full content of message ID `AAMkAGI2...`."
# List Attachments
Source: https://documentation.deepmask.io/connectors/Outlook/list-attachments
`outlook_list_attachments`
Lists all attachments on an Outlook message — file attachments, linked files, and attached items. Returns the name, type, size, and ID for each attachment. Use `outlook_get_attachment` with an attachment ID to retrieve the actual file content.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ----------------------------------------------------------------- |
| `message_id` | string | Yes | Message ID from `outlook_list_messages` or `outlook_get_message`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of attachment objects, each with `id`, `name`, `contentType`, `size`, `isInline`, and `lastModifiedDateTime`.
## Example use
> "What files are attached to message ID `AAMkAGI2...`?"
# List Calendars
Source: https://documentation.deepmask.io/connectors/Outlook/list-calendars
`outlook_list_calendars`
Lists all calendars in the user's Outlook account. Returns calendar IDs, names, colours, and whether the signed-in user can edit each one. Use the returned `id` as `calendar_id` in event actions to target a specific calendar.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of calendar objects, each with `id`, `name`, `color`, `isDefaultCalendar`, and `canEdit`.
## Example use
> "What calendars do I have in Outlook?"
# List Contacts
Source: https://documentation.deepmask.io/connectors/Outlook/list-contacts
`outlook_list_contacts`
Lists contacts from the user's Outlook address book, ordered alphabetically by display name. Use `outlook_get_contact` to retrieve full details for a specific contact, or `outlook_search_contacts` to find a contact by name or email.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `limit` | number | No | Max contacts to return (1–100, default 20). |
| `offset` | number | No | Contacts to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of contact objects, each with `id`, `displayName`, `givenName`, `surname`, `emailAddresses`, `mobilePhone`, `businessPhones`, `jobTitle`, `companyName`, `department`, and `officeLocation`.
## Example use
> "List the first 20 contacts in my address book."
# List Events
Source: https://documentation.deepmask.io/connectors/Outlook/list-events
`outlook_list_events`
Lists calendar events, optionally filtered by date range. When both `start_datetime` and `end_datetime` are provided, returns recurring event instances within the range. Without a date range, returns upcoming events in chronological order with offset paging.
Dates must be ISO 8601 UTC strings — e.g. `"2024-06-01T00:00:00Z"`.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ------------------------------------------------------------------------------------------ |
| `calendar_id` | string | No | Calendar ID from `outlook_list_calendars`. Omit to use the default calendar. |
| `start_datetime` | string | No | Return events at or after this time (ISO 8601 UTC). Must be paired with `end_datetime`. |
| `end_datetime` | string | No | Return events at or before this time (ISO 8601 UTC). Must be paired with `start_datetime`. |
| `limit` | number | No | Max events to return (1–100, default 20). |
| `offset` | number | No | Events to skip for pagination (default 0). Only applies when no date range is set. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of event objects, each with `id`, `subject`, `start`, `end`, `isAllDay`, `attendees`, `organizer`, `location`, `isOnlineMeeting`, `onlineMeetingUrl`, and `webLink`.
## Example use
> "What meetings do I have this week?"
> "Show me all events on Monday 9 June."
# List Folders
Source: https://documentation.deepmask.io/connectors/Outlook/list-folders
`outlook_list_folders`
Lists all mail folders in the user's mailbox. Returns folder IDs, display names, and item counts. Use the returned `id` as the `folder_id` in `outlook_list_messages`.
You can also skip this action and pass well-known folder names directly to mail actions: `inbox`, `drafts`, `sentitems`, `deleteditems`, `junkemail`.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of folder objects, each with `id`, `displayName`, `totalItemCount`, `unreadItemCount`, `isHidden`, and `parentFolderId`.
## Example use
> "What mail folders do I have in my mailbox?"
# List Messages
Source: https://documentation.deepmask.io/connectors/Outlook/list-messages
`outlook_list_messages`
Lists messages in an Outlook mail folder, ordered newest-first. Returns subject, sender, received date, read status, and a short body preview. Use `outlook_get_message` to retrieve the full body of a specific message.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
| `folder_id` | string | No | Folder to list. Supports well-known names: `"inbox"`, `"drafts"`, `"sentitems"`, `"deleteditems"`, `"junkemail"`. Defaults to `"inbox"`. |
| `limit` | number | No | Max messages to return (1–100, default 20). |
| `offset` | number | No | Messages to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of message objects, each with `id`, `subject`, `from`, `toRecipients`, `receivedDateTime`, `isRead`, `bodyPreview`, `importance`, `hasAttachments`, `webLink`, and `conversationId`.
## Example use
> "Show me my last 10 unread emails in the inbox."
# Overview
Source: https://documentation.deepmask.io/connectors/Outlook/msft-outlook-overview
A reference for every Outlook action available in DeepMask — what each one does, when to use it, and what parameters it accepts.
## Overview
DeepMask exposes a set of read-only Outlook actions that let users read mail, browse calendars, look up contacts, and get intelligent summaries of their inbox — all scoped to what the signed-in user already has access to in your Microsoft 365 tenant.
Actions are grouped into four categories:
| Category | Actions | Purpose |
| ------------ | ------- | -------------------------------------------------- |
| **Mail** | 6 | Read folders, messages, and attachments |
| **Calendar** | 4 | Browse calendars and events |
| **Contacts** | 3 | Look up people in the address book |
| **Compound** | 2 | Intelligent inbox summaries and priority detection |
All Outlook actions are read-only. DeepMask cannot send, delete, or modify mail, events, or contacts. Every action runs on behalf of the signed-in user — they can only access content they already have permission to see.
### Mail
| Action | Description |
| -------------------------- | ------------------------------------------------------------------------------------ |
| `outlook_list_folders` | Lists all mail folders in the mailbox (Inbox, Drafts, Sent Items, etc.). |
| `outlook_list_messages` | Lists messages in a folder, newest-first, with subject, sender, and a short preview. |
| `outlook_get_message` | Returns the full content of a message by ID, including the complete body. |
| `outlook_search_messages` | Full-text KQL search across all messages — subject, body, sender, and recipients. |
| `outlook_list_attachments` | Lists all attachments on a message — file name, type, and size. |
| `outlook_get_attachment` | Downloads a specific attachment and returns its content. |
### Calendar
| Action | Description |
| ------------------------ | ----------------------------------------------------------------------------------- |
| `outlook_list_calendars` | Lists all calendars in the user's account with edit permissions. |
| `outlook_list_events` | Lists calendar events, optionally filtered by date range. |
| `outlook_get_event` | Returns full details of a calendar event by ID, including attendees and recurrence. |
| `outlook_search_events` | Full-text KQL search across calendar events. |
### Contacts
| Action | Description |
| ------------------------- | --------------------------------------------------------------------- |
| `outlook_list_contacts` | Lists contacts from the address book, alphabetically by display name. |
| `outlook_get_contact` | Returns full details of a contact by ID. |
| `outlook_search_contacts` | Search contacts by name, email, company, or other fields. |
### Compound
| Action | Description |
| -------------------------- | ------------------------------------------------------------------------------------ |
| `outlook_catch_me_up` | Returns a prioritized summary of recent inbox activity and upcoming calendar events. |
| `outlook_priority_replies` | Scans the inbox for messages that need a reply and threads waiting for a response. |
***
## Prerequisites
Before using Outlook actions in DeepMask, confirm the following are in place.
### Required Access
* **Microsoft 365 account (Business Basic or higher)** — Your organization must have an active Microsoft 365 or Office 365 subscription with Exchange Online and Outlook enabled.
* **User signed in via Microsoft** — Each user must sign in with their own Microsoft account through DeepMask's connector. Actions run under that user's identity and are limited to content they already have access to.
### What You Do Not Need
* No developer tools, code, or command-line experience
* No changes to existing mailbox permissions
* No service account or shared credentials
DeepMask uses OAuth 2.0 delegated authentication. When a user connects their account, they sign in with their own Microsoft credentials. DeepMask never stores passwords or receives broader access than the user already has in your tenant.
***
## Common Parameters
Several parameters appear across multiple actions.
| Parameter | Type | Description |
| ----------------- | -------------------------- | ----------------------------------------------------------------------------------------------------------------- |
| `response_format` | `"markdown"` \| `"json"` | Controls the output format. `"markdown"` (default) returns human-readable text; `"json"` returns structured data. |
| `limit` | number (1–100, default 20) | Maximum number of results to return in a single call. |
| `offset` | number (default 0) | Number of results to skip. Use with `limit` to page through large result sets. |
### Well-Known Folder IDs
Mail actions accept either a folder ID from `outlook_list_folders` or these well-known names:
| Name | Folder |
| -------------- | ------------- |
| `inbox` | Inbox |
| `drafts` | Drafts |
| `sentitems` | Sent Items |
| `deleteditems` | Deleted Items |
| `junkemail` | Junk Email |
***
## Typical Workflows
### Reading recent emails
Call `outlook_list_folders` to see all available mail folders and their IDs.
Call `outlook_list_messages` with a `folder_id` (e.g. `"inbox"`) to get messages newest-first.
Call `outlook_get_message` with a `message_id` from the list to get the full body.
Call `outlook_list_attachments` with the `message_id`, then `outlook_get_attachment` with an `attachment_id` to download a file.
### Getting a daily briefing
Call `outlook_catch_me_up` with `hours_back` set to how far back to scan (default 24h) and `include_calendar=true` to also surface upcoming events.
Use the `message_id` values returned in `flagged` to call `outlook_get_message` for full context on the most urgent threads.
### Finding a specific email thread
Call `outlook_search_messages` with a KQL query — e.g. `"from:alice@example.com subject:budget"`.
Pass a `message_id` from the results to `outlook_get_message`.
### Checking the calendar
Call `outlook_list_events` with `start_datetime` and `end_datetime` in ISO 8601 UTC format — e.g. `"2024-06-01T00:00:00Z"`.
Call `outlook_get_event` with an `event_id` to see the full attendee list, body, and recurrence info.
***
## Pagination
Actions that return lists support pagination via `limit` and `offset`.
| Field in response | Description |
| ----------------- | ----------------------------------------------------------- |
| `count` | Number of items in this page |
| `has_more` | `true` if there are more pages |
| `next_offset` | Pass this as `offset` in the next call to get the next page |
`outlook_search_contacts` does not support offset paging. Results are unordered and a single page is returned up to the `limit`.
***
## Security & Privacy
DeepMask is designed so that you retain full control of your data and your users' access. This section explains the key security properties of the Outlook integration.
### Delegated Authentication
DeepMask uses OAuth 2.0 delegated permissions, not application-level (app-only) permissions. This means:
* Every action performed by DeepMask is done on behalf of the signed-in user.
* A user can only read mail, calendar events, and contacts they already have access to.
* DeepMask cannot bypass Exchange Online's existing access controls.
* Removing a user's mailbox access in Microsoft 365 immediately removes their access in DeepMask.
### No Stored Credentials
DeepMask does not store your Microsoft password, your Client Secret, or raw mailbox content. Authentication is handled entirely through short-lived OAuth access tokens and refresh tokens, which are encrypted at rest.
### No Service Account
Unlike some integrations that use a single shared service account to access all data, DeepMask authenticates each user individually. This ensures audit logs in your Microsoft 365 tenant accurately reflect which user accessed which content.
### Read-Only Permissions
All active Outlook actions are strictly read-only. DeepMask requests no write, send, or delete permissions — it cannot send mail, create events, or modify contacts.
The delegated Microsoft Graph permissions used by these actions are:
| Permission | Actions that use it |
| ---------------- | --------------------------------------------- |
| `User.Read` | All actions (identity context) |
| `Mail.Read` | All mail actions |
| `Mail.ReadBasic` | Folder listing and message previews |
| `Calendars.Read` | All calendar actions |
| `Contacts.Read` | All contact actions |
| `offline_access` | Keeps the user's session active between calls |
If a user reports missing mail or calendar events, confirm that admin consent has been granted for all required permissions in your Azure AD app registration. See the [Microsoft Enterprise](/msft-enterprise) guide for details.
### Revoking Access
To disconnect DeepMask from Outlook at any time:
* In DeepMask → **Connectors** → **Microsoft (Enterprise)**, click **Disconnect**.
Either action immediately revokes all access tokens. No data is retained after disconnection.
Questions about data residency, compliance, or security? Contact DeepMask support at [support@deepmask.io](mailto:support@deepmask.io).
# Priority Replies Needed
Source: https://documentation.deepmask.io/connectors/Outlook/priority-replies-needed
Scans the inbox for messages that need a reply and threads waiting for a response.
# Search Contacts
Source: https://documentation.deepmask.io/connectors/Outlook/search-contacts
`outlook_search_contacts`
Searches contacts by name, email address, company, or other fields. Returns matching contacts up to the specified limit.
Contact search does not support offset paging — results are unordered and only a single page is returned up to `limit`. Use `outlook_list_contacts` with pagination if you need to browse the full address book.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ----------------------------------------------------------------- |
| `query` | string | Yes | Search term — e.g. `"Alice"`, `"Contoso"`, `"alice@example.com"`. |
| `limit` | number | No | Max contacts to return (1–100, default 20). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of matching contact objects with the same fields as `outlook_list_contacts`.
## Example use
> "Find the contact details for Alice from Contoso."
# Search Events
Source: https://documentation.deepmask.io/connectors/Outlook/search-events
`outlook_search_events`
Full-text search across the user's calendar events using KQL (Keyword Query Language) syntax. Searches subject, body, location, and attendees. Returns event previews with IDs you can pass to `outlook_get_event` for full details.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
| `query` | string | Yes | Search query using KQL syntax — e.g. `"quarterly review"`, `"\"team standup\""`, `"organizer:alice@company.com"`, `"location:\"Conference Room A\""`. |
| `limit` | number | No | Max results to return (1–100, default 20). |
| `offset` | number | No | Results to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of matching event objects with preview data.
KQL supports phrases (`"exact phrase"`), field-scoped queries (`organizer:user@domain.com`, `location:"Room Name"`), and boolean operators. For simple keyword searches, plain text works without any special syntax.
## Example use
> "Find all calendar events related to the Q3 planning review."
# Search Messages
Source: https://documentation.deepmask.io/connectors/Outlook/search-messages
`outlook_search_messages`
Full-text search across the user's Outlook messages using KQL (Keyword Query Language) syntax. Searches subject, body, sender, and recipients. Returns message previews with IDs you can pass to `outlook_get_message` for the full content.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------------------------------------------------------------------------- |
| `query` | string | Yes | Search query using KQL syntax — e.g. `"budget"`, `"\"Q3 report\""`, `"from:alice@example.com"`, `"subject:invoice"`. |
| `limit` | number | No | Max results to return (1–100, default 20). |
| `offset` | number | No | Results to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An array of matching message objects with preview data, including `id`, `subject`, `from`, `receivedDateTime`, `isRead`, `bodyPreview`, and `conversationId`.
KQL supports phrases (`"exact phrase"`), field-scoped queries (`from:user@domain.com`, `subject:keyword`), and boolean operators (`invoice AND 2024`). For simple keyword searches, plain text works without any special syntax.
## Example use
> "Find all emails from [alice@example.com](mailto:alice@example.com) with 'budget' in the subject."
# Get Site Details
Source: https://documentation.deepmask.io/connectors/SharePoint/get-site-details
`sharepoint_get_site`
Returns full metadata for a single SharePoint site by its ID. Use this when you already have a site ID — for example from `sharepoint_search_sites` or `sharepoint_get_root_site`.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | ---------------------------------------------------------------- |
| `site_id` | string | Yes | SharePoint site ID. Format: `"tenant.sharepoint.com,guid,guid"`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
The site object with `id`, `displayName`, `webUrl`, `name`, `description`, `createdDateTime`, and `lastModifiedDateTime`.
## Example use
> "Give me the details for site ID `contoso.sharepoint.com,abc123,def456`."
# List Document Libraries
Source: https://documentation.deepmask.io/connectors/SharePoint/list-drives
`sharepoint_list_drives`
Lists all document libraries (drives) within a given SharePoint site. Returns the drive IDs you need to browse files and folders with `sharepoint_list_folder` or `sharepoint_get_drive`.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | --------------------------------------------------------- |
| `site_id` | string | Yes | ID of the site whose document libraries you want to list. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
An object containing a `drives` array, a `count`, and the `site_id`. Each drive entry includes its `id`, `name`, `driveType`, and `webUrl`.
## Example use
> "List all document libraries in the Finance SharePoint site."
# List Folder Contents
Source: https://documentation.deepmask.io/connectors/SharePoint/list-folder-contents
List files and sub-folders inside a SharePoint drive folder.
`sharepoint_list_folder`
Lists the files and sub-folders inside a folder within a drive. Pass `item_id="root"` to start at the top of a document library, then drill into sub-folders by passing their `item_id`.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | --------------------------------------------------------------------- |
| `drive_id` | string | Yes | ID of the drive containing the folder. |
| `item_id` | string | No | Item ID of the folder to list. Defaults to `"root"` (the drive root). |
| `limit` | number | No | Max items to return (1–100, default 20). |
| `offset` | number | No | Items to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A paginated list of items (files and folders), each with `id`, `name`, `webUrl`, `size`, `lastModifiedDateTime`, and type indicators. Includes `total_count`, `count`, `offset`, `has_more`, and `next_offset`.
Results are ordered alphabetically by name.
## Example use
> "Show me everything in the root of the Marketing document library."
# Get Drive Details
Source: https://documentation.deepmask.io/connectors/SharePoint/msft-sharepoint-enterprise-get-drive-details
Returns metadata for a specific document library by its drive ID, including storage quota.
`sharepoint_get_drive`
Returns metadata for a specific document library by its drive ID, including name, type, web URL, and storage quota.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `drive_id` | string | Yes | The drive (document library) ID. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A drive object with `id`, `name`, `driveType`, `webUrl`, and `quota` (containing `total`, `used`, and `remaining` in bytes).
## Example use
> "How much storage is left in drive ID `b!abc123`?"
# Get File Content
Source: https://documentation.deepmask.io/connectors/SharePoint/msft-sharepoint-enterprise-get-file-content
`sharepoint_get_file_content`
Downloads and returns the text content of a SharePoint file. Best suited for plain-text formats: `.txt`, `.csv`, `.json`, `.xml`, `.md`. Binary formats (images, PDFs, Office files) are not supported for content extraction. Large files are automatically truncated.
## Parameters
| Parameter | Type | Required | Description |
| ----------- | ------ | -------- | ---------------------------------------------------------- |
| `drive_id` | string | Yes | ID of the drive containing the file. |
| `item_id` | string | Yes | Item ID of the file to read. |
| `max_chars` | number | No | Maximum characters to return (100–25,000, default 10,000). |
## Returns
An object with:
* `content` — the extracted text
* `truncated` — `true` if the file was cut off at `max_chars`
* `length` — the number of characters returned
If `truncated` is `true`, only a portion of the file was returned. Increase `max_chars` (up to 25,000) or use `sharepoint_search_files` to locate a more targeted excerpt.
## Example use
> "Read the contents of the Q3 budget CSV in the Finance library."
# Get File Metadata
Source: https://documentation.deepmask.io/connectors/SharePoint/msft-sharepoint-enterprise-get-file-metadata
Returns metadata for a SharePoint file or folder without downloading its content.
`sharepoint_get_file_metadata`
Returns metadata for a specific file or folder — name, size, MIME type, timestamps, and URLs — without downloading the actual content. Use this when you only need to know what a file is, not what it contains.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `drive_id` | string | Yes | ID of the drive containing the file. |
| `item_id` | string | Yes | Item ID of the file or folder. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A `DriveItem` object with `id`, `name`, `webUrl`, `size`, `createdDateTime`, `lastModifiedDateTime`, `file.mimeType` (for files), `folder.childCount` (for folders), and `parentReference`.
## Example use
> "What is the size and last-modified date of item ID `01ABC123` in drive `b!xyz`?"
# Get SharePoint Root Site
Source: https://documentation.deepmask.io/connectors/SharePoint/msft-sharepoint-enterprise-get-root-site
Returns the tenant's root SharePoint site — the recommended starting point for discovering site IDs.
`sharepoint_get_root_site`
Returns the tenant's root SharePoint site. This is the recommended starting point when you do not yet know any site IDs. The root site's `id` can be passed to `sharepoint_list_drives` to begin browsing document libraries.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
The root site object, including `id`, `displayName`, `webUrl`, `name`, `description`, `createdDateTime`, and `lastModifiedDateTime`.
## Example use
> "What is the root SharePoint site for this tenant?"
# Overview - SharePoint (Enterprise)
Source: https://documentation.deepmask.io/connectors/SharePoint/msft-sharepoint-enterprise-overview
A reference for every SharePoint action available in DeepMask — what each one does, when to use it, and what parameters it accepts.
## Overview
DeepMask exposes a set of read-only SharePoint actions that let users explore sites, browse document libraries, and retrieve file content — all scoped to what the signed-in user already has access to in your Microsoft 365 tenant.
Actions are grouped into four categories:
| Category | Actions | Purpose |
| ---------- | ------- | --------------------------------------------------------- |
| **Sites** | 3 | Discover and inspect SharePoint sites |
| **Drives** | 2 | List and inspect document libraries within a site |
| **Files** | 3 | Browse folders, read metadata, and download file content |
| **Search** | 1 | Full-text search across all accessible SharePoint content |
All SharePoint actions are read-only. DeepMask cannot create, modify, or delete files, folders, or sites. Every action runs on behalf of the signed-in user — they can only access content they already have permission to see.
### Sites
| Action | Description |
| -------------------------- | ------------------------------------------------------------------------------------------------- |
| `sharepoint_get_root_site` | Returns the tenant's root SharePoint site. Good starting point when you don't yet have a site ID. |
| `sharepoint_search_sites` | Search for SharePoint sites by keyword. Returns site IDs, display names, and URLs. |
| `sharepoint_get_site` | Returns full metadata for a single site by its ID. |
### Drives
| Action | Description |
| ------------------------ | ----------------------------------------------------------------------------------------------- |
| `sharepoint_list_drives` | Lists all document libraries (drives) within a site. Returns drive IDs for use in file actions. |
| `sharepoint_get_drive` | Returns metadata for a specific drive, including name, type, URL, and storage quota. |
### Files
| Action | Description |
| ------------------------------ | -------------------------------------------------------------------------------------------------- |
| `sharepoint_list_folder` | Lists files and sub-folders inside a drive folder. Use `item_id="root"` to start at the top. |
| `sharepoint_get_file_metadata` | Returns name, size, MIME type, and timestamps for a file or folder — without downloading content. |
| `sharepoint_get_file_content` | Downloads and returns the text content of a file. Best for `.txt`, `.csv`, `.json`, `.xml`, `.md`. |
### Search
| Action | Description |
| ------------------------- | -------------------------------------------------------------------------------------------------------------------- |
| `sharepoint_search_files` | Full-text KQL search across all SharePoint files the user can access. Returns `driveId` and `itemId` for each match. |
***
## Prerequisites
Before using SharePoint actions in DeepMask, confirm the following are in place.
### Required Access
* **Microsoft 365 account (Business Basic or higher)** — Your organization must have an active Microsoft 365 or Office 365 subscription with SharePoint Online enabled.
* **User signed in via Microsoft** — Each user must sign in with their own Microsoft account through DeepMask's connector. Actions run under that user's identity and are limited to content they already have access to.
### What You Do Not Need
* No developer tools, code, or command-line experience
* No changes to existing SharePoint sites or permissions
* No service account or shared credentials
DeepMask uses OAuth 2.0 delegated authentication. When a user connects their account, they sign in with their own Microsoft credentials. DeepMask never stores passwords or receives broader access than the user already has in your tenant.
***
## Common Parameters
Several parameters appear across multiple actions.
| Parameter | Type | Description |
| ----------------- | -------------------------- | ----------------------------------------------------------------------------------------------------------------- |
| `response_format` | `"markdown"` \| `"json"` | Controls the output format. `"markdown"` (default) returns human-readable text; `"json"` returns structured data. |
| `limit` | number (1–100, default 20) | Maximum number of results to return in a single call. |
| `offset` | number (default 0) | Number of results to skip. Use with `limit` to page through large result sets. |
***
## Typical Workflows
### Browsing a site you know by name
Use `sharepoint_search_sites` with the site name as the query to get its `site_id`.
Call `sharepoint_list_drives` with the `site_id` to get the available drives and their `drive_id` values.
Call `sharepoint_list_folder` with a `drive_id` and `item_id="root"` to see top-level contents. Drill into sub-folders by passing the folder's `item_id`.
Call `sharepoint_get_file_content` with the `drive_id` and `item_id` of the target file.
### Finding a file by keyword
Call `sharepoint_search_files` with your search term. The results include `driveId` and `itemId` for each match.
Pass the `driveId` and `itemId` to `sharepoint_get_file_metadata` (for details) or `sharepoint_get_file_content` (for text content).
***
## Pagination
Actions that return lists (`sharepoint_search_sites`, `sharepoint_list_folder`, `sharepoint_search_files`) support pagination via `limit` and `offset`.
| Field in response | Description |
| ----------------- | ----------------------------------------------------------- |
| `count` | Number of items in this page |
| `total_count` | Total matching items (where available) |
| `has_more` | `true` if there are more pages |
| `next_offset` | Pass this as `offset` in the next call to get the next page |
***
## Security & Privacy
DeepMask is designed so that you retain full control of your data and your users' access. This section explains the key security properties of the SharePoint integration.
### Delegated Authentication
DeepMask uses OAuth 2.0 delegated permissions, not application-level (app-only) permissions. This means:
* Every action performed by DeepMask is done on behalf of the signed-in user.
* A user can only read SharePoint content they already have permission to access.
* DeepMask cannot bypass SharePoint's existing role-based access controls.
* Removing a user's SharePoint access in Microsoft 365 immediately removes their access in DeepMask.
### No Stored Credentials
DeepMask does not store your Microsoft password, your Client Secret, or raw SharePoint content. Authentication is handled entirely through short-lived OAuth access tokens and refresh tokens, which are encrypted at rest.
### No Service Account
Unlike some integrations that use a single shared service account to access all data, DeepMask authenticates each user individually. This ensures audit logs in your Microsoft 365 tenant accurately reflect which user accessed which content.
### Read-Only Permissions
All SharePoint actions are strictly read-only. DeepMask requests no write, delete, or administrative permissions — it cannot create, modify, or delete files, folders, or sites.
The delegated Microsoft Graph permissions used by these actions are:
| Permission | Actions that use it |
| ---------------- | --------------------------------------------- |
| `User.Read` | All actions (identity context) |
| `Sites.Read.All` | All site and drive actions |
| `Files.Read.All` | All file and search actions |
| `offline_access` | Keeps the user's session active between calls |
If a user reports missing sites or files, confirm that admin consent has been granted for all required permissions in your Azure AD app registration. See the [Microsoft Enterprise](/msft-enterprise) guide for details.
### Revoking Access
To disconnect DeepMask from SharePoint at any time:
* In DeepMask → **Connectors** → **Microsoft (Enterprise)**, click **Disconnect**.
Either action immediately revokes all access tokens. No data is retained after disconnection.
Questions about data residency, compliance, or security? Contact DeepMask support at [support@deepmask.io](mailto:support@deepmask.io).
# Search Files
Source: https://documentation.deepmask.io/connectors/SharePoint/search-files
`sharepoint_search_files`
Runs a full-text search across all SharePoint content the user can access, using Microsoft's Keyword Query Language (KQL). Returns file identifiers (`driveId` and `itemId`) that can be passed directly to `sharepoint_get_file_content` or `sharepoint_get_file_metadata`.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------------------------------------------------------------------------------------- |
| `query` | string | Yes | Full-text search query. Supports KQL syntax — e.g. `"budget"`, `"\"Q3 report\""`, `"filename:roadmap.docx"`. |
| `limit` | number | No | Max results to return (1–100, default 20). |
| `offset` | number | No | Results to skip for pagination (default 0). |
| `entity_types` | array | No | Graph entity types to include. Options: `"driveItem"`, `"listItem"`, `"list"`, `"site"`, `"drive"`. Defaults to `["driveItem"]`. |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A paginated list of search results. Each result includes `driveId`, `itemId`, `name`, `webUrl`, `mimeType`, `size`, `lastModifiedDateTime`, `parentPath`, and `siteId`. Includes `count`, `offset`, `has_more`, and `next_offset`.
KQL supports phrases (`"exact phrase"`), field-scoped queries (`filename:report.xlsx`), and boolean operators (`budget AND 2024`). For simple keyword searches, plain text works without any special syntax.
## Example use
> "Search for any SharePoint files mentioning 'annual review'."
# Search Sites
Source: https://documentation.deepmask.io/connectors/SharePoint/search-sites
`sharepoint_search_sites`
Searches for SharePoint sites by keyword. Returns site IDs, display names, and URLs. Use this when you know part of a site's name but not its ID. Pass the returned `id` to `sharepoint_get_site` or `sharepoint_list_drives` to go deeper.
## Parameters
| Parameter | Type | Required | Description |
| ----------------- | ------ | -------- | -------------------------------------------------- |
| `query` | string | Yes | Search term, e.g. `"marketing"` or `"finance"`. |
| `limit` | number | No | Max results to return (1–100, default 20). |
| `offset` | number | No | Results to skip for pagination (default 0). |
| `response_format` | string | No | Output format: `"markdown"` (default) or `"json"`. |
## Returns
A paginated list of matching sites. Each result includes `id`, `displayName`, and `webUrl`. The response also includes `total_count`, `count`, `offset`, `has_more`, and `next_offset` for pagination.
## Example use
> "Find me all SharePoint sites related to HR."
# DeepMask Security and Compliance Overview
Source: https://documentation.deepmask.io/enterprise/security-compliance
DeepMask runs on EU sovereign cloud infrastructure with GDPR compliance, enterprise-grade encryption, EU data residency, and a strict no-training policy.
DeepMask is a sovereign European AI workspace built with enterprise security requirements as a first principle, not an afterthought. Your data stays in the EU, is encrypted in transit and at rest, and is never used to train AI models. This page covers what that means in practice: the compliance certifications we hold and are working toward, how data is stored and processed, and the infrastructure partners that underpin our hosting.
## Security pillars
DeepMask is fully aligned with the General Data Protection Regulation. Your data is processed according to GDPR principles: lawfulness, purpose limitation, data minimization, and user rights. We do not transfer personal data outside the EU.
ISO 27001 certification is currently in progress. This internationally recognized standard for information security management systems sets the framework for how we manage risk, control access, and respond to incidents across the organization.
Your conversations, uploaded files, and project data are never used to train, fine-tune, or improve any AI model—by DeepMask or by any underlying model provider. This applies permanently and without exception.
All data is encrypted in transit using TLS and encrypted at rest using AES-256. This covers conversations, files, project instructions, and any other data you store or transmit through DeepMask.
DeepMask enforces strict data residency controls. Your data is processed and stored within the European Union, with German sovereign cloud infrastructure as the primary hosting environment.
Primary hosting runs on StackIT, the German sovereign cloud operated by Schwarz Group—one of Europe's largest infrastructure providers. This means your data is not subject to US cloud provider jurisdiction or extraterritorial data access laws.
## EU hosting infrastructure
DeepMask operates across a layered EU infrastructure designed for redundancy, compliance, and sovereignty.
StackIT is the German sovereign cloud operated by Schwarz Group, the corporate group behind Lidl and Kaufland. It is DeepMask's primary hosting partner for both cloud infrastructure and select LLM inference. Data processed on StackIT stays within Germany and is governed by German and EU law, not US cloud provider terms. Certain EU-hosted models—including Qwen3 (StackIT), Gemma 3 27B (StackIT), and GPT-OSS 120B (StackIT)—run directly on this infrastructure.
For select models not natively hosted on StackIT, DeepMask routes inference through EU-region deployments to ensure data processing remains within the European Union. This covers redundancy scenarios and scalability requirements while maintaining the same data residency guarantees.
Additional model endpoints are served through infercom's EU-hosted infrastructure with strict data residency controls. These endpoints provide access to additional models in the DeepMask catalog while keeping all inference within EU jurisdiction.
Models labeled "(StackIT)" or "(DeepMask)" in the model selector run on EU-sovereign infrastructure. You can identify these in the model list—for example, Kimi K2 (DeepMask), Qwen (DeepMask), Qwen3 (StackIT), Gemma 3 27B (StackIT), and GPT-OSS 120B (StackIT).
## GDPR compliance in detail
Being GDPR compliant means specific obligations are met at the infrastructure, product, and contractual level.
Enterprise customers receive a Data Processing Agreement (DPA) that documents how DeepMask processes personal data on your behalf, in line with GDPR Article 28 requirements. Contact the sales team to request a DPA for your organization.
DeepMask supports GDPR data subject rights including the right to access, rectify, and erase personal data. Requests can be directed to [contact@deepmask.io](mailto:contact@deepmask.io).
DeepMask does not transfer personal data to countries outside the EEA without appropriate safeguards. Our primary infrastructure (StackIT) and EU-region deployments ensure data stays within EU jurisdiction by default.
Data you submit to DeepMask is used solely to provide the service—to generate AI responses in your workspace. It is not analyzed for advertising, used for model training, or shared with third parties for their own purposes.
## ISO 27001 certification
ISO 27001 certification is currently in progress. DeepMask operates according to ISO 27001 principles and is actively undergoing the certification process. We will update this documentation when certification is formally achieved.
ISO 27001 defines a systematic approach to managing sensitive company and customer information. Our certification process covers risk assessment, access control policies, incident response procedures, and supplier security management. Enterprise customers who require evidence of our security posture during the certification period can request supporting documentation from the sales team.
## Questions and enterprise security reviews
For security questionnaires, DPA requests, penetration test reports, or any compliance-related inquiry, contact [contact@deepmask.io](mailto:contact@deepmask.io). For organizations requiring a dedicated security review session, [book a 30-minute call](https://cal.com/hissan-usmani/30min) with the team. We respond to all enterprise inquiries within 42 hours.
# Team
Source: https://documentation.deepmask.io/enterprise/team-management
Invite team members, assign admin and member roles, share projects, and oversee AI token usage across your entire organization from one workspace.
DeepMask Enterprise gives your organization a single workspace where admins can invite members, govern access, share projects across teams, and track how AI is being used—all without managing separate accounts or subscriptions per person. Instead of AI scaling with individuals, it scales with your company.
Enterprise plans are configured through our sales team. Contact us at [contact@deepmask.io](mailto:contact@deepmask.io) or [book a 30-minute call](https://cal.com/hissan-usmani/30min) to discuss seat counts, custom contracts, and pricing for your organization size.
## How enterprise differs from individual plans
Individual plans give a single user access to DeepMask's 25+ AI models, Projects, and MCP connectors. Enterprise plans add a management layer on top: your organization gets a shared workspace with centralized billing, team-level visibility into usage, and the ability to coordinate projects across departments.
One user, one workspace. Models, Projects, and MCP connectors are personal and not shared across colleagues.
Organization-wide workspace. Admins manage members, share projects, and monitor token usage across the whole team.
## Inviting team members
Navigate to your organization's settings from the sidebar. You need admin or owner permissions to manage members.
Enter the email addresses of the colleagues you want to invite. You can add multiple members at once.
Choose a role for each invitee—member or admin. Admins can invite others and access usage analytics; members can use the workspace and collaborate on shared projects.
DeepMask sends each invitee an email with a link to join the workspace. Pending invitations appear in your settings until accepted.
## Shared projects
Projects in DeepMask are persistent workspaces that store instructions, uploaded files, and conversation history across sessions. On Enterprise plans, you can share projects with teammates so the whole team works from the same context.
Upload requirement documents, briefs, and reference files once. Every team member working in the project has access to the same files and instructions.
Teammates can continue conversations inside a shared project, keeping all work organized in one place instead of scattered across individual chats.
Set project-level instructions that apply to every conversation in that project. Useful for enforcing tone, output format, or domain-specific context.
Each conversation in a project can use any of DeepMask's 25+ models. Switch mid-conversation to match the task—writing, coding, research, or analysis.
## Usage controls and oversight
Enterprise admins have access to the usage dashboard, which shows token consumption across the organization broken down by model. This gives you the data you need to understand how AI is being used and where spend is concentrated.
The usage dashboard shows Total Tokens consumed, split into Input Tokens and Output Tokens. A Model Breakdown view shows which models your team used most—useful for identifying whether usage is concentrated on a few high-cost models or distributed across the catalog.
Usage analytics are available to admins and owners. Standard members can see their own usage but not organization-wide figures.
Usage limits and budget controls are part of custom enterprise configuration. Contact the sales team to discuss options for your organization.
Use the Model Breakdown in the usage dashboard to identify which models drive the most token consumption. Shifting lower-stakes tasks to faster, more efficient models like Haiku 4.5 or Gemini 2.0 Flash can meaningfully reduce costs without impacting output quality.
## Get started with enterprise
DeepMask Enterprise plans are available for organizations of all sizes—from teams of 10 to companies with 1,000+ employees.
Send your requirements to [contact@deepmask.io](mailto:contact@deepmask.io). The team responds within 42 hours.
