The latest two releases (2026.08 & 2026.10) introduce significant improvements across the chat experience, knowledge management, agent collaboration, and AI model infrastructure.
The chat interface has been modernized with a cleaner layout, improved prompt and tool access, voice input powered by Whisper, and new safeguards such as confirmation for external prompts and a floating hallucination indicator.
On the analytics and automation side, admins gain new capabilities including NPS export and an optional code execution tool.
Knowledge management has also advanced with enhanced RAG features such as email file ingestion, automated report generation, file-level access control, and improved metadata extraction. Collaboration within Agentic Tables is strengthened through granular permissions, persistent agent context, and more controlled source processing.
Additionally, the platform now supports customizable end-user spaces, better file sharing across agents and sub-agents, model-agnostic token counting, and several new AI models. Together, these updates focus on improving usability, security, collaboration, and analytical power for complex enterprise workflows.
1. Better Chat Experience
What's New (Business/End-User Focus)
UserExternal Prompt Links Require Confirmation: Prompts opened via external links will be pre-filled but won’t auto-submit; users must explicitly confirm before sending, reducing unintended execution and aligning with modern security best practices.
Admin(experimental)Voice Input with Whisper Speech-to-Text:A new Whisper-based voice input mode is now available in the chat input. It features real-time audio waveform visualization during recording and a processing indicator while transcription is in progress.
Admin(experimental)Code Execution Tool: Space admins can now enable it by toggling the tool on in Space Admin settings.
The code execution tool will be automatically disabled if an admin selects an incompatible model. Code execution currently only supports models in the OpenAI responses API via Azure or LiteLLM. Click here to view the complete list.
Required Feature Flags
For code execution to be available after this release, the following three feature flags must be enabled: FEATURE_FLAG_USE_OPENAI_V1_13819, FEATURE_FLAG_ENABLE_CODE_EXECUTION_UN_17498,SOURCE_AVAILABLE_TOOLS_REMOTELY
The Code Execution tool can easily be toggled on and off
The code execution tool presents with an error when an invalid model is selected
AdminAdd NPS Rating to Analytics: Space Admins can now download the Net Promoter Score rating from the chat analytics page.
UI / UX Refresh of Chat Interface
User (Beta)Chat Welcome View:
We have introduced a warmer, more inviting entry: “How can I help you?” with a short description.
The large, always-on upload area is removed; upload is available via drag-and-drop into the chat box or via a less prominent control, so the main focus is on asking a question.
This new UI is not enabled by default (but can be enabled via the corresponding feature flag mentioned in the Infrastructure Chages) as it part of the UI refresh we are working on.
User (Beta)Buttons inside the chat box: Chat input actions (upload, voice, tools, send) are moved inside the chat input area for a cleaner, more consistent look. Upload and voice are de‑emphasized so they’re not confused with the primary send action, while still easy to find when needed.
User (Beta)Simpler prompt menu: My Prompts and Suggested Prompts are in a compact bar below the input with clear labels, so you can open and pick prompts faster.
User (Beta)Easier tool selection: Tool selection is available via a “+” menu, with grouped tools (MCP tools) in dropdown groups for easier separation of concerns.
User(experimental) Cleaner answer layout: User messages use a bubble style on the right; assistant messages retain a flat style on the left for a clear question-and-answer distinction. Assistant message avatars are no longer shown, making threads easier to read, especially on mobile.
User(experimental) Sources and content with the answer: Files and sources linked to an answer are shown as clickable badges directly under that answer’s text instead of as separate blocks between messages.
User(experimental)Unified message actions: Copy, feedback (thumbs up/down), references, and debug sit in a single action row for assistant messages. References and debug open in a side panel so you can keep reading the thread while viewing references or debug info. On user messages, actions appear on hover to keep the default view minimal.
User(experimental) Tables with copy: Tables in assistant messages use an updated design with a copy button in the lower-right of the table (visible on hover on desktop, always visible on mobile) so you can copy table data quickly.
User(experimental) Updated suggested follow-up questions: Follow-up questions use a clearer design with a list-style layout so suggested next questions are easier to scan and tap.
User(experimental)Floating hallucination indicator: The previous Answer Assessment button is removed. If no hallucination is detected, nothing is shown. When the model flags potential inaccuracies (moderate or severe hallucination), a floating button appears in the lower right (yellow for moderate, red for severe). Hover over the button to view the reason for the hallucination. It stays visible while you scroll so you can still see and act on the warning.
Requesting Approval in Chat
The new elicitation behavior (in-chat approval and query review) is experimental and off by default. “Query elicitation for Web Search” can be enabled for selected organizations or environments through advanced configurations in Space Admin. Full documentation here.
User(experimental) In-chat approval form: Tools and assistants can now ask for your approval or input in the middle of a conversation. When they do, the form appears in the same message thread with a clear “Requesting Approval” label. You can Confirm or Cancel, so you stay in control before the tool continues. Your response is sent back to the tool so it only proceeds with what you approved.
User(experimental) Query review for Web Search: When using Web Search with query elicitation enabled, you can review and edit the search queries before they run. A form in the chat shows the proposed queries; you can add, remove, or change them, then confirm or cancel. This gives you a chance to correct or refine searches before they execute.
User(experimental) Consistent form experience: The approval form supports different field types (standard MCP Elicitation schema) and validation, so behavior is consistent whether the request comes from an MCP tool or an internal feature like Web Search query review.
Improvements
AdminWhisper OX API key auth: The speech service now supports optional API key authentication for Whisