What's new

MCP Server Support

Assist now lets you create MCP servers to expose your workspace tools to external AI clients like Claude Desktop, ChatGPT, Cursor, Windsurf, and n8n. Clients connect securely using the OAuth 2.1 protocol, so each connection is authorized and scoped to a specific server.

• Create MCP servers — Choose a name and select which tools to include, then share the unique server URL with your AI client
• OAuth 2.1 authorization — Clients go through a secure authorization flow before accessing tools, with automatic token refresh
• Scoped access — Connected clients can only call tools included in the server, nothing else in your workspace
• Revocable connections — Revoke an MCP server at any time to immediately terminate all active connections
• Works with popular clients — Setup guides for Claude Desktop, Cursor, ChatGPT, Windsurf, and n8n

Streaming Timer

A new visual timer shows how long a response has been generating, so you always know your request is being processed.

• Live elapsed time -- See exactly how many seconds have passed since the response started streaming
• Animated glow -- A subtle gradient animation provides a visual cue that the assistant is actively working
• Peace of mind -- No more wondering if something is stuck -- the timer confirms progress at a glance

One-Time Scheduled Triggers

You can now schedule an agent to run once at a specific date and time -- not just on a recurring schedule.

• Pick a date and time -- Set exactly when you want the agent to run, down to the minute
• Automatic cleanup -- The trigger deactivates itself after running, so there is nothing to remember to turn off
• Great for reminders and deadlines -- Perfect for one-off tasks like sending a report on a specific date or checking in before a deadline

Workspace Admin Access

Workspace admins now have visibility into everything happening across the workspace.

• View all chats -- Admins can see every conversation in the workspace for oversight and support
• View all sandcastles -- Browse and inspect any sandcastle app created by workspace members

Notification Recaps

Notifications have been rewritten to be friendlier and easier to scan.

• First-person recaps -- Notifications now read like a quick update from your assistant instead of dry status reports
• Natural language -- Results are summarized conversationally so you can understand what happened at a glance

Sandcastle User Permissions

Sandcastle apps can now adapt their interface based on who is using them.

• Role-aware apps -- Apps can detect whether the current user has view, edit, or admin access and adjust what they see
• Show and hide features -- Restrict sensitive actions to admins while keeping the app usable for everyone else

Sandcastle Execute Permissions

Sandcastle apps now support a proper permission tier -- view, edit, and execute -- so you can share apps with the right level of access.

• Execute access -- Grant someone the ability to use a sandcastle app without being able to modify its code or configuration
• Clear permission levels -- View lets people see the app, execute lets them interact with it, and edit gives full control
• Share with confidence -- Give teammates or stakeholders access to your apps without worrying about accidental changes

Better Slack Bot Responses

Conversations with Assist in Slack now feel more natural and human.

• Conversational tone -- Responses read like a helpful teammate instead of robotic action summaries
• Friendlier updates -- Status messages and action confirmations are written in plain language

Dedicated AI Agents for Sandcastle Apps

Every sandcastle app now gets its own AI agent that understands the app's data and capabilities.

• Automatic provisioning -- When you create a sandcastle, a dedicated AI agent is created for it with knowledge of the app's collections and structure
• Scoped access -- Each agent only sees and modifies data belonging to its sandcastle, preventing cross-app interference
• Custom instructions -- Add your own instructions to shape how the agent interacts with the app's data
• Schema awareness -- Agents inspect collection schemas before making changes, catching errors before they happen

Smarter State Management for Sandcastle Apps

The tools that AI agents use to read and write sandcastle data have been significantly upgraded.

• Schema introspection -- Agents can inspect collection names, document counts, and field types before making changes
• Write safety -- Agents must read the current schema before writing, preventing accidental data corruption from stale assumptions
• Discoverable tools -- Sandcastle state tools now appear in tool search, so agents can find and use them without explicit configuration

Refreshed Login Page

The login page has a new animated background with a distinctive ASCII art style that responds to mouse movement.

Single Sign-On (SSO)

Workspaces can now require members to sign in through their company's identity provider.

• Any OIDC provider -- Works with Okta, Azure AD, Google Workspace, and any OIDC-compatible identity provider
• SCIM directory sync -- Automatically provision and deprovision user accounts when people join or leave your organization
• Workspace enforcement -- Require all members to authenticate through SSO so no one bypasses your identity policies

More AI Models

Assist now supports models from additional providers, giving you more flexibility in how your tools and agents run.

• New providers -- Google Gemini, Cerebras, Perplexity, and OpenAI models are now available alongside Anthropic
• Choose per workflow -- Select the best model for each agent or tool based on speed, cost, or capability

Sandcastle Persistent State

AI-built apps (sandcastles) can now store and retrieve data that persists between sessions.

• Read and write data -- Sandcastle apps can save records, update them, and query them with pagination
• Survives sessions -- Data persists across browser refreshes and between different users

Slack Long Message Handling

AI responses in Slack are no longer truncated when they exceed the message size limit. Long responses are automatically split into multiple messages so nothing is lost.

Security Improvements

• DNS rebinding protection for all outbound requests
• Workspace deactivation now enforced across Slack, MCP, and all API endpoints

Dynamic Tool Reliability Fix

Fixed an issue where custom tools created by users could fail to load with a "Version not found" error. Tools now load reliably regardless of when they were created.

Browser Automation

Your AI agents can now browse the web. Agents can visit pages, interact with elements, and extract information from websites as part of their workflows -- useful for research tasks, data collection, and monitoring.

• Web browsing for agents -- Agents can navigate to URLs, click elements, and read page content
• Background delegation -- Assign browsing tasks to a sub-agent that works in the background while you continue your conversation

Slack Channel Routing

You can now control which AI agent responds in each Slack channel directly from the subagent settings page.

• Channel mapping UI -- View, add, and remove Slack channel-to-agent mappings from the subagent edit page
• Per-channel routing -- Route different Slack channels to different specialized agents