Schedule a 30-minute call to discuss your team's needs, seat counts, and custom pricing.
# Token Usage
Source: https://documentation.deepmask.io/enterprise/usage-analytics
Monitor total token consumption, input and output breakdowns, and per-model usage to control AI spend and demonstrate ROI across your organization.
DeepMask's usage analytics dashboard gives enterprise admins a clear view of how your team consumes AI across the platform. You can see total token usage, how tokens split between input and output, and a per-model breakdown showing which models your team relies on most. This data helps you manage AI spend, justify costs to stakeholders, and make informed decisions about which models to route different workloads to.
## What the usage dashboard shows
The dashboard surfaces three core metrics for your organization:
The combined count of all tokens processed across every conversation and project in your workspace during the selected period.
Tokens sent to the model—your messages, uploaded file contents, system instructions, and project context. Input tokens typically represent the majority of consumption.
Tokens generated by the model in its responses. Output tokens are smaller in volume but often priced higher by underlying model providers.
A per-model view showing how token consumption distributes across the 25+ models available in DeepMask—for example, Mistral Large, Opus 4.5, GPT-4, DeepSeek V3, and others your team has used.
## Reading the model breakdown
The model breakdown is the most actionable part of the usage dashboard. It shows you which models consumed the most tokens during a given period, so you can answer questions like:
* Are teams defaulting to the most expensive models for every task, including simple ones?
* Which departments or projects are the heaviest users of premium reasoning models?
* Has usage shifted since you onboarded a new team or started a new project?
A high input token count usually means your team is uploading large documents, using long system prompts, or working within projects with extensive context. Review project instructions and file uploads to ensure you're only passing context that the model actually needs.
If the breakdown shows most tokens concentrated on a single high-cost model, consider whether all those tasks require that capability. Switching routine writing or summarization tasks to a faster model reduces cost without degrading output quality for those use cases.
Spikes in output tokens can indicate conversations asking for very long responses—full reports, detailed code, or extended analysis. Check whether those outputs are being used or whether response length can be constrained with project instructions.
## Optimizing AI spend with usage data
Token usage data translates directly into cost. Here is how to use it to get better ROI from your DeepMask enterprise plan:
Review your first month of usage to understand your team's natural consumption patterns. Note which models dominate the breakdown and what total token volumes look like week over week.
DeepMask gives you access to 25+ models. Use the breakdown to identify tasks running on premium reasoning models that could be handled by faster, lower-cost alternatives. Routine drafting, translation, and summarization rarely need extended thinking models.
Input tokens often grow as projects accumulate files and instructions. Periodically review project files and instructions to remove outdated context. Lean project context means lower input token counts on every conversation.
Share usage data with department leads quarterly. Usage patterns often reflect workflow inefficiencies—teams that use AI heavily for tasks that could be templatized or automated with MCP connectors can reduce token usage while getting faster results.
Models like Haiku 4.5 and Gemini 2.0 Flash are well-suited for high-volume, lower-complexity tasks. Reserving extended thinking models like Opus 4.6 for genuinely complex reasoning tasks keeps your model breakdown balanced and your costs predictable.
## ROI context for enterprise teams
DeepMask is designed so that AI scales with your company, not just individuals. Usage analytics give you the organizational visibility to measure the return on your AI investment—not just whether people are using it, but how, and at what cost.
Token usage data is most useful when connected to business outcomes. Consider tracking usage alongside output metrics for your teams: volume of reports produced, code shipped, campaigns drafted, or research completed. This gives finance and leadership teams a concrete basis for evaluating the value DeepMask delivers against what it costs.
For custom usage reporting, consolidated billing, or help interpreting your analytics data, contact the sales team at [contact@deepmask.io](mailto:contact@deepmask.io) or [book a 30-minute call](https://cal.com/hissan-usmani/30min).
# AI Chat Workspace
Source: https://documentation.deepmask.io/features/chat-workspace
Start conversations, switch between 25+ AI models mid-chat, upload files, and use extended thinking — all from one unified interface.
The DeepMask chat workspace is your central hub for interacting with 25+ AI models from providers including OpenAI, Anthropic, Google, Mistral, DeepSeek, Grok, Meta, and Alibaba. You can start a conversation, swap models at any point without losing context, upload documents and images for analysis, and toggle advanced capabilities on demand — all without managing separate subscriptions or logins.
## Starting a new chat
Navigate to the DeepMask workspace. Click **New Chat** in the left sidebar to open a fresh conversation.
Use the model selector at the top of the chat input area to choose your starting model. A description and capability list appears for each model before you confirm.
Enter your prompt in the input field and press **Enter** or click the send button to begin the conversation.
## Switching models mid-conversation
You are never locked in to the model you started with. You can switch to a different model at any point in a conversation and the new model receives the full conversation history as context.
Click the model name displayed above the chat input field at any point during the conversation.
Select any available model from the list. The switch takes effect on your next message.
Switch to a reasoning-focused model like Kimi K2 or GPT-o3 Mini when a task requires multi-step analysis, then return to a faster model for follow-up questions.
## Extended thinking
Some models support extended thinking, which gives the AI additional processing time to reason through complex problems before responding.
Click the **Extended thinking** toggle in the chat toolbar, below the message input field.
Submit your message as normal. The model will show a thinking indicator while it works through the problem, then deliver a more thorough response.
Extended thinking is available on models that support it. The toggle is hidden automatically when the selected model does not support this capability.
## Uploading files
You can upload documents and images directly into a chat for analysis, summarization, or reference.
Click the attachment icon in the chat input area, or drag and drop a file into the chat window.
Type a prompt alongside the file to tell the model what you want — for example, "Summarize this document" or "What does this chart show?"
Submit the message. The model processes the file and responds based on its contents.
Uploaded files are processed within EU infrastructure and are never used to train AI models.
## Model capabilities
Each model in DeepMask lists the capabilities it supports. You can see these in the model selector before choosing.
Query the web in real time and receive cited, up-to-date answers without leaving the conversation.
Upload and analyze PDFs, Word files, text files, and other document formats.
Send images and ask the model to describe, extract data from, or reason about visual content.
Upload spreadsheets and CSVs for AI-generated charts, insights, and summaries.
Deep, multi-step logical reasoning for complex research, coding, or analytical tasks.
An interactive editing surface for drafting and refining long-form content collaboratively with the model.
# Data and Charts
Source: https://documentation.deepmask.io/features/data-visualization
Upload spreadsheets, CSVs, or documents and let AI generate charts, insights, and polished reports instantly — no coding or BI tools required.
DeepMask turns your raw data into decisions by letting you upload spreadsheets, CSVs, or documents directly into the chat and asking the AI to analyze them. The AI reads the data, identifies patterns and anomalies, generates charts and visualizations, and produces written insights or structured reports — all in plain language, with no coding, formulas, or external tools required.
## Supported file types
You can upload data in any of the following formats:
* **CSV files** — comma-separated values from any source
* **Spreadsheets** — Excel (.xlsx, .xls) and similar formats
* **Documents** — PDFs and Word files containing tabular data or figures
## How to analyze data
Start a new chat or open an existing conversation. You can also upload data files inside a Project to make them available across multiple threads.
Click the attachment icon in the chat input area and select your CSV, spreadsheet, or document. The file name and size appear as a preview before you send.
Type a prompt that tells the AI what to produce. You can ask for specific chart types, summary statistics, trend analysis, anomaly detection, or a full report.
The AI responds with charts, written insights, and any structured output you requested. You can ask follow-up questions to refine the analysis or request additional visualizations.
## Example: sales data to a line graph
This example from a real DeepMask session shows how a CSV upload becomes a visual report in a single prompt.
> **User:** Can you generate me the line graph for the visual representation of sales per month by analyzing the data I have just uploaded?
>
> **File attached:** MyShop\_Sales\_Sheet.CSV (12.4 KB)
>
> **Model:** Haiku 4.5
The AI analyzed the monthly sales figures and generated a line graph plotting values from January through December, with a peak of 119K marked at 12 July 2025. No formulas, no pivot tables, no charting software — just the file and a plain-language request.
Ask the AI to explain what it found as well as visualize it. Combining charts with written insights gives you a report you can share directly with stakeholders.
## What you can ask for
Line graphs, bar charts, scatter plots, and other visualizations generated directly from your uploaded data.
Written summaries of trends, outliers, and patterns — including percentage changes and notable data points.
Structured reports combining charts, narrative summaries, and KPI tables ready to share with your team.
The AI flags unusual values, unexpected drops or spikes, and data quality issues in your dataset.
## Which models support data analysis
Models that list **Data analyst** in their capability set support file-based data analysis and chart generation. You can check a model's capabilities in the model selector before starting a conversation.
Haiku 4.5, Kimi K2 (DeepMask), and other data-capable models are listed with the **Data analyst** capability badge in the DeepMask model selector.
Very large files may take longer to process. If your dataset is especially large, consider uploading a filtered or sampled version first to validate the analysis approach before running the full dataset.
# MCP Connectors
Source: https://documentation.deepmask.io/features/mcp-connectors
Link Microsoft Enterprise, OneDrive, Outlook, SharePoint and more to your AI workspace using the open MCP framework — directly from the chat interface.
MCP (Model Context Protocol) connectors let you link the tools your team already uses directly to DeepMask. Once connected, the AI can read, reference, and act on data from those systems during a conversation — without you having to copy-paste content or switch between tabs. Connections are managed from within the chat interface and are available across your projects and chat threads.
## Supported integrations
The **Connectors** panel lists all integrations currently available in your workspace.
Connect your organization's Microsoft 365 environment so the AI can work with your enterprise data.
Access documents and content stored in your organization's SharePoint sites. The AI can retrieve, summarize, and work with those files.
Access files and folders stored in your personal OneDrive account.
Reference emails and calendar events so the AI can help draft replies and surface relevant threads.
Generate and work with slide decks directly from your conversation.
Pull external web and market data into your workflow for research and analysis.
Give the AI access to files and folders in Google Drive to summarize documents, extract data, or draft content.
Connect your own MCP-compatible server to bring internal tools and data sources into DeepMask.
## How to add a connector
In the chat interface, click **Add connectors** in the toolbar below the message input. The connector panel lists all available integrations.
Click **Connect** next to the tool you want to link — for example, SharePoint or Outlook.
Follow the OAuth or credential flow for the selected service. DeepMask requests only the permissions it needs to read and interact with your data.
Once authenticated, the integration appears as an active connector. It is now available in your current chat and across your workspace.
Active connectors are shown with a count in the chat toolbar. For example, "2 Connectors" indicates two integrations are currently linked and available to the AI.
## Managing active connectors
Click **Manage connectors** in the chat toolbar to view all currently active integrations.
You can see which tools are connected and disconnect any integration at any time. Disconnecting removes the AI's access to that service immediately.
Connect only the tools relevant to the task at hand. You can add and remove connectors at any time, so there is no need to keep all integrations active simultaneously.
## Privacy and data access
Connector data is processed within DeepMask's EU-hosted infrastructure. DeepMask is fully GDPR compliant and never uses your connected data to train AI models. You control which integrations are active and can revoke access at any time.
When you connect a service, the AI gains access to the data that service exposes. Review the permissions requested during authentication and connect only accounts and scopes appropriate for your use case.
# Projects
Source: https://documentation.deepmask.io/features/projects
Create persistent AI workspaces in DeepMask that retain custom instructions, uploaded files, and multiple chat threads across every session for ongoing work.
Projects give you a persistent, context-rich workspace that remembers your instructions, files, and preferences across sessions. Instead of re-explaining background and re-uploading reference materials every time you start a chat, a project preserves everything in one place. You can run multiple focused chat threads within a single project and invite teammates to collaborate — making Projects the foundation for sustained, team-level AI work.
## What a project contains
Define standing context — goals, tone, constraints — that every chat thread in the project inherits automatically.
Attach reference documents, briefs, and data files once and reuse them across all threads in the project.
Run parallel conversations on distinct sub-topics, all sharing the same project context and files.
Invite teammates so everyone works from the same instructions, files, and conversation history.
## Creating a project
In the left sidebar, click **Projects** to view your existing projects or create a new one.
Click **New Project**, enter a name that reflects the work area (for example, "Automation Research Project"), and add an optional description.
Click **Create** to open the project workspace.
## Setting custom instructions
Custom instructions appear in every chat thread within the project, giving the AI consistent context without you having to repeat it each time.
Inside the project, locate the **Instructions** section in the project sidebar.
Describe the project's purpose, the AI's role, preferred output formats, and any constraints. For example: "Build an end-to-end automated reporting workflow that collects data from multiple sources, analyzes it, generates insights, and delivers structured reports with minimal manual effort."
Your instructions are saved immediately and applied to all current and future threads in this project.
Keep instructions focused on goals and constraints rather than procedural steps. The AI applies them as persistent context, not a one-time command.
## Uploading files to a project
Files uploaded to a project are available in every chat thread, so you only need to add them once.
In the project sidebar, click the **Files** area.
Drag and drop files or click to browse. Supported formats include PDFs, Word documents, spreadsheets, CSVs, and text files.
Start a chat thread and refer to the uploaded files by name or ask the AI to work with them directly — for example, "Analyze the Project Requirement Document."
## Working with multiple chat threads
Each chat thread within a project is a separate conversation that still inherits the project's instructions and files.
Inside the project, click **New Chat** to open a fresh thread. Give it a descriptive name so you can distinguish it from others.
Use separate threads for distinct sub-tasks. For example, an automation project might have threads named "Automation Workflow Setup," "Insights and Trend Analysis," "Report Structure & KPIs," and "Data Collection & Integration."
Click any thread name in the project sidebar to resume it from where you left off.
Each thread maintains its own conversation history. The project's custom instructions and files are shared across all threads, but individual thread contexts remain separate.
## Sharing with teammates
In the project, click the settings or share icon in the project header.
Enter your teammates' email addresses or select them from your organization's team directory.
Invited teammates can view the project's instructions, files, and chat threads, and start new threads within the shared project.
Project sharing is available on team plans. All shared data remains within your organization's EU-hosted DeepMask environment.
# Web Search
Source: https://documentation.deepmask.io/features/web-search
Enable Perplexity-powered web search in any chat to get cited, real-time answers grounded in current information without leaving the conversation.
DeepMask integrates Perplexity-powered web search directly into your AI conversations, delivering real-time, cited answers grounded in the latest information without leaving the chat. When web search is active, the model retrieves current content from the web before composing its response, so you get answers that go beyond the model's training data — including breaking news, recent research, live pricing, or any fast-moving topic.
## How to enable web search
Start a new chat or continue an existing one. Web search can be toggled at any point during a conversation.
In the chat toolbar below the message input, click **Web Search** to activate it. The toggle highlights when enabled.
Type your prompt and send it. The model searches the web, retrieves relevant sources, and includes citations in its response.
You can turn web search on or off between messages. Disable it when your question relies on the model's existing knowledge rather than current data — this tends to produce faster responses.
## What cited answers look like
When web search is active, the model's response includes numbered references to the sources it consulted. You can see which websites informed the answer and follow links to read the original content directly.
Citations are included automatically when web search is enabled. You do not need to ask the model to provide sources — they appear inline with every web-informed response.
## When to use web search
Ask about recent developments, policy changes, or events that occurred after the model's training cutoff. Web search ensures the answer reflects what is happening now, not what the model learned months or years ago.
Look up current pricing, product releases, company news, or industry reports. The AI synthesizes multiple sources into a single, coherent summary with citations you can verify.
Request the latest figures — exchange rates, population data, software version numbers, regulatory thresholds. Web search fetches live data rather than relying on potentially outdated training information.
Find recent papers, documentation updates, or technical specifications. Web search surfaces content published after the model's knowledge cutoff so your research stays current.
Cross-reference claims or double-check figures against live sources. The model retrieves and cites the sources it used, so you can review the evidence directly.
## Privacy and data handling
Web search queries are processed through Perplexity's infrastructure. Like all DeepMask features, your conversation data is never used to train AI models. DeepMask is fully GDPR compliant and EU-hosted.
# Welcome to DeepMask
Source: https://documentation.deepmask.io/introduction
DeepMask unifies 25+ leading AI models in one GDPR-compliant EU workspace. No vendor lock-in, no model training on your data, built for enterprise teams.
DeepMask is a unified AI workspace that gives your team access to more than 25 of the world's leading AI models — including GPT, Claude, Gemini, Mistral, DeepSeek, Kimi, Qwen, MiniMax, GLM, and Gemma — from a single interface at [chat.deepmask.io](https://chat.deepmask.io). This page explains what DeepMask is, what it can do, and who it is built for, so you can quickly understand whether it fits your organization's needs.
## What DeepMask is
DeepMask is a sovereign European AI workspace SaaS. It runs on StackIT — the German cloud platform operated by the Schwarz Group — alongside EU-region infrastructure from other hosting partners, ensuring your data never leaves European jurisdiction.
Fully aligned with GDPR for strong privacy and data protection. Your conversations and files are never used to train AI models.
Primary hosting on StackIT (Schwarz Group German cloud). EU-region deployments for select models with strict data residency controls.
Data encrypted in motion and at rest. ISO 27001 certification currently in progress.
DeepMask never uses your personal data to train or fine-tune AI models. User data remains private and secure.
## Key features
### 25+ AI models, one interface
DeepMask unifies top AI models so you can switch, compare, and choose the best tool mid-conversation — without juggling logins or separate subscriptions. Available models include GPT-4o, GPT-5.2, Claude Opus and Sonnet, Gemini 2.0 Flash and 2.5 Flash, Mistral, DeepSeek, Grok, Kimi K2, Qwen, Gemma, and more. EU-hosted variants of select models (Qwen3 via StackIT, Gemma 3 27B via StackIT, GPT-OSS 120B via StackIT) are available for stricter data residency requirements.
### Projects
Create persistent, context-rich workspaces that remember your instructions, files, and preferences across sessions. Upload documents, set custom instructions, and share projects with teammates for collaborative work.
### MCP connectors
Connect DeepMask to the tools your team already uses — Microsoft Enterprise, OneDrive, Outlook, SharePoint, and more — through the open MCP (Model Context Protocol) framework.
### AI-powered web search
DeepMask integrates Perplexity-powered web search directly into your AI conversations, delivering real-time, cited answers grounded in the latest information without leaving the chat.
### Response style controls
Choose from five response styles — Normal, Concise, Explanatory, Learning, or Formal — and DeepMask adapts its output instantly to match the task.
### Data visualization
Upload spreadsheets, CSVs, or documents and let AI handle the analysis. Get instant charts, insights, and polished reports — no coding required.
### Enterprise team management
Manage teams, projects, and usage across your organization. Track total token consumption and model distribution to understand ROI and enforce team-level policies.
## Who DeepMask is for
DeepMask is designed for organizations that need AI across multiple departments without building or maintaining their own infrastructure.
Draft campaigns, write product launch emails, generate social posts, and iterate on brand messaging — all with models tuned for fluent, natural writing.
Build workflows, automate reporting, analyze spreadsheets, and let AI handle repetitive operational tasks end-to-end.
Write code, debug, plan features, document systems, and manage product workflows inside a single AI workspace.
Analyze reports, extract key insights, and visualize trends — powered by web search and extended thinking.
Turn complexity into clarity. Analyze data, explore scenarios, generate insights, and support high-stakes decisions.
Deploy AI securely at scale with GDPR-compliant infrastructure, role-based team management, and full usage visibility.
## Explore the documentation
Sign up, create your workspace, pick a model, and send your first message in minutes.
Create your account, choose a plan, and configure your team or organization.
Explore every available model, understand its strengths, and learn how to choose the right one.
Build persistent workspaces with shared context, files, and team instructions.
Connect DeepMask to Microsoft Enterprise, OneDrive, Outlook, SharePoint, Google Drive (coming soon) and your other tools.
Manage teams, add new members, track usage, and review security & compliance settings.
# Choosing Right Model
Source: https://documentation.deepmask.io/models/choosing-a-model
Match your task to the right AI model in DeepMask. Compare models by use case — coding, research, writing, speed, reasoning, and EU-only data residency.
DeepMask is built around the principle that no single model is best for every task. You can switch models at any point — including mid-conversation — without losing your context. Use the tabs below to find the right model for what you're working on right now.
These models perform well on software engineering tasks including code generation, debugging, architecture design, and long-horizon autonomous development.
| Model | Why it fits |
| ----------------------- | ------------------------------------------------------------------------------------------------ |
| **GPT-5.2 / 5.3 / 5.4** | Most capable OpenAI models; strong reasoning and tool use across all coding tasks |
| **Sonnet 4.5 / 4.6** | Gold standard for autonomous coding; handles 30+ hour engineering sessions with 1M token context |
| **Gemini 2.5 Pro** | Large context window suits large codebase analysis and multi-file refactors |
| **Kimi K2 (DeepMask)** | Agent Swarm Mode enables 100 parallel sub-agents for complex, multi-step builds |
| **DeepSeek V3** | 671B MoE model with frontier-level coding and math; strong on STEM and security analysis |
| **MiniMax M2 / M2.1** | Built specifically for elite multi-language coding and advanced agent workflows |
These models excel at synthesizing large volumes of information, reasoning over documents, and producing structured analytical outputs.
| Model | Why it fits |
| ----------------------- | ----------------------------------------------------------------------------------------------------------- |
| **Kimi K2 (DeepMask)** | Searches hundreds of sources simultaneously in Agent Swarm Mode; 2M token context for massive document sets |
| **Opus 4.5 / 4.6** | Anthropic's highest-capability tier; strong on complex multi-document reasoning and subtle logical analysis |
| **GPT-5.2 / 5.3 / 5.4** | Full document and image analysis with strong reasoning across all content types |
| **DeepSeek V3** | Outperforms most frontier models on AIME and MATH-500; strong for STEM research and symbolic math |
| **Gemini 2.5 Pro** | 1M+ token context; suited for summarizing large research corpora |
| **Gemini 2.5 Flash** | Handles real-time summarization of hundreds of PDFs or hour-long recordings in one pass |
These models produce high-quality long-form text, adapt to different tones, and handle creative and professional writing tasks.
| Model | Why it fits |
| -------------------- | ------------------------------------------------------------------------------------------------------------------- |
| **GPT-5.2** | Highly capable for creative and persuasive writing; shown in-product generating marketing copy and strategy content |
| **Opus 4.5** | Nuanced, high-fidelity writing with strong narrative coherence; suited for strategy, legal, and financial documents |
| **Mistral Large 3** | Elite multilingual writing across 40+ languages; good for international marketing and professional content |
| **Sonnet 4.5** | Balanced between quality and speed; well-suited for content workflows requiring document context |
| **Mistral Medium 3** | Frontier-level writing output at significantly lower cost; good for high-volume content generation |
When you need quick responses, high throughput, or a cost-efficient model for simple queries and automation pipelines, these models deliver.
| Model | Why it fits |
| -------------------- | -------------------------------------------------------------------------------------------------------- |
| **Haiku 4.5** | Anthropic's fastest model at 180+ tokens/sec with 0.20s latency; designed for enterprise-scale workloads |
| **Gemini 2.5 Flash** | 185 tokens/sec with a 1M token context window; most cost-effective for large-volume document processing |
| **GLM-4.7 Flash** | Lightweight MoE model with strong reasoning and coding accuracy at high speed |
| **Mistral Medium 3** | 8× lower cost than frontier-tier models with strong general performance |
These models apply extended or structured thinking to work through complex, multi-step problems including math, logic, planning, and ambiguous tasks.
| Model | Why it fits |
| ----------------------- | ------------------------------------------------------------------------------------------------------------ |
| **Kimi K2 (DeepMask)** | High reasoning effort; 87.6% GPQA Diamond; decomposes tasks into 100 parallel sub-tasks |
| **Sonnet 4.5 / 4.6** | Adaptive reasoning (standard/high); 83.4% GPQA Diamond; significant gains on graduate-level math and science |
| **GPT-5.2 / 5.3 / 5.4** | Strong reasoning built into the latest GPT-5 generation |
| **GPT-o3 Mini** | Dedicated reasoning model for document analysis and tool use |
| **DeepSeek V3** | Adaptive non-thinking/thinking mode; 80.7% GPQA Diamond; excellent for mathematical proofs |
| **GLM-4.7** | Interleaved thinking with elite agent workflows; strong on complex real-world tasks |
If your organization requires that all data processing occurs within the European Union, every model in this tab runs exclusively on EU infrastructure.
| Model | Host | Notes |
| ---------------------------- | -------------------------------- | ------------------------------------------------- |
| **Qwen3 (StackIT)** | StackIT (Schwarz Group, Germany) | Reasoning, tool use, document and image analysis |
| **GPT-OSS 120B (StackIT)** | StackIT (Schwarz Group, Germany) | Document analysis and research; no image support |
| **Gemma 3 27B (StackIT)** | StackIT (Schwarz Group, Germany) | Lightweight; chat and document/image analysis |
| **DeepSeek V3.1 (Infercom)** | Infercom (EU-hosted endpoints) | Document analysis, tool use, complex writing |
| **GPT-OSS 120B (Infercom)** | Infercom (EU-hosted endpoints) | Document analysis and research; no image support |
| **MiniMax M2.5 (Infercom)** | Infercom (EU-hosted endpoints) | Coding, document analysis, tool use; 164K context |
| **Kimi K2 (DeepMask)** | DeepMask EU infrastructure | Full capability including Agent Swarm Mode |
| **Qwen (DeepMask)** | DeepMask EU infrastructure | Reasoning, tool use, multilingual chat |
StackIT is operated by the Schwarz Group (parent company of Lidl and Kaufland) and is certified as a German sovereign cloud. Infercom provides EU-hosted LLM endpoints with strict data residency controls. DeepMask's own infrastructure is also EU-based.
You can switch models at any point in a conversation. If a response isn't working well, select a different model from the model picker and continue your conversation — DeepMask carries your context forward automatically.
## Model capability quick reference
Extended thinking (sometimes shown as "reasoning mode" or "thinking mode" in the UI) causes a model to work through a problem step-by-step before producing its final answer. The model generates an internal chain of reasoning that it uses to improve accuracy on complex tasks.
Models in DeepMask that support extended or adaptive thinking include **Sonnet 4.5 / 4.6**, **Haiku 4.5**, **Kimi K2 (DeepMask)**, **GLM-4.7**, and **DeepSeek V3** (in its thinking mode). GPT-o3 Mini is also specifically optimized for reasoning tasks.
Extended thinking increases response time but significantly improves results for graduate-level math, multi-step logic, planning, and any task where intermediate reasoning matters.
The following models in DeepMask can accept images as input and reason about their contents:
* **OpenAI:** GPT-4o, GPT-4.1, GPT-5.2, GPT-5.3, GPT-5.4
* **Anthropic:** Opus 4.5 / 4.6, Sonnet 4.5 / 4.6, Haiku 4.5
* **Google:** Gemini 2.5 Pro, Gemini 2.5 Flash, Gemma 3 27B (StackIT)
* **MoonshotAI:** Kimi K2 (DeepMask), Kimi K2.5 (via MoonViT multimodal)
* **Alibaba:** Qwen (DeepMask), Qwen3 (StackIT)
* **Mistral:** Mistral Large 3
Models that do **not** support image input include GPT-OSS 120B (both StackIT and Infercom variants), DeepSeek V3 / V3.1, GPT-o3 Mini, and MiniMax M2.5 (Infercom).
Tool use (also called function calling) lets a model invoke external tools, APIs, or data connectors during a conversation. DeepMask exposes tool use through its MCP (Model Context Protocol) connector framework, which supports Microsoft Enterprise, OneDrive, Outlook, SharePoint and more.
Models with strong tool use support include:
* **Anthropic:** Haiku 4.5 (95% success rate on complex JSON schemas), Sonnet 4.5 / 4.6, Opus 4.5 / 4.6
* **OpenAI:** GPT-4o, GPT-4.1, GPT-5.x, GPT-o3 Mini
* **MoonshotAI:** Kimi K2 (DeepMask) — maintains coherence across 300+ sequential tool calls
* **Google:** Gemini 2.5 Pro, Gemini 2.5 Flash
* **Alibaba:** Qwen (DeepMask), Qwen3 (StackIT)
* **DeepSeek:** DeepSeek V3, DeepSeek V3.1 (Infercom)
* **MiniMax:** M2, M2.1, M2.5 (Infercom)
* **Z.ai:** GLM-4.7
Gemma 3 27B (StackIT) does not support tool use in DeepMask.
The context window is the maximum amount of text (measured in tokens, where 1 token ≈ 0.75 words) that a model can read and reason over in a single conversation. Larger context windows let you work with longer documents, more conversation history, and bigger codebases without losing earlier information.
Context window sizes for key models in DeepMask:
| Model | Context window |
| ----------------------- | ------------------------ |
| Kimi K2 (DeepMask) | 2,000,000 tokens |
| Sonnet 4.5 / 4.6 | 1,000,000 tokens |
| Gemini 2.5 Flash | 1,040,000 tokens |
| Gemini 2.5 Pro | \~1,000,000 tokens |
| Haiku 4.5 | 200,000 tokens |
| DeepSeek V3 / V3.1 | 128,000 – 164,000 tokens |
| MiniMax M2.5 (Infercom) | 164,000 tokens |
| GPT-4o | 128,000 tokens |
| GPT-4.1 | 128,000 tokens |
For very long documents or multi-session projects, prefer Kimi K2, the Sonnet series, or the Gemini 2.5 models. Use DeepMask Projects to persist files and instructions across sessions regardless of the model you choose.
# DeepSeek
Source: https://documentation.deepmask.io/models/deepseek
DeepSeek V3 and V3.1 (Infercom) on DeepMask. 671B MoE architecture delivering frontier-class coding, math, and document analysis without image support.
DeepSeek's V3 family redefines the cost-to-intelligence ratio using a 671B parameter Mixture-of-Experts architecture. DeepSeek V3 sets a new standard for efficient coding and STEM reasoning, while DeepSeek V3.1 (Infercom) builds on that foundation with a hybrid thinking mode and EU-hosted infrastructure via Infercom. Neither model supports image input, making both ideal for text and document-heavy workflows.
## About
DeepSeek V3 is a 671B parameter Mixture-of-Experts model that has set a new industry standard for efficiency. Its innovative Multi-head Latent Attention (MLA) architecture delivers frontier-class coding and math performance at a fraction of the hardware cost. It is widely regarded as the best model for developers who need maximum reasoning power at the lowest possible price.
DeepSeek V3 does not support image inputs. Use it for text, code, and document-based tasks.
## Key Capabilities
Outperforms most frontier models on the AIME and MATH-500 benchmarks for complex symbolic reasoning.
Highly effective at identifying vulnerabilities in C++, Rust, and Python codebases.
Delivers highly consistent reasoning across all query types, reducing unexpected output drift.
Efficient decoding architecture accelerates response times without losing precision.
## Use Cases
* **Low-cost coding agents** — Build production-grade code generators and automation pipelines with minimal per-task cost.
* **STEM research** — Solve complex engineering problems and symbolic math equations at scale.
* **Bulk data transformation** — Reformat and clean massive datasets with structural precision.
* **Document analysis** — Extract structured information from dense technical or legal documents.
DeepSeek V3 is your best choice when you need strong logic and coding ability at the lowest token cost. For EU-hosted deployments or hybrid thinking modes, use DeepSeek V3.1 (Infercom).
## Specifications
| Specification | Value |
| ---------------- | --------------------------------------------- |
| Model Provider | DeepSeek |
| Main Use Cases | High-Efficiency Agents, STEM, Bilingual Logic |
| Reasoning Effort | Adaptive (Non-Thinking / Thinking) |
| GPQA Diamond | 80.7% |
| Max Context | 128K – 164K Tokens |
| Latency (TTFT) | 0.41s |
| Throughput | 74 Tokens/sec |
## About
DeepSeek V3.1 (Infercom) is the updated "Terminus" release of the DeepSeek V3 family, refined for high-scale managed APIs (MaaS). It is a hybrid model that supports both a high-speed "Non-Thinking" mode for general chat and a deep "Thinking" mode for complex reasoning. The Infercom variant is specifically optimized for sub-second responses in autonomous agent and API integration scenarios.
DeepSeek V3.1 is EU-hosted via Infercom, making it a strong option for organizations with European data residency requirements. Image input is not supported.
## Key Capabilities
Toggle between fast non-thinking mode for quick answers and deep thinking mode for complex multi-step logic — all in one model.
The 3.1 update reduced time-to-answer for reasoning queries by 30% compared to earlier iterations.
Achieves 93.1% on AIME 2024, making it a price-performance leader for technical problem-solving.
Reliable document-to-JSON conversion with high accuracy for data pipeline workflows.
## Use Cases
* **High-volume API integration** — Provide smart reasoning for thousands of simultaneous users at a fraction of the cost of US-based models.
* **Bilingual RAG** — Excel at English-Chinese technical documentation and cross-border business intelligence.
* **Structured data extraction** — Run reliable document-to-JSON pipelines with the Infercom managed API.
* **Autonomous agents** — Deploy cost-efficient agentic loops that alternate between fast chat and deep reasoning as needed.
Choose DeepSeek V3.1 (Infercom) when you need EU data residency, hybrid thinking modes, or high-throughput API deployments. Its 32K tokens/sec throughput makes it one of the fastest options available on DeepMask.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------------- |
| Model Provider | DeepSeek |
| Main Use Cases | Bilingual API Dev, Low-Cost Reasoning |
| Hosting | EU-hosted via Infercom |
| Reasoning Effort | Hybrid (Think / Non-Think) |
| GPQA Diamond | 93.1% |
| Max Context | 164K Tokens |
| Latency (TTFT) | 0.21s |
| Throughput | 32K Tokens/sec |
# Gemini 2.5 Flash & Pro
Source: https://documentation.deepmask.io/models/gemini
Explore Gemini 2.5 Flash and 2.5 Pro on DeepMask. From high-throughput multimodal processing to deep reasoning, both models offer a 1M+ token context window.
Google's Gemini models bring two distinct capabilities to your DeepMask workspace: Gemini 2.5 Flash delivers high-throughput, low-cost multimodal processing at scale, while Gemini 2.5 Pro applies a native "Thinking" architecture to tackle complex reasoning, coding, and research tasks. Both models share a massive context window and full multimodal support for text, images, video, and audio.
## About
Gemini 2.5 Flash is Google's most efficient multimodal model, engineered for scale. It provides a massive 1-million-token context window at a fraction of the cost of Pro-tier models, and is specifically optimized for high-volume tasks such as real-time video summarization, large-scale document OCR, and high-speed data extraction. It is the most cost-effective way to process native audio and video inputs via API.
Gemini 2.5 Flash is served via Google's infrastructure. Your data is processed under DeepMask's EU data-handling agreements.
## Key Capabilities
Maintains near-perfect accuracy (99%+) when finding specific data points across a million tokens.
Processes video at 1 frame per second and audio at 16 kHz for high-fidelity temporal reasoning.
Store massive datasets — such as a 100-video training course — for rapid, cost-efficient recurring queries.
Supports real-time, low-latency multimodal interactions for voice assistants and live monitoring pipelines.
## Use Cases
* **Real-time customer support** — Power conversational bots that can understand user-uploaded screenshots or voice notes instantly.
* **Large-scale document synthesis** — Summarize hundreds of PDFs or hour-long meeting recordings in a single pass.
* **Multimodal agents** — Build assistants that can navigate your data across Gmail, Photos, and Workspace to perform complex cross-app tasks.
* **High-speed data extraction** — Process and reformat massive structured or semi-structured datasets with high throughput.
Use Gemini 2.5 Flash when you need to process large volumes of multimodal content at low cost. For complex reasoning or tasks requiring step-by-step logic, switch to Gemini 2.5 Pro.
## Specifications
| Specification | Value |
| ---------------- | --------------------------------------------------------------- |
| Model Provider | Google |
| Main Use Cases | Data Extraction, Real-time Summarization, Large Codebase Search |
| Reasoning Effort | Adaptive (Balanced) |
| GPQA Diamond | 68.3% |
| Max Context | 1.04M Tokens |
| Latency (TTFT) | 0.15s |
| Throughput | 185 Tokens/sec |
## About
Gemini 2.5 Pro is Google's most advanced reasoning model. Unlike the Flash variant, which prioritizes speed, the Pro model is engineered for "Thinking" — an internal process where it explores multiple solutions and verifies its own logic before responding. It features a standardized 1-million-token context window (scalable to 2M for select enterprise tiers) and is the primary model behind Deep Research and Agentic Coding workflows in the Gemini ecosystem.
Gemini 2.5 Pro is served via Google's infrastructure. Your data is processed under DeepMask's EU data-handling agreements.
## Key Capabilities
Built-in chain-of-thought reasoning lets the model pause and plan for complex queries without needing specialized prompts.
Ingests over 30,000 lines of code or 1,500 pages of text while maintaining perfect needle-in-a-haystack recall.
Leverages visual reasoning to interact with web browsers and software UIs, performing multi-step administrative tasks autonomously.
Processes up to 3,000 images, 1 hour of video, or 8 hours of audio simultaneously to find cross-modal patterns.
## Use Cases
* **Autonomous software engineering** — Debug and refactor entire code repositories by understanding full project architecture, not just individual functions.
* **Deep research and strategy** — Synthesize dozens of academic papers or financial reports into a comprehensive, cited brief.
* **Enterprise decision support** — Analyze dense legal contracts or medical records to identify subtle risks that smaller models might miss.
* **Personal intelligence** — Act as a proactive agent that manages your Google Workspace (Gmail, Docs, Drive) to organize schedules and complex data.
Use Gemini 2.5 Pro when accuracy and multi-step reasoning matter more than cost. Its Thinking architecture makes it ideal for tasks where you need the model to verify its own logic before responding.
## Specifications
| Specification | Value |
| ---------------- | -------------------------------------------- |
| Model Provider | Google |
| Main Use Cases | Professional Coding, Agentic Browser Control |
| Reasoning Effort | High |
| GPQA Diamond | 84.4% |
| Max Context | 1.04M – 2.0M Tokens |
| Latency (TTFT) | 0.45s |
| Throughput | 128 Tokens/sec |
# GLM-4.7 & GLM-4.7
Source: https://documentation.deepmask.io/models/glm
GLM-4.7 and GLM-4.7 Flash on DeepMask. Z.ai's bilingual models for advanced reasoning, agentic coding, UI generation, and high-speed automation workflows.
Z.ai's GLM-4.7 family brings two complementary models to DeepMask: the 358B flagship GLM-4.7 for deep reasoning, preserved thinking across long agentic workflows, and high-fidelity UI generation; and the lightweight GLM-4.7 Flash for high-volume, real-time automation where hundreds of small decisions are needed per minute. Both models offer strong bilingual (English/Chinese) performance and native support for interleaved thinking.
## About
GLM-4.7 is the 358B parameter flagship model from Z.ai. It achieves coding scores aligned with Claude Sonnet 4.5 and features "Preserved Thinking" for agentic workflows — maintaining a complex logical plan across hundreds of individual tool calls without losing track of the goal. It is particularly strong at bilingual English/Chinese reasoning, full-stack prototype generation, and high-fidelity UI/UX code generation.
GLM-4.7 is an open-source model from Z.ai. Its 200K context window and preserved thinking architecture make it a strong choice for long-horizon agentic tasks.
## Key Capabilities
Focuses on task completion rather than snippets — builds whole executable frameworks and app skeletons.
Strong understanding of UI/UX principles, producing well-structured and visually polished web layouts.
Leading performance in technical and legal English/Chinese translation and cross-language reasoning.
Executes 300+ sequential tool calls without losing track of the original goal or accumulated context.
## Use Cases
* **Full-stack prototype generation** — Create structurally complete, ready-to-run application skeletons from a description or diagram.
* **Multi-document content creation** — Generate 16:9 presentations and posters with coherent visual and logical structure.
* **Technical research** — Synthesize cross-border research papers across multiple languages into unified summaries.
* **Complex workflow automation** — Execute long multi-step agent workflows involving search, code execution, and document generation.
GLM-4.7 is the right choice when you need a model that can sustain a complex plan across many tool calls. Its preserved thinking architecture makes it particularly reliable for multi-step agentic tasks that would cause other models to drift.
## Specifications
| Specification | Value |
| ---------------- | --------------------------------------------------------- |
| Model Provider | Z.ai |
| Main Use Cases | Expert Coding, Complex Workflow Automation, STEM Research |
| Reasoning Effort | Adaptive (Standard/High) |
| GPQA Diamond | 85.7% |
| Max Context | 200K Tokens |
| Latency (TTFT) | 0.65s |
| Throughput | 76 Tokens/sec |
## About
GLM-4.7 Flash is the lightweight, high-speed variant of Z.ai's 4.7 series. It is engineered for action-first scenarios where a model needs to make hundreds of small decisions per minute. Its interleaved thinking capability allows it to output reasoning steps while performing tasks with minimal speed penalty, making it one of the most affordable and fastest options available on DeepMask for agent swarm deployments.
GLM-4.7 Flash is optimized for high-volume parallel deployments. Its low cost-per-token makes it practical to run dozens of instances simultaneously in agent swarm configurations.
## Key Capabilities
Outputs reasoning steps while performing tasks without a major speed penalty — reasoning and action in one pass.
Optimized for efficient generation in both English and Chinese for bilingual workflows.
Specifically tuned for repetitive search-and-extract workflows and high-frequency tool-calling patterns.
Designed for real-time chat and interactive applications where response speed is critical.
## Use Cases
* **Real-time data entry** — Process thousands of invoices or forms into structured databases at high throughput.
* **Massive web scrapers** — Summarize hundreds of search results in parallel across multiple agent instances.
* **Bilingual customer support** — Provide instant, context-aware translation and support in English and Mandarin.
* **Agent swarms** — Run dozens of parallel GLM-4.7 Flash instances for distributed task execution at low cost.
Use GLM-4.7 Flash when you need high-frequency, low-cost reasoning — especially for search-and-extract loops, bilingual support bots, or any scenario where you run many model instances in parallel. For tasks requiring deeper reasoning or UI generation, use GLM-4.7 instead.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------------------------------ |
| Model Provider | Z.ai |
| Main Use Cases | Real-time Agents, Local UI Gen, High-Speed Translation |
| Reasoning Effort | Standard |
| GPQA Diamond | 58.1% |
| Max Context | 203K Tokens |
| Latency (TTFT) | 0.59s |
| Throughput | 91 Tokens/sec |
# GPT-4.1
Source: https://documentation.deepmask.io/models/gpt-4-1
OpenAI's precision and context specialist. 1M token window, GPQA 66.6%, 0.62s TTFT, 91 tok/s. Best for large document processing and spreadsheets.
GPT-4.1 is OpenAI's 2025/2026 reliability update to the GPT-4 family, built for situations where precision and context depth matter more than raw reasoning power. With a standardized 1-million-token context window and industry-leading instruction adherence, it is the model to reach for when you need something that reliably "follows the rules" without over-explaining.
## About GPT-4.1
Where newer models focus on "thinking," GPT-4.1 focuses on **precision and context**. It achieves 99%+ "Needle in a Haystack" performance across its full 1M token range and scores 38% higher than GPT-4o on MultiChallenge — a benchmark measuring the ability to follow multi-turn constraints. It is significantly faster and more cost-efficient than the older GPT-4o, making it the preferred choice for developers who need reliable, high-volume processing.
## Key Capabilities
99%+ needle-in-a-haystack performance across the full 1M token range, ensuring nothing gets lost in long documents.
Scores 38% higher than GPT-4o on MultiChallenge — ideal for workflows with strict multi-turn constraints.
Native support for 110+ languages with culturally specific nuance, suited for enterprise localization pipelines.
Highly reliable at generating valid structured data for system integrations, reducing downstream parsing failures.
## Best For
GPT-4.1 is the right choice when you need to process very large documents, maintain strict output formats, or follow complex multi-turn instructions without drift. It handles spreadsheets, codebase audits, and content moderation pipelines particularly well. For tasks that require real-time voice or visual reasoning, GPT-4o is more appropriate. For the highest reasoning depth, consider the GPT-5 series.
When ingesting multiple large documents, batch them into a single request using GPT-4.1's 1M context window rather than chaining multiple calls — this preserves cross-document context and reduces cost.
## Use Cases
* **Log analysis** — Ingest months of server logs in a single pass to find root causes of errors.
* **Repository audits** — Index and summarize an entire company codebase for technical debt reviews.
* **Content moderation** — Process large batches of text and images with consistent judgment.
## Specifications
| Specification | Value |
| -------------- | ----------------------------------------------------------- |
| Provider | OpenAI |
| Context Window | 1.0M tokens |
| Reasoning | Medium-High |
| GPQA Diamond | 66.6% |
| Latency (TTFT) | 0.62s |
| Throughput | 91 tokens/sec |
| Key use cases | Long documents, spreadsheets, code refactoring, translation |
[Try GPT-4.1 in DeepMask →](https://chat.deepmask.io/)
# GPT-4o
Source: https://documentation.deepmask.io/models/gpt-4o
OpenAI's high-frequency multi-modal model for real-time voice, vision, and text interactions. 128K context, 0.12s TTFT, 112 tok/s, GPQA 74.0%.
GPT-4o is OpenAI's "high-frequency" multi-modal model, unifying text, audio, and vision in a single neural network optimized for low-latency, real-time interactions. With an average first-token latency of just 0.12 seconds, it comes close to matching human response times — making it the go-to choice whenever immediacy matters.
## About GPT-4o
Unlike models that bolt on audio or vision as afterthoughts, GPT-4o processes all three modalities natively. That means you get tighter coherence between what it sees, hears, and says. It is the high-frequency variant of the GPT-4o series, trading some raw reasoning depth for a level of speed and interactivity that no previous generation could match.
## Key Capabilities
Understands tone, background noise, and multiple speakers natively — without transcription as an intermediate step.
Expresses diverse speaking styles and emotions in real-time voice, making interactions feel natural rather than robotic.
Can watch a screen or camera feed and assist with tasks like math homework, software debugging, or live navigation.
Near-instant bidirectional translation across 50+ languages, suitable for live conversation and customer-facing apps.
## Best For
Choose GPT-4o when your application requires real-time, multi-modal interaction — live voice assistants, interactive tutors, accessibility tools, or gaming NPCs that need to see and respond within milliseconds. If you need deeper reasoning or very large context, consider GPT-4.1 or the GPT-5 series instead.
For voice applications, GPT-4o's native audio understanding means you can pass raw audio directly rather than pre-transcribing — this reduces latency and preserves prosodic signals like hesitation or emotion.
## Use Cases
* **Interactive tutors** — Provide real-time, voice-based feedback to students via audio and vision simultaneously.
* **Accessible assistants** — Help visually impaired users navigate their surroundings using a live camera feed.
* **Gaming NPCs** — Power non-player characters that can see, hear, and react to players in real time.
## Specifications
| Specification | Value |
| -------------- | --------------------------------------------- |
| Provider | OpenAI |
| Context Window | 128K tokens |
| Reasoning | Standard (Balanced) |
| GPQA Diamond | 74.0% |
| Latency (TTFT) | 0.12s |
| Throughput | 112 tokens/sec |
| Key use cases | Real-time voice, vision analysis, translation |
[Try GPT-4o in DeepMask →](https://chat.deepmask.io/)
# GPT-5 Series
Source: https://documentation.deepmask.io/models/gpt-5
GPT-5.2, 5.3, and 5.4 cover the full spectrum from production workhorse to autonomous agent with computer use. GPQA scores up to 92.4%.
The GPT-5 series represents OpenAI's frontier generation of models available in DeepMask. Each variant is tuned for a different point on the speed-intelligence-autonomy spectrum — GPT-5.2 as a fast, reliable production model; GPT-5.3 as an agentic coding specialist; and GPT-5.4 as the most capable general-purpose agent with native computer control. All three handle chat, document and image analysis, and tool use at the highest level.
## About the GPT-5 Series
GPT-5.2 is the "Goldilocks" model of the series — 25% faster than the original GPT-5 while maintaining the precision required for professional environments. GPT-5.3 is purpose-built for autonomous software engineering, with terminal-first training that lets it manage environments, run tests, and iterate on compiler errors. GPT-5.4 introduces a configurable reasoning effort parameter, native OS-level computer control, and a 33% reduction in hallucinations compared to earlier GPT-5 variants.
## Key Capabilities
Excels at complex Excel automation and multi-dimensional financial modeling, with reliable adherence to negative constraints.
GPT-5.3 natively understands shell commands, git workflows, and CI/CD logs — it doesn't just write code, it manages environments.
GPT-5.4 clicks, types, and navigates GUI applications natively without third-party wrappers, acting as a digital employee.
GPT-5.4 offers five reasoning levels from "none" (fast chat) to "xhigh" (deep research and verification), per request.
## Best For
Use **GPT-5.2** for everyday enterprise productivity — corporate communications, financial data extraction, and strategic brainstorming — where speed and reliability are equally important. Use **GPT-5.3** when you need autonomous software engineering: legacy migration, autonomous QA, or multi-language architecture work. Use **GPT-5.4** when you need the most capable, self-directed agent for end-to-end enterprise automation, deep scientific research, or massive cross-language refactoring.
For agentic coding tasks, GPT-5.3's terminal-first training makes it significantly more reliable than general-purpose models — set up your system prompt with your repo structure and let it iterate autonomously rather than prompting step by step.
## Use Cases
* **Corporate communications** — Drafting READMEs, documentation, and internal reports with precise tone control.
* **Financial data extraction** — Pulling granular data from dense PDF reports into structured JSON.
* **Legacy codebase migration** — Converting entire systems from outdated frameworks to modern stacks.
* **End-to-end enterprise automation** — Filing expense reports, configuring software, and running test suites via GUI (GPT-5.4).
* **Deep scientific research** — Multi-step data synthesis and architecture decisions using maximum reasoning effort.
## Specifications
| Specification | GPT-5.2 | GPT-5.3 | GPT-5.4 |
| -------------- | ---------------------------- | ------------------------------ | ------------------------------------ |
| Provider | OpenAI | OpenAI | OpenAI |
| Context Window | 400K tokens | 400K–1M tokens | 1.1M tokens |
| Reasoning | High | Medium | X-High (configurable) |
| GPQA Diamond | 92.4% | 91.5% | 92.0% |
| Latency (TTFT) | 0.55s | 0.18s | 0.55s |
| Throughput | 68 tokens/sec | 150+ tokens/sec | 65 tokens/sec |
| Key use cases | Agentic tasks, STEM research | Chat, agentic coding, research | Global codebases, PhD-level research |
[Try the GPT-5 series in DeepMask →](https://chat.deepmask.io/)
# GPT-o3 Mini
Source: https://documentation.deepmask.io/models/gpt-o3-mini
OpenAI's compact reasoning model for STEM and coding tasks. 200K context window, GPQA 79.7%, 0.25s TTFT, 141 tok/s. No image input support.
GPT-o3 Mini is OpenAI's compact, cost-efficient reasoning model — built to deliver PhD-level performance in STEM subjects at the speed of a small model. It replaces o1-mini in the 2026 lineup with higher rate limits, better tool integration, and an adaptive reasoning effort parameter that lets you dial in the right balance of speed and depth for each request.
## About GPT-o3 Mini
The o3 series brings chain-of-thought reasoning to a weight class that can run at scale. GPT-o3 Mini achieves o1-level reasoning at nearly 5x the speed, making it practical for production workloads that require real analytical depth — not just pattern matching. It supports function calling and Structured Outputs natively, and achieves an elite Codeforces rating that outperforms previous "mini" reasoning models. Note that image input is not supported.
## Key Capabilities
Delivers o1-level reasoning at nearly 5x the speed — making deep analytical thinking viable in latency-sensitive pipelines.
Native function calling and Structured Outputs support, with near-perfect reliability for API-driven workflows.
Solves 32%+ of research-level math problems on the first attempt when paired with Python tool execution.
Achieves elite Codeforces scores, outperforming previous mini-class reasoning models on algorithmic problem solving.
## Best For
GPT-o3 Mini is ideal for STEM-intensive tasks where you need genuine reasoning depth but cannot afford the latency or cost of a full frontier model. It is well-suited for real-time tutoring, fast debugging, and structured data extraction. It does not support image inputs — for multi-modal reasoning, use GPT-4o or the GPT-5 series instead.
Use the adaptive reasoning effort parameter deliberately: set it to "low" for simple classification or extraction tasks, and reserve "high" for math-heavy or multi-step logic problems. This can cut costs significantly without sacrificing output quality where it counts.
## Use Cases
* **Real-time tutoring** — Instant feedback on complex physics or calculus problems during live sessions.
* **Fast debugging** — Identifying logic errors in scripts with minimal latency.
* **Structured data extraction** — Pulling complex variables from messy text into precise JSON via function calling.
## Specifications
| Specification | Value |
| -------------- | -------------------------------------------------- |
| Provider | OpenAI |
| Context Window | 200K tokens |
| Reasoning | Adaptive (Low, Medium, High) |
| GPQA Diamond | 79.7% (High effort) |
| Latency (TTFT) | 0.25s |
| Throughput | 141 tokens/sec |
| Image support | No |
| Key use cases | STEM tasks, competitive coding, structured outputs |
[Try GPT-o3 Mini in DeepMask →](https://chat.deepmask.io/)
# GPT-OSS 120B
Source: https://documentation.deepmask.io/models/gpt-oss-120b
OpenAI's open-source 120B model, EU-hosted via StackIT and Infercom. GPQA 80.9%, transparent chain-of-thought, adaptive reasoning effort levels.
GPT-OSS 120B is OpenAI's 2026 open-source contribution to the frontier model ecosystem, available in DeepMask through two EU-hosted infrastructure providers: StackIT and Infercom. It delivers GPT-4-tier intelligence under an Apache 2.0 license, with full chain-of-thought transparency and adjustable reasoning effort — making it the model of choice for organizations that need frontier-class AI without black-box opacity or data leaving European infrastructure.
## About GPT-OSS 120B
Built on a Mixture-of-Experts (MoE) architecture, GPT-OSS 120B uses sparse activation to stay fast and efficient — activating only a fraction of its parameters per request. The Infercom variant is optimized for high-throughput deployments, reaching up to 544 tokens/sec. The StackIT variant is tuned for sovereign enterprise deployments with a focus on transparent reasoning and strict schema enforcement. Neither variant supports image inputs.
Both the StackIT and Infercom variants of GPT-OSS 120B are hosted entirely within the European Union, making them suitable for use cases governed by GDPR and sector-specific data residency requirements.
## Key Capabilities
Full visibility into internal reasoning steps — critical for legal, medical, and compliance use cases where "black box" AI is unacceptable.
Switch between Low (fast), Medium (balanced), and High (deep analytical thinking) per request to control cost and latency.
Native strict schema enforcement ensures near-perfect reliability for API-driven agents and structured output pipelines.
The Infercom variant exceeds 500 tokens/sec on optimized stacks — one of the fastest models in its weight class.
## Best For
GPT-OSS 120B is ideal when you need frontier-level reasoning on-premises or within EU-hosted infrastructure. It is the right choice for legal and clinical workflows where reasoning transparency is mandatory, for privacy-sensitive production environments in finance and healthcare, and for high-volume agentic pipelines that need both speed and analytical depth. It does not support image inputs or tool use in the DeepMask interface — for those capabilities, see GPT-4o or the GPT-5 series.
For legal and compliance workflows, use High reasoning effort to maximize analytical depth. For high-volume document classification or extraction pipelines, Medium effort typically provides the best cost-per-quality tradeoff.
## Use Cases
* **Clinical summarization** — Processing patient histories locally under HIPAA- or GDPR-equivalent data residency requirements.
* **Legal research** — Analyzing sensitive litigation documents without any cloud exposure outside the EU.
* **Local coding assistants** — Running a high-intelligence coding model entirely on private, EU-resident infrastructure.
* **STEM and technical research** — Graduate-level science and mathematics reasoning with verifiable reasoning steps.
## Specifications
| Specification | StackIT | Infercom |
| -------------- | ---------------------------------- | -------------------------------------------- |
| Provider | OpenAI (open-source) | OpenAI (open-source) |
| Hosting | EU (StackIT) | EU (Infercom) |
| Context Window | 131K tokens | 131K tokens |
| Reasoning | High | Adaptive (Low, Medium, High) |
| GPQA Diamond | 80.9% | 80.9% |
| Latency (TTFT) | 0.27s | 0.37s |
| Throughput | 262 tokens/sec | 313–544 tokens/sec |
| Image support | No | No |
| Key use cases | Agentic security, sovereign DevOps | High-speed agents, API orchestration, coding |
[Try GPT-OSS 120B in DeepMask →](https://chat.deepmask.io/)
# Haiku 4.5
Source: https://documentation.deepmask.io/models/haiku
Claude Haiku 4.5 by Anthropic — fastest Anthropic model. 200K context, GPQA 73.2%, 0.20s TTFT, 180+ tok/s. Ideal for high-volume tasks and real-time agents.
Haiku 4.5 is Anthropic's most efficient model — designed to deliver 2025 flagship-class intelligence at a fraction of the latency and cost. With a first-token latency of 0.20 seconds and throughput exceeding 180 tokens/sec, it is the first choice for real-time UI agents, high-frequency content moderation, and enterprise workloads that require millions of requests per hour without performance degradation.
## About Haiku 4.5
Released in late 2025, Claude Haiku 4.5 brings enhanced computer use capabilities to the lightweight tier — making it practical not just for chat and classification, but also for simple navigation and form-filling in web browsers and desktop environments. It achieves a 73.2% GPQA Diamond score and a 95% success rate on complex JSON schemas, making it reliable for structured tool-calling pipelines at high volume.
## Key Capabilities
Optimized for high-speed function execution with a 95% success rate on complex JSON schemas — reliable enough for production API pipelines.
Handles simple navigation and form-filling within web browsers and desktop environments without requiring a heavier model.
Sustains 180+ tokens/sec across enterprise workloads that require millions of requests per hour without performance degradation.
Context-aware safety filters reduce false positives by 40% compared to previous versions, making moderation pipelines more precise.
## Best For
Haiku 4.5 is the right choice when speed and cost efficiency matter more than maximum reasoning depth. It is ideal as the execution layer in multi-agent systems — acting as the fast "hands and eyes" for a larger orchestrating model like Opus or Sonnet. For tasks that require deeper reasoning, complex document analysis, or autonomous multi-step workflows, Sonnet 4.5 or 4.6 is the appropriate step up.
In multi-agent architectures, consider using Haiku 4.5 to handle repetitive, low-level sub-tasks (data labeling, form extraction, moderation checks) while reserving Sonnet or Opus for planning, summarization, and complex reasoning steps. This pattern can reduce overall costs by 60–80%.
## Use Cases
* **Customer support chatbots** — Near-instant, high-quality responses for global user bases at scale.
* **Large-scale data labeling** — Classifying millions of records with high accuracy for research and model training pipelines.
* **Sub-agent swarms** — Acting as the fast execution layer for larger orchestrating models handling repetitive, low-level tasks.
* **Real-time content moderation** — Context-aware filtering for user-generated content platforms requiring high throughput.
## Specifications
| Specification | Value |
| -------------- | ------------------------------------------------------ |
| Provider | Anthropic |
| Context Window | 200K tokens |
| Reasoning | Adaptive (Standard/High) |
| GPQA Diamond | 73.2% |
| Latency (TTFT) | 0.20s |
| Throughput | 180+ tokens/sec |
| Key use cases | High-volume support, coding sub-agents, real-time data |
[Try Haiku 4.5 in DeepMask →](https://chat.deepmask.io/)
# Kimi K2 & K2.5
Source: https://documentation.deepmask.io/models/kimi-k2
Kimi K2 and Kimi K2.5 on DeepMask. 1-trillion parameter MoE models with Agent Swarm Mode, multimodal reasoning, and 2M token context for complex agentic tasks.
MoonshotAI's Kimi K2 family introduces a new paradigm for agentic AI. Both models are built on a 1-trillion parameter Mixture-of-Experts architecture and natively support Agent Swarm Mode — coordinating up to 100 parallel sub-agents to decompose and execute complex tasks. Kimi K2 (DeepMask) is optimized for deep reasoning and long-context retrieval at 2M tokens, while Kimi K2.5 extends this with a focus on visual-to-code generation and full-stack prototyping from UI screenshots and video.
## About
Kimi K2 (DeepMask) is MoonshotAI's massive Mixture-of-Experts breakthrough. It is the first model to natively support Agent Swarm Mode, allowing the main model to coordinate up to 100 specialized sub-agents working in parallel. Despite its 1-trillion total parameters, it activates only 32B per request — making it exceptionally efficient. It supports a 2M-token context window and uses MoonViT for native multimodal processing across text, images, and video.
Kimi K2 (DeepMask) is available directly on DeepMask with a 2M-token context window and 256K-token support on standard hardware via Multi-Head Latent Attention.
## Key Capabilities
Decomposes a complex task into up to 100 parallel sub-tasks and executes them simultaneously.
Processes text, images, and video with equal fluency through MoonshotAI's multimodal architecture.
Handles up to 2 million tokens, enabling analysis of massive documents and long-running agent sessions.
Maintains coherence across 300+ sequential tool calls without logic drift.
## Use Cases
* **Massive research synthesis** — Search hundreds of web sources simultaneously to compile comprehensive reports.
* **Vision-to-code** — Upload a UI walkthrough video and have Kimi K2 rebuild the entire website structure.
* **Batch data processing** — Analyze thousands of legal or medical records in a single swarm session.
* **Multi-step agent workflows** — Run long-horizon tasks across tools, documents, and APIs without losing track of the goal.
Kimi K2 (DeepMask) is the strongest choice on DeepMask when you need to run complex, long-horizon tasks that span multiple tools and document types. Its 0.31s TTFT and 111 tokens/sec throughput make it fast enough for interactive use.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------ |
| Model Provider | MoonshotAI |
| Main Use Cases | Visual Debugging, Agent Swarms |
| Reasoning Effort | High |
| GPQA Diamond | 87.6% |
| Max Context | 2.0M Tokens |
| Latency (TTFT) | 0.31s |
| Throughput | 111 Tokens/sec |
## About
Kimi K2.5 by MoonshotAI is a 1-trillion parameter MoE model built for the agentic era. It shifts from single-prompt interactions to a self-directed paradigm where it decomposes a project into parallel sub-tasks. Trained on 15 trillion mixed visual and text tokens, it is the top choice for visual-to-code generation, complex UI automation, and full-stack prototyping directly from screenshots or Figma links.
Kimi K2.5 is a powerful open-source model. Its multimodal training on 15 trillion tokens makes it particularly strong for tasks that combine visual and code reasoning.
## Key Capabilities
Decomposes one goal into a coordinated team of sub-agents working in parallel across search, code, and terminal tools.
Generates functional code directly from a UI screenshot or design link.
Browses the live web visually — clicking buttons and filling forms like a human user.
Integrates document analysis with real-time image search to solve data-heavy, cross-modal problems.
## Use Cases
* **Visual asset refinement** — Autonomously find, edit, and place assets into a web layout.
* **Complex tool-augmented search** — Manage 100+ steps across search and terminal tools to verify scientific data.
* **Full-stack prototyping** — Build functional web and mobile apps from zero to one, including backend logic.
* **UI automation** — Navigate live interfaces and perform multi-step workflows like a human operator.
Use Kimi K2.5 when your task starts with a visual input — a screenshot, video walkthrough, or Figma design — and ends with running code or a deployed prototype. It is the strongest model on DeepMask for visual-to-code generation.
## Specifications
| Specification | Value |
| ---------------- | -------------------------------------- |
| Model Provider | MoonshotAI |
| Main Use Cases | Visual Debugging, Coding, Agent Swarms |
| Reasoning Effort | High |
| GPQA Diamond | 87.6% |
| Max Context | 262K Tokens |
| Latency (TTFT) | 0.52s |
| Throughput | 138 Tokens/sec |
# MiniMax M2, M2.1 & M2.5
Source: https://documentation.deepmask.io/models/minimax
MiniMax M2, M2.1, and M2.5 on DeepMask. Expert MoE models for full-stack coding, mobile development, and EU-hosted agentic workflows with interleaved thinking.
MiniMax's M2 model family brings a distinct agentic philosophy to DeepMask: all three models are built around Interleaved Thinking, maintaining coherent state across multi-turn tool interactions without logic drift. MiniMax M2 focuses on full-stack development and office automation. M2.1 extends this to mobile app development and 3D visualization. MiniMax M2.5 (Infercom) adds EU hosting via Infercom with a massive 1M-token context window optimized for long-running autonomous agents.
## About
MiniMax M2 is an expert-level Mixture-of-Experts model built from the ground up for the agent universe. It introduces Interleaved Thinking, where it natively uses internal planning steps to separate its reasoning from its final output. Trained via a Forge RL framework across 200,000+ complex environments, it is highly optimized for agentic loops — tasks where the model must search, act, and reason repeatedly to solve a problem.
MiniMax M2 provides native support for generating and editing high-fidelity Office documents (Word, PowerPoint, Excel) — a capability not found in most other models on DeepMask.
## Key Capabilities
Optimized for reliable task execution across complex, real-world agentic environments.
Maintains coherent state across multi-turn tool interactions, reducing logic drift in long agentic loops.
Sees UI screenshots and translates them into executable code or precise navigation steps.
Natively generates and edits Word, PowerPoint, and Excel files from natural language instructions.
## Use Cases
* **Autonomous office assistants** — Build complex financial models in Excel or strategy decks in PowerPoint from natural language instructions.
* **Full-stack web development** — Write 1,000+ line TypeScript files with an 80%+ first-run pass rate.
* **Strategy consulting** — Synthesize massive market datasets into professional presentations automatically.
* **Agent scaffolding** — Build reliable multi-step agentic systems that loop across search, code execution, and document generation.
Use MiniMax M2 when your workflow involves repeated search-act-reason cycles, especially tasks that produce Office documents or require long-horizon coherence across many tool calls.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------------------------- |
| Model Provider | MiniMax |
| Main Use Cases | Efficient Coding, Agent Scaffolding, MoE Research |
| Reasoning Effort | Interleaved Thinking |
| GPQA Diamond | 78.2% |
| Max Context | 196.6K Tokens |
| Latency (TTFT) | 0.35s |
| Throughput | 95 Tokens/sec |
## About
MiniMax M2.1 is a specialized model designed to close the mobile development gap in AI. While most models focus on Python and web, M2.1 is fine-tuned for Swift (iOS), Kotlin (Android), and 3D visualization (Three.js). It is the premier model for "Vibe Coding" — describing an app's aesthetic and interaction logic and having the AI build the entire functional package, including backend logic.
MiniMax M2.1 is the strongest model on DeepMask for native iOS and Android app development. Its training focus on Swift and Kotlin sets it apart from general-purpose coding models.
## Key Capabilities
Outperforms other models in building functional Android and iOS application logic from natural language descriptions.
Generates complex 3D web scenes with physics simulation and collision detection.
Bridges backend and frontend languages in a single, coherent development workflow.
Understands aesthetic feel and UX principles, generating UIs that are visually polished and user-friendly.
## Use Cases
* **Rapid app prototyping** — Turn a two-paragraph idea into a downloadable iOS or Android mockup.
* **Game development tools** — Create browser-based 3D simulations and mini-games with physics-aware logic.
* **Enterprise office automation** — Develop custom internal tools for complex Excel and CRM data management.
* **Full-stack vibe coding** — Describe an app's look and feel in natural language and receive complete, functional source code.
MiniMax M2.1 is the best choice on DeepMask for mobile app development and 3D web experiences. If your project targets iOS, Android, or requires Three.js, this model will significantly outperform general-purpose alternatives.
## Specifications
| Specification | Value |
| ---------------- | --------------------------------------------------- |
| Model Provider | MiniMax |
| Main Use Cases | Multilingual AppDev, Full-Stack Agents, Vibe Coding |
| Reasoning Effort | Adaptive (Concise Thinking) |
| GPQA Diamond | 83.0% |
| Max Context | 205K Tokens |
| Latency (TTFT) | 0.25s |
| Throughput | 113 Tokens/sec |
## About
MiniMax M2.5 (Infercom) is a 229B parameter Mixture-of-Experts model utilizing a breakthrough Hybrid Attention architecture — a 7:1 ratio of Lightning to SoftMax attention — to provide linear scaling for long contexts. The Infercom variant is EU-hosted and specifically optimized for sub-second responses in messaging-based autonomous agents and high-traffic production systems.
MiniMax M2.5 (Infercom) is EU-hosted via Infercom, providing European data residency for organizations with compliance requirements. Image input is not supported on this variant.
## Key Capabilities
Industry-leading retrieval across its 1M token context window, virtually eliminating lost-in-the-middle errors.
Optimized for multi-step tool-calling sequences for high-reliability task execution in production agent loops.
Delivers 100+ tokens/sec while maintaining EU data residency via Infercom hosting.
Well-suited for long-running AI assistants that need to retain context across extended sessions.
## Use Cases
* **24/7 messaging agents** — Run high-traffic customer support and sales bots where cost-per-token is a critical business factor.
* **Full-stack vibe coding** — Prototype and iterate on code generation tasks with a 1M-token context for large codebases.
* **Persistent memory systems** — Build long-running AI assistants that remember context across extended sessions.
* **Efficient RAG** — Power retrieval-augmented generation pipelines at scale with EU data residency.
Use MiniMax M2.5 (Infercom) for production agentic systems that need EU hosting, a massive context window, and high throughput at reasonable cost. Its 1M-token context and linear scaling make it well-suited for persistent, long-running assistants.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------------------------ |
| Model Provider | MiniMax |
| Hosting | EU-hosted via Infercom |
| Main Use Cases | Multi-step Agents, Efficient RAG, Knowledge Work |
| Reasoning Effort | Adaptive (Concise) |
| GPQA Diamond | 80.0% |
| Max Context | 1.0M Tokens |
| Latency (TTFT) | 1.17s |
| Throughput | 100+ Tokens/sec |
# Mistral Large 3
Source: https://documentation.deepmask.io/models/mistral
Mistral Large 3 and Mistral Medium 3 on DeepMask. State-of-the-art open-weight models for multilingual reasoning, coding, and scalable enterprise workflows.
Mistral AI's Large 3 and Medium 3 models give you a choice between maximum capability and maximum efficiency. Mistral Large 3 is a 675B parameter open-weight powerhouse optimized for multilingual enterprise workflows and data sovereignty, while Mistral Medium 3 delivers frontier-level performance at a fraction of the cost — ideal for high-throughput production systems. Both models are built for European enterprise requirements.
## About
Mistral Large 3 is a state-of-the-art 675B parameter Mixture-of-Experts (MoE) model from Mistral AI. It is currently the top-ranked open-weight model globally, with 41B active parameters and a 256K context window. It provides a no-compromise open-source alternative to proprietary frontier models for organizations requiring full data sovereignty, high-fidelity reasoning, and multilingual capability across 40+ languages.
As a Mistral AI model, Mistral Large 3 is available for on-premise and private cloud deployments, making it one of the strongest choices for European data sovereignty requirements.
## Key Capabilities
Exceptional performance across 40+ languages, with strong results in French, German, Spanish, and Arabic.
Extracts structured data from complex financial reports and scanned documents with high fidelity.
Achieves approximately 92% on HumanEval, rivaling frontier models in clean, idiomatic code generation.
The premier choice for deployments where data privacy and EU residency are non-negotiable.
## Use Cases
* **Global enterprise automation** — Manage multilingual customer support and legal workflows across international borders.
* **Technical document synthesis** — Digest 200+ page engineering manuals to provide precise architectural guidance.
* **Private RAG systems** — Power internal knowledge bases where data must remain behind a corporate firewall.
* **Complex coding workflows** — Build and refactor production codebases with near-frontier coding performance.
Mistral Large 3 is the right choice when you need the strongest possible open-weight reasoning, especially for multilingual tasks or deployments where full data control is required.
## Specifications
| Specification | Value |
| ---------------- | --------------------------------- |
| Model Provider | Mistral AI |
| Main Use Cases | Complex Coding, Multi-Step Agents |
| Reasoning Effort | Adaptive (Standard/High) |
| GPQA Diamond | 78.9% |
| Max Context | 262K Tokens |
| Latency (TTFT) | 0.55s |
| Throughput | 36 Tokens/sec |
## About
Mistral Medium 3 is designed as the perfect balance for enterprise production. It achieves or exceeds 90% of the benchmark performance of much larger models while being significantly less expensive. It is specifically optimized for professional use cases — coding, STEM, and multimodal understanding — where latency and cost are as important as raw intelligence. It supports seamless switching between cloud API and local VPC or on-premises deployments.
Mistral Medium 3 supports hybrid deployment across cloud APIs and on-premises infrastructure, enabling flexible European data residency configurations.
## Key Capabilities
Can be fine-tuned into private knowledge bases for domain-specific mastery and custom use cases.
Performs exceptionally well in physics and engineering tasks, nearly matching Large-class models at a fraction of the cost.
High-reliability function calling and structured output for enterprise systems integration.
Available via cloud API or on-premises setups — supports European data residency configurations.
## Use Cases
* **Domain-specific experts** — Deploy custom fine-tuned instances for legal advice, medical diagnostics, or technical support.
* **High-throughput coding assistants** — Power enterprise-wide code generation with low latency and cost.
* **Complex data analysis** — Enrich customer service pipelines with deep context from massive enterprise datasets.
* **Math reasoning** — Handle STEM-heavy tasks in physics and engineering at near-Large model quality.
Use Mistral Medium 3 when you need reliable frontier-class performance at scale without the cost of a full Large model. It is a strong default for production deployments where throughput and cost efficiency matter.
## Specifications
| Specification | Value |
| ---------------- | --------------------------------------------------- |
| Model Provider | Mistral AI |
| Main Use Cases | Enterprise Chat, Math Reasoning, On-Prem Deployment |
| Reasoning Effort | Standard (Native) |
| GPQA Diamond | 57.8% |
| Max Context | 131K Tokens |
| Latency (TTFT) | 0.42s |
| Throughput | 49 Tokens/sec |
# Opus (4.6, 4.5)
Source: https://documentation.deepmask.io/models/opus
Claude Opus 4.5 and 4.6 from Anthropic. Up to 91.3% GPQA, 1M context, adaptive reasoning. Best for demanding coding, research, and agents.
The Opus family represents the ceiling of Anthropic's model intelligence within DeepMask. Both Opus 4.5 and Opus 4.6 are designed for the most demanding agentic, research, and software engineering tasks — where accuracy, long-horizon coherence, and the ability to handle ambiguity without losing goal-state are non-negotiable. If your task is complex, high-stakes, and requires sustained autonomous effort, Opus is the right choice.
## About Opus 4.5 and Opus 4.6
**Opus 4.5** is Anthropic's heavyweight frontier model, optimized for heavy-duty agentic workflows, complex software engineering, and deep research. It introduces a dynamic "Effort Control" parameter so you can minimize token spend on standard tasks or maximize reasoning depth for difficult problems — with state-of-the-art results at GPQA Diamond 88.9%. It handles context windows up to 500K tokens and excels at 3D visualization, financial modeling, and sustained autonomous coding sessions.
**Opus 4.6** (released February 4, 2026) takes the series further with "Adaptive Thinking" toggles across four effort levels (Low, Medium, High, Max), "Context Compaction" for near-infinite agent sessions, and native parallel sub-agent orchestration. Its GPQA Diamond score reaches 91.3%, and it introduces the ability to spin up independent sub-tasks in parallel — making it uniquely suited for cybersecurity investigations, complex codebase refactors, and high-stakes research where edge-case analysis is mandatory.
## Key Capabilities
Outperforms human candidates on elite engineering exams. Handles codebase migrations, complex refactoring, and 30-minute autonomous coding sessions.
Opus 4.5 and 4.6 both offer configurable reasoning depth — minimize cost for routine tasks, maximize capability for demanding problems.
Opus 4.6 natively spins up independent sub-tasks to run tools in parallel — essential for cybersecurity and large-scale coding investigations.
Opus 4.6 automatically summarizes older context to sustain "infinite-feeling" agent sessions without losing the original goal-state.
## Best For
Choose **Opus 4.5** for deep research, large-scale document analysis, enterprise automation, and production-grade coding agents — particularly where you need a balance of maximum capability and cost efficiency. Choose **Opus 4.6** for the most complex tasks in your stack: multi-day codebase refactors, high-stakes legal or financial research, and cybersecurity defense workflows where parallel sub-agent execution makes a meaningful difference. For tasks that don't require Opus-level intelligence, Sonnet 4.5 or 4.6 offers strong performance at lower cost.
For long-running agentic sessions, Opus 4.6's Context Compaction feature prevents context-window exhaustion automatically. Enable it explicitly in your agent loop rather than relying on manual context management.
## Use Cases
* **Deep research and analysis** — Digest massive datasets, financial documents, and technical reports using up to 500K (4.5) or 1M (4.6) context.
* **Production-grade coding agents** — Build autonomous agents that create, test, and iterate on entire codebases with up to 75% fewer build/lint errors.
* **Enterprise automation** — Automate complex Excel workflows, financial modeling, and multi-agent systems that refine their own capabilities.
* **Cybersecurity defense** — Perform end-to-end vulnerability investigations with 90%+ success rates in blind tests (Opus 4.6).
* **High-stakes research** — Legal, financial, and scientific discovery where edge-case analysis and multi-step verification are mandatory.
## Specifications
| Specification | Opus 4.5 | Opus 4.6 |
| -------------- | ------------------------------------------------- | --------------------------------------------- |
| Provider | Anthropic | Anthropic |
| Context Window | 200K–500K tokens | 1.0M tokens |
| Reasoning | Adaptive (Standard/High) | Adaptive (Low/Medium/High/Max) |
| GPQA Diamond | 88.9% | 91.3% |
| Latency (TTFT) | 0.10s | — |
| Throughput | 22 tokens/sec | 22 tokens/sec |
| Key use cases | Expert logic, legal analysis, scientific research | PhD-level research, code review, legal audits |
[Try Opus in DeepMask →](https://chat.deepmask.io/)
# AI Models
Source: https://documentation.deepmask.io/models/overview
Browse the full catalog of 25+ AI models in DeepMask, organized by provider — from OpenAI and Anthropic to EU-hosted options via StackIT and Infercom.
DeepMask gives you access to more than 25 AI models from 11 leading providers in a single workspace. You can switch between models at any time without losing your conversation context, and several models run entirely on EU-hosted infrastructure so your data never leaves European soil.
**GPT-5.2 · GPT-5.3 · GPT-5.4** — Most capable models for chat, document and image analysis, and tool use with strong reasoning.
**GPT-4o** — Low-latency multimodal model optimized for real-time voice and vision tasks.
**GPT-4.1** — Well-suited for long-context tasks, spreadsheet analysis, and tool use.
**GPT-o3 Mini** — Focused reasoning model for document analysis and tool use.
**GPT-OSS 120B (StackIT)** · **GPT-OSS 120B (Infercom)** — Open-weight model for document analysis and research, EU-hosted.
**Opus 4.5 · Opus 4.6** — Anthropic's most capable tier; best for demanding chat, complex documents, image analysis, and tool use.
**Sonnet 4.5 · Sonnet 4.6** — Balanced performance for autonomous coding, agentic workflows, and long-horizon tasks. Context window up to 1M tokens.
**Haiku 4.5** — Fastest Anthropic model; designed for high-volume support, real-time data, and sub-agent workloads.
**Gemini 2.5 Pro** — High-capability model for chat, document and image analysis, and tool use with a large context window.
**Gemini 2.5 Flash** — Industry-leading throughput at 185 tokens/sec with a 1M token context window. Optimized for large-scale document processing and real-time summarization.
**Gemma 3 27B (StackIT)** — Lightweight open model for chat and document/image analysis, EU-hosted via StackIT (Schwarz Group).
**DeepSeek V3** — 671B MoE model delivering frontier-level coding and math performance. Strong for complex questions, writing, and document analysis. Context window up to 164K tokens.
**DeepSeek V3.1 (Infercom)** — Same capability as V3 with EU-hosted endpoints via Infercom for strict data residency requirements.
**Kimi K2 (DeepMask)** — 1T parameter MoE model with native Agent Swarm Mode, supporting up to 100 parallel sub-agents. Handles 2M token context. EU-hosted via DeepMask infrastructure.
**Kimi K2.5** — Open-source multimodal model that converts text, images, and video into production-ready code, built for large-scale agent swarm workflows.
**Mistral Large 3** — Elite reasoning, multimodal understanding, and best-in-class multilingual performance across 40+ languages.
**Mistral Medium 3** — Frontier-level performance at significantly lower cost; designed for fast, scalable enterprise AI deployments across cloud, hybrid, and on-premises environments.
**Qwen (DeepMask)** — Versatile model with reasoning and tool use, strong at document and image analysis and multilingual chat. EU-hosted via DeepMask infrastructure.
**Qwen3 (StackIT)** — Same capability profile as Qwen (DeepMask) with EU-hosting via StackIT (German sovereign cloud by Schwarz Group).
**MiniMax M2** · **MiniMax M2.1** — Built for elite multi-language coding, advanced agent workflows, and high-quality reasoning across development and office tasks.
**MiniMax M2.5 (Infercom)** — Strong for document analysis, coding, and tool use with a 164K token context. EU-hosted via Infercom.
**GLM-4.7** — Advanced reasoning and coding model featuring interleaved thinking, elite agent workflows, and high-fidelity UI generation for complex real-world tasks.
**GLM-4.7 Flash** — High-performance lightweight MoE variant delivering strong reasoning and coding accuracy with exceptional speed.
**Grok 3 Mini** — Compact reasoning model for fast, cost-effective chat and analysis tasks.
**Grok 4 Fast Non-Reasoning** — High-speed model optimized for rapid response without extended thinking overhead.
Models marked with **(StackIT)** run on StackIT, the German sovereign cloud operated by the Schwarz Group. Models marked with **(Infercom)** use EU-hosted endpoints via Infercom with strict data residency controls. Models marked with **(DeepMask)** are hosted directly on DeepMask's own EU infrastructure. If your organization requires that all data processing stays within the European Union, filter for EU-hosted models in the model selector.
# Qwen
Source: https://documentation.deepmask.io/models/qwen
Qwen (DeepMask) and Qwen3 (StackIT) on DeepMask. Alibaba's flagship models with dual-mode reasoning, 1M token context, and EU-hosted infrastructure via StackIT.
Alibaba's Qwen3 models bring frontier-class reasoning and repository-scale coding to DeepMask in two deployment configurations. Qwen (DeepMask) is the 235B flagship with dual-mode inference, spatial-visual logic, and a 1M-token context window. Qwen3 (StackIT) is a StackIT-tuned variant co-developed for European cloud environments, with native infrastructure awareness and the same powerful reasoning core. Both models excel at multilingual tasks, document analysis, and agentic workflows.
## About
Qwen (DeepMask) is Alibaba's 235B Qwen3 flagship model. It features Dual-Mode Inference, allowing you to toggle between "Instant" mode for fast chat and "Thinking" mode for deep, PhD-level problem solving. It leads on repository-scale coding — able to reason across tens of thousands of lines of code without context drift — and supports a 1M-token context window with efficient hardware use via a tiered KV cache.
Qwen (DeepMask) is hosted on DeepMask infrastructure with a 1M-token context window. Your data remains within DeepMask's EU-compliant environment.
## Key Capabilities
Toggles between fast chat and deep Thinking mode for complex, multi-step problem solving.
Excels at understanding complex diagrams, maps, technical blueprints, and spatial relationships.
Handles up to 1 million tokens, enabling analysis of very large codebases and document sets.
Understands the architectural intent behind a codebase, enabling whole-repo reasoning and refactoring.
## Use Cases
* **Enterprise software architecture** — Plan and refactor multi-repository backend systems with full structural awareness.
* **Global fintech analytics** — Process large volumes of financial data for predictive market analysis.
* **Creative design suite** — Leverage native support for high-fidelity image understanding and natural speech tasks.
* **Multilingual RAG** — Build retrieval-augmented generation pipelines across multiple languages with strong reasoning.
Qwen (DeepMask) is your best choice for repository-scale coding and complex reasoning tasks on DeepMask. Use Thinking mode for difficult problems and Instant mode for interactive chat at 0.22s TTFT.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------- |
| Model Provider | Alibaba |
| Main Use Cases | Agents Coding, Multilingual RAG |
| Reasoning Effort | High (Instant & Thinking) |
| GPQA Diamond | 89.3% |
| Max Context | 1M Tokens |
| Latency (TTFT) | 0.22s (Non-Thinking Mode) |
| Throughput | 145 Tokens/sec |
## About
Qwen3 (StackIT) is a specialized variant of Alibaba's Qwen3 series, co-developed with StackIT for European enterprise cloud environments. It features Hybrid Thinking Modes, allowing it to alternate between a high-compute "Deep Logic" mode and a lightweight "Fast Action" mode via API toggle. This model is specifically tuned for infrastructure-as-code, cloud-native application management, and complex technical project workflows.
Qwen3 (StackIT) is EU-hosted via StackIT, making it a strong choice for teams with European data residency requirements. It provides the same Qwen3 reasoning core with infrastructure-aware tuning and a 1M+ token context window.
## Key Capabilities
A single model that can think step-by-step or respond instantly — no separate model needed.
Strong at understanding cloud topologies and technical system configurations.
Maintains long-term project memory efficiently, suitable for large-scale document and code tasks.
Built-in vision to recognize and reason about architecture diagrams and visual inputs.
## Use Cases
* **Cloud infrastructure management** — Generate and debug complex multi-cloud deployment scripts with infrastructure-native understanding.
* **Repository-scale refactoring** — Analyze 10,000+ line codebases and propose structural architectural changes.
* **Technical project management** — Convert visual whiteboard sketches into technical PRDs and Jira tickets.
* **Agentic workflows** — Run complex reasoning-heavy agent pipelines with EU data residency guarantees.
Choose Qwen3 (StackIT) when you need EU-hosted Qwen3 reasoning for cloud infrastructure, DevOps, or enterprise workflows. It shares the same reasoning core as Qwen (DeepMask) with added StackIT infrastructure awareness.
## Specifications
| Specification | Value |
| ---------------- | ------------------------------------ |
| Model Provider | Alibaba |
| Hosting | EU-hosted via StackIT |
| Main Use Cases | Agentic Workflows, Complex Reasoning |
| Reasoning Effort | High |
| GPQA Diamond | 87.4% |
| Max Context | 1.01M Tokens |
| Latency (TTFT) | 0.35s |
| Throughput | 95 Tokens/sec |
# Sonnet (4.6, 4.5)
Source: https://documentation.deepmask.io/models/sonnet
Claude Sonnet 4.5 and 4.6 by Anthropic. 1M token context, GPQA up to 84.4%, native computer use, and 30-hour autonomous coding. Best for engineering agents.
The Sonnet family from Anthropic strikes the most practical balance between intelligence and throughput for professional software engineering and enterprise agentic workflows. Sonnet 4.5 pioneered long-horizon autonomous coding and native computer use; Sonnet 4.6 extends those capabilities with Opus-class performance at Sonnet-class cost, making it the current default for demanding production deployments.
## About Sonnet 4.5 and Sonnet 4.6
**Sonnet 4.5** is widely considered one of the most balanced models in the world for professional engineering. Built specifically to handle "long-horizon" tasks, it can work autonomously for 30+ hours on a single coding objective without losing coherence. It was the first model to achieve a 61.4% score on the OSWorld benchmark for real-world computer use, and its 1M token context window combined with a GPQA Diamond score of 83.4% makes it exceptional for research agents and full-stack engineering.
**Sonnet 4.6** (released February 17, 2026) is the current default model on Claude.ai. It delivers Opus-tier performance for "economically valuable office tasks" at Sonnet-tier cost — with significant advances in computer-use capabilities (OSWorld benchmark), instruction-following consistency, and up to 90% cost savings on high-volume tasks through prompt caching and batch processing. Its GPQA Diamond score reaches 84.4%.
## Key Capabilities
Sees screens, moves cursors, and types in standard desktop applications — enabling true browser-based and GUI automation without external tools.
Sonnet 4.5 can manage multi-day engineering sprints with self-correction and testing, maintaining coherence across the full session.
Sonnet 4.6 delivers significant improvements in consistency and nuance, making it a reliable choice for agent-in-the-loop systems.
Sonnet 4.6's prompt caching and batch processing discounts enable up to 90% cost savings versus standard API calls for high-volume workloads.
## Best For
Choose **Sonnet 4.5** for autonomous software engineering tasks, complex multi-app research workflows, and legal or financial forensics where massive document sets need sustained coherent analysis. Choose **Sonnet 4.6** for enterprise-grade agents, full-stack development lifecycle management, and browser-based automation where Opus-level accuracy is needed at Sonnet cost. For tasks with lower complexity requirements, Haiku 4.5 offers faster throughput. For the most demanding reasoning challenges, Opus 4.6 sets the ceiling.
For long-running Sonnet 4.5 agentic sessions, use the Context Editing API feature to let the model rewrite parts of its own memory — this keeps the active context efficient without losing goal-state over 30-hour sessions.
## Use Cases
* **Autonomous software engineering** — Building, testing, and deploying full-stack features from a single prompt.
* **Complex multi-app workflows** — Researching data in a browser and then populating a local Excel sheet and PowerPoint.
* **Legal and financial forensics** — Analyzing massive document sets for subtle logical contradictions.
* **Enterprise workflow automation** — Analyzing financial data, synthesizing internal insights, and generating professional content.
* **Browser-based agents** — Automating procurement, competitive analysis, and customer onboarding via digital interaction.
## Specifications
| Specification | Sonnet 4.5 | Sonnet 4.6 |
| -------------- | ------------------------------------------------------ | ----------------------------------------------------- |
| Provider | Anthropic | Anthropic |
| Context Window | 1.0M tokens | 1.0M tokens |
| Reasoning | Adaptive (Standard/High) | Adaptive (Standard/High) |
| GPQA Diamond | 83.4% | 84.4% |
| Latency (TTFT) | 0.42s | 0.42s |
| Throughput | 38 tokens/sec | 38 tokens/sec |
| Key use cases | Full-stack engineering, research agents, system design | Enterprise agents, UI automation, professional coding |
[Try Sonnet in DeepMask →](https://chat.deepmask.io/)
# Enterprise Integration Setup
Source: https://documentation.deepmask.io/msft-enterprise
Connect your Microsoft 365 tenant to DeepMask in 15 minutes by registering an Azure AD app and pasting two IDs.
## Overview
This guide walks you through registering an Azure AD application in your Microsoft 365 tenant. Once complete, you will have a **Client ID** and **Tenant ID** to paste into DeepMask's Microsoft (Enterprise) connector — enabling your users to interact with SharePoint and Outlook date inside DeepMask.
| What You'll Get | Details |
| -------------------------- | -------------------------------------------------------------------------------------------------------------------------------- |
| **Azure App Registration** | A single app entry in your Azure AD tenant that DeepMask uses to request access on behalf of your users. |
| **Delegated Permissions** | Read-only Microsoft Graph permissions. Each user can only see content they already have access to — no elevated service account. |
| **Client ID** | A unique identifier for your registered app. Paste this into DeepMask. |
| **Tenant ID** | Your organization's Azure AD directory identifier. Paste this into DeepMask. |
DeepMask uses OAuth 2.0 delegated authentication. When a user connects their account, they sign in with their own Microsoft credentials. DeepMask never stores passwords or receives broader access than the user already has in your tenant.
***
## Prerequisites
Before you begin, confirm you have the following:
### Required Access
* **Microsoft 365 account (Business Basic or higher)** — Your organization must have an active Microsoft 365 or Office 365 subscription with SharePoint Online and Outlook enabled.
* **Azure AD Global Administrator or Application Administrator role** — You need sufficient privileges to register applications and grant admin consent. If you are unsure of your role, ask your IT administrator.
* **Access to Azure Portal** — Navigate to [portal.azure.com](https://portal.azure.com) and confirm you can sign in with your admin account.
### What You Do Not Need
* No developer tools, code, or command-line experience required
* No changes to existing SharePoint sites or permissions
* No service account or shared credentials
The **Grant admin consent** step requires Global Administrator or Application Administrator privileges. If you do not have this role, you will need to coordinate with your IT/Azure administrator before proceeding.
***
## Step 1 — Register an Application in Azure AD
You will create a new App Registration in your Azure Active Directory. This is the entry point that DeepMask uses to identify itself with Microsoft.
### Navigate to App Registrations
Open a browser and go to [portal.azure.com](https://portal.azure.com). Sign in with your Global Administrator or Application Administrator account.
In the top search bar, type **Azure Active Directory** and click on it.
In the left-hand sidebar, click **App registrations**.
Click **+ New registration** at the top of the page.
### Fill in Registration Details
Enter the following values in the registration form:
| Field | Value |
| --------------------------- | -------------------------------------------------------------- |
| **Name** | `DeepMask SP Connector` (or any name you prefer) |
| **Supported account types** | Accounts in this organizational directory only (Single tenant) |
| **Redirect URI — Platform** | Mobile and desktop applications |
| **Redirect URI — URL** | `https://chat.deepmask.io/api/user/connectors/oauth/callback` |
Click **Register**. Azure will create the app and take you to its Overview page.
You should now see the app's Overview page showing an **Application (client) ID** and **Directory (tenant) ID**. Do not close this tab — you will return here in Step 4.
***
## Step 2 — Configure Authentication Settings
With the app registered, you need to enable the correct token and client flow settings so that DeepMask can complete authentication on behalf of your users.
From your app's Overview page, click **Authentication** in the left sidebar.
At the top of the Authentication page, click the **Settings** tab.
Under **Implicit grant and hybrid flows**, check both boxes:
* ☑ **Access tokens** (used for implicit flows)
* ☑ **ID tokens** (used for implicit and hybrid flows)
Scroll down to **Allow public client flows** and toggle it to **Enabled**.
Click **Save** at the top of the page to apply all changes.
These settings allow DeepMask to receive tokens directly after the user signs in and to operate without requiring a client secret on the user's device — which is the correct behaviour for a delegated, user-facing integration.
***
## Step 3 — Configure API Permissions
DeepMask requires a small set of delegated Microsoft Graph permissions to read SharePoint and Outlook content on behalf of your users. No write permissions are requested.
### Add Permissions
From your app's Overview page, click **API permissions** in the left sidebar.
Click **+ Add a permission**, select **Microsoft Graph**, then choose **Delegated permissions**.
| Permission | Type | Purpose |
| ------------------ | --------- | ---------------------------------------------------------- |
| `User.Read` | Delegated | Read the signed-in user's profile. |
| `Sites.Read.All` | Delegated | Read all SharePoint sites the user has access to. |
| `Files.Read.All` | Delegated | Read files in SharePoint document libraries. |
| `offline_access` | Delegated | Maintain access via refresh token (keeps users signed in). |
| `Sites.Selected` | Delegated | **Optional** — restrict access to specific sites only. |
| `Calendars.Read` | Delegated | Read user calendars |
| `ChatMessage.Read` | Delegated | Read user chat messages |
| `Contacts.Read` | Delegated | Read user contacts |
| `Mail.Read` | Delegated | Read user mail |
| `Mail.Read.Shared` | Delegated | Read user and shared mail |
| `Mail.ReadBasic` | Delegated | Read user basic mail |
| `MailboxItem.Read` | Delegated | Read a user's mailbox items |
Click **Add permissions** after selecting all required scopes.
### Grant Admin Consent
After adding all permissions, you must grant admin consent so users are not prompted for individual approval on first sign-in.
On the API permissions page, click **Grant admin consent for \[Your Organization]**.
A confirmation dialog will appear. Click **Yes**.
All permissions will show a green checkmark in the **Status** column.
Without this step, every user who signs in will see a **"Need admin approval"** prompt and will be unable to complete authentication. This step must be performed by a Global Administrator or Application Administrator.
***
## Step 4 — Find Your Client ID and Tenant ID
Both identifiers are visible on the app's Overview page in Azure Portal.
In the Azure Portal, go to **Azure Active Directory** → **App registrations** and click the name of the app you just registered (e.g. DeepMask SP Connector).
You will land on the **Overview** tab. You will see two important values:
| Azure Portal Label | What to Copy → Where to Paste in DeepMask |
| --------------------------- | ----------------------------------------- |
| **Application (client) ID** | → **Client ID** field in DeepMask |
| **Directory (tenant) ID** | → **Tenant ID** field in DeepMask |
Each value looks like a UUID, for example: `f47ac10b-58cc-4372-a567-0e02b2c3d479`
These identifiers are not secrets — they do not grant access on their own. However, treat them like internal configuration values and avoid sharing them publicly.
***
## Step 5 — Enter Credentials in DeepMask
With your Client ID and Tenant ID copied, you can now connect Microsoft (Enterprise) inside DeepMask.
Open your browser and go to [chat.deepmask.io](https://chat.deepmask.io). Sign in to your DeepMask account.
In the left navigation, click **Connectors**.
Locate the **Microsoft (Enterprise)** tile and click **Connect**.
* Paste your **Client ID** into the Client ID field.
* Paste your **Tenant ID** into the Tenant ID field.
Click **Save & Connect**.
A Microsoft sign-in window will open. Sign in with any user in your tenant and grant the requested permissions when prompted (this only appears once per user).
You will be redirected back to DeepMask. The connector tile will show a green **Connected** status.
DeepMask is now connected to your M365 tenant. Users can search files, sites, and list items from within DeepMask — limited to content they already have access to.
***
## Troubleshooting
If you run into issues during setup or after connecting, refer to the scenarios below.
### "Need Admin Approval" screen appears during sign-in
**Cause:** Admin consent was not granted for the app, or the tenant's user consent policy is set to block all user-initiated consent.
**Resolution:**
1. In Azure Portal, go to **Azure Active Directory** → **App registrations**.
2. Open your app and click **API permissions** in the sidebar.
3. Click **Grant admin consent for \[Your Organization]** and confirm.
4. Ask the affected user to try signing in again.
***
### "Insufficient privileges" or 403 error when browsing SharePoint
**Cause:** The `Sites.Read.All` or `Files.Read.All` permission is missing, or admin consent was not completed.
**Resolution:**
* Verify all four required permissions are listed on the API permissions page.
* Confirm the **Status** column shows a green checkmark (admin consent granted) for each.
* If a permission is missing, click **+ Add a permission**, add it, then re-grant admin consent.
***
### Wrong tenant — users from another organization are being prompted
**Cause:** The Tenant ID entered in DeepMask does not match your organization's Azure AD directory.
**Resolution:**
* In Azure Portal, go to **Azure Active Directory** → **Overview** and copy the Tenant ID shown there.
* Compare it to what is configured in DeepMask.
* Also confirm that **Supported account types** is set to **Single tenant** (not multi-tenant).
***
### Redirect URI mismatch error after sign-in
**Cause:** The Redirect URI in your app registration does not exactly match the one DeepMask sends.
**Resolution:**
1. In Azure Portal, open your app and click **Authentication** in the sidebar.
2. Under **Mobile and desktop applications → Redirect URIs**, confirm the entry is exactly: `https://chat.deepmask.io/api/user/connectors/oauth/callback`
3. No trailing slash. No `http://` variant. Save and retry.
***
### User can sign in but sees no SharePoint sites
**Cause:** The signed-in user does not have any SharePoint site memberships in the tenant, or sites are restricted.
**Resolution:**
* Confirm the user has been added as a member to at least one SharePoint site.
* In **SharePoint Admin Center**, verify the site is not restricted or archived.
* If `Sites.Selected` was added, ensure the specific sites have been explicitly granted to the app via the SharePoint Admin API.
***
## Security & Privacy
DeepMask is designed so that you retain full control of your data and your users' access. This section explains the key security properties of the SharePoint integration.
### Delegated Authentication
DeepMask uses OAuth 2.0 delegated permissions, not application-level (app-only) permissions. This means:
* Every action performed by DeepMask is done on behalf of the signed-in user.
* A user can only read SharePoint content they already have permission to access.
* DeepMask cannot bypass SharePoint's existing role-based access controls.
* Removing a user's SharePoint access in Microsoft 365 immediately removes their access in DeepMask.
### No Stored Credentials
DeepMask does not store your Microsoft password, your Client Secret, or raw SharePoint content. Authentication is handled entirely through short-lived OAuth access tokens and refresh tokens, which are encrypted at rest.
### No Service Account
Unlike some integrations that use a single shared service account to access all data, DeepMask authenticates each user individually. This ensures audit logs in your Microsoft 365 tenant accurately reflect which user accessed which content.
### Read-Only Permissions
The permissions configured in this guide are strictly read-only. DeepMask requests no write, delete, or administrative permissions. It cannot create, modify, or delete files, lists, or sites.
### Revoking Access
To disconnect DeepMask from your tenant at any time:
* **Option A:** In DeepMask → **Integrations** → **SharePoint**, click **Disconnect**.
* **Option B:** In Azure Portal → **Enterprise Applications**, find `DeepMask SP Connector` and delete it.
Either action immediately revokes all access tokens. No data is retained after disconnection.
Questions about data residency, compliance, or security? Contact DeepMask support at [support@deepmask.io](mailto:support@deepmask.io).
# Get started
Source: https://documentation.deepmask.io/quickstart
Sign up, create your workspace, select an AI model, and send your first message in minutes. This guide walks you through every step to get productive with DeepMask immediately.
DeepMask lives at [chat.deepmask.io](https://chat.deepmask.io) — a unified AI workspace where you can access 25+ leading models, run perplexity-powered web searches, visualize data, and organize work into persistent Projects. This quickstart guide takes you from a blank browser tab to your first productive AI conversation, and introduces Projects so you can see how DeepMask scales beyond one-off chats.
Go to [deepmask.io](https://deepmask.io) and sign up for an account. You can register as an individual or as part of an enterprise organization.
If your organization already has a DeepMask enterprise account, ask your administrator to invite you directly. You will receive an email with a link to join your team's workspace.
Once registered, you will be taken to the main chat interface at [chat.deepmask.io](https://chat.deepmask.io).
At the top of the chat interface, open the model selector and choose the AI model you want to use. DeepMask offers 25+ models from providers including OpenAI (GPT-5.4, GPT-5.3, GPT-5.2, GPT-4.1, GPT-4o), Anthropic (Claude Opus 4.6/4.5, Sonnet 4.6/4.5, Haiku 4.5), Google (Gemini 2.0 Flash, Gemini 2.5 Flash), Mistral, DeepSeek, Grok, Kimi K2, Qwen, Gemma, MiniMax and more.
Not sure which model to pick? See the [model guide](/models/choosing-a-model) for a breakdown of each model's strengths — from long-document reasoning to fast coding assistance. EU-hosted variants of select models (marked with "StackIT" or "DeepMask") keep your data within European infrastructure.
Type your first message in the chat input and press Enter. DeepMask sends your prompt to the selected model and streams the response back in real time.
You can enhance any conversation with additional capabilities available from the toolbar below the input:
* **Web Search** — enable Perplexity-powered search to ground responses in real-time, cited web results
* **Extended thinking** — activate deeper reasoning for complex, multi-step problems
* **Response style** — switch between Normal, Concise, Explanatory, Learning, or Formal output styles
* **MCP connectors** — attach connected tools like SharePoint or Outlook to give the model access to your data
A typical first conversation might look like this: you upload a spreadsheet, ask the model to analyze trends and generate a chart, then follow up by asking for a summary in a formal tone — all without leaving the chat window.
If you want a different perspective or need a model better suited to a follow-up task, change the model at any time during a conversation. DeepMask preserves the conversation context so the new model has full awareness of what was discussed.
For example, you might start with Gemini 2.5 Flash for a quick answer, then switch to Claude Opus for a more in-depth analysis of the same topic.
For ongoing work — a research initiative, a product launch campaign, or an engineering sprint — create a Project to give your conversations a persistent home.
To create a Project:
1. Click **Projects** in the left sidebar.
2. Select **New Project**.
3. Give the project a name and write a set of custom instructions to prime every conversation in that project (for example, "You are analyzing our Q3 operations data. Always respond in a formal tone.").
4. Upload relevant files — documents, spreadsheets, briefs — so every conversation in the project has access to them.
5. Invite teammates to collaborate in the same project workspace.
Projects persist across sessions. Every conversation you start inside a project inherits the project's instructions, files, and shared context automatically.
## What to do next
Learn which models excel at coding, writing, research, and reasoning — and when to switch between them.
Choose a plan, configure team settings, and review security options for your organization.
Use MCP connectors to give DeepMask access to Microsoft Enterprise, OneDrive, Outlook, SharePoint and more.
Upload CSVs or spreadsheets and generate charts and insights without writing any code.
# Contact DeepMask Support and Sales
Source: https://documentation.deepmask.io/support/contact
Reach the DeepMask team by email or book a call for support, sales questions, enterprise plans, and partnership inquiries. Response within 48 hours.
The DeepMask team is based in Munich, Germany and is available to help with product questions, enterprise deployments, billing, and general support. Choose the option below that best fits your need and the team will get back to you within 48 hours.
Send a message to [contact@deepmask.io](mailto:contact@deepmask.io) for support questions, billing inquiries, partnership requests, or anything else. The team responds within 48 hours.
Schedule a free 30-minute call to discuss your team's needs, get a product walkthrough, or explore enterprise options. Pick a time that works for you at [cal.com/hissan-usmani/30min](https://cal.com/hissan-usmani/30min).
For large team deployments or enterprise contracts, booking a call is the fastest way to get a tailored proposal. The team can discuss custom pricing, compliance documentation (including DPAs), and dedicated onboarding support.
## What to include in your message
To help the team respond quickly, include the following in your email:
* Your name and company
* Team size or number of intended users
* A brief description of your use case or question
* Any compliance or data residency requirements
## Security and privacy
DeepMask is operated by DeepMask GmbH, registered in Munich, Germany. All data is stored and processed within the EU. DeepMask is fully GDPR compliant and never uses your data to train AI models.
For full details on how your data is handled, read the [Privacy Policy](https://www.deepmask.io/privacy).
If you have a security concern or vulnerability to report, email [contact@deepmask.io](mailto:contact@deepmask.io) with the subject line "Security Report" and the team will prioritise your message.
# FAQ
Source: https://documentation.deepmask.io/support/faq
Answers to common questions about DeepMask's AI models, data privacy, GDPR compliance, Projects, MCP connectors, response styles, and enterprise pricing.
DeepMask is a sovereign European AI workspace that unifies 25+ AI models in a single interface. If you have a question not answered here, reach out to the team at [contact@deepmask.io](mailto:contact@deepmask.io) or visit the [contact page](/support/contact).
DeepMask gives you access to 25+ leading AI models from providers including OpenAI, Anthropic, Google, Mistral, xAI, Meta, DeepSeek, Moonshot, and Alibaba. The current lineup includes models such as GPT-4o, GPT-4.1, GPT-5.2, Claude Sonnet 4.5, Claude Opus 4.5, Claude Haiku 4.5, Gemini 2.0 Flash, Gemini 2.5 Flash, Grok 3 Mini, DeepSeek V3, Kimi K2, Qwen, and EU-hosted variants like Qwen3 (StackIT) and Gemma 3 27B (StackIT).
The model list is updated regularly. You can see the full current selection inside [chat.deepmask.io](https://chat.deepmask.io).
Yes. You can change the active AI model at any point during a conversation without losing your chat history. This lets you start a task with one model and switch to a better-suited model for follow-up steps — for example, beginning with a fast model for drafting and switching to a reasoning-focused model for analysis. No context switching between apps or separate logins required.
No. DeepMask never uses your conversations, files, or any other data you submit to train AI models — either its own systems or those of the underlying model providers. Your data is yours. This commitment is a core part of DeepMask's privacy policy and applies to all plans.
You can review the full data handling policy at [deepmask.io/privacy](https://www.deepmask.io/privacy).
All data is hosted within the European Union. DeepMask's primary infrastructure partner is the Schwarz Group's German sovereign cloud (StackIT), with additional EU-region deployments for redundancy, scalability, and strict data residency. Your data does not leave EU jurisdiction.
Yes. DeepMask is fully aligned with the EU General Data Protection Regulation (GDPR). Data is processed and stored exclusively within the EU, you retain ownership of your data, and DeepMask never trains AI models on user content.
ISO 27001 certification is currently in progress. Data in transit and at rest is protected with enterprise-grade encryption.
If your organisation requires a Data Processing Agreement (DPA) or has specific compliance requirements, contact the team at [contact@deepmask.io](mailto:contact@deepmask.io).
A Project is a persistent, context-rich workspace within DeepMask. You can set custom instructions that apply across all conversations in the project, upload reference documents, and organise related chats together. Projects retain context between sessions, so the AI always has the background it needs without you repeating yourself.
Projects also support team collaboration — you can share a project with teammates so everyone works from the same instructions and files.
MCP (Model Context Protocol) connectors let you link DeepMask to external tools and data sources your team already uses. Supported integrations include Microsoft Enterprise, OneDrive, Outlook, SharePoint and more.
Once a connector is active, the AI can read from and interact with those services directly within your conversations — pulling in relevant documents, data, or records without you having to copy and paste content manually. You manage your connectors from the connector panel inside the app.
DeepMask lets you choose from five response styles to match the tone and format of AI output to your task:
* **Normal** — balanced, general-purpose responses
* **Concise** — brief, direct answers with minimal elaboration
* **Explanatory** — detailed explanations with reasoning and context
* **Learning** — structured for education, with examples and step-by-step breakdowns
* **Formal** — professional tone suited to business writing and reports
You can switch styles at any time during a conversation. The model adapts its output immediately.
Yes. DeepMask includes enterprise team management features designed to scale across an organisation. Administrators can manage team members, oversee usage across the team, track token consumption by model, and organise work into shared Projects. This gives leadership visibility into AI adoption and cost without disrupting individual workflows.
Enterprise pricing is available for larger teams and organisations with specific deployment, compliance, or support requirements. To get a quote or discuss your needs, you can:
* Email the team at [contact@deepmask.io](mailto:contact@deepmask.io)
* Book a 30-minute call at [cal.com/hissan-usmani/30min](https://cal.com/hissan-usmani/30min)
The team typically responds within 42 hours.
Extended thinking is a mode available on select models (such as Claude Opus) where the model takes additional time to reason through a problem before responding. Instead of producing an immediate answer, the model works through multiple steps internally, which improves accuracy on complex, multi-step tasks like research synthesis, strategic analysis, or difficult coding problems.
You can enable extended thinking from the toolbar inside any conversation when a compatible model is selected.
Yes. You can upload documents, spreadsheets, CSVs, images, and other files directly into your conversations or Projects. Once uploaded, the AI can read, analyse, and reference the content — for example, summarising a report, generating charts from a CSV, or answering questions about an uploaded document.
Files attached to a Project remain available across all conversations in that project, so you do not need to re-upload them each session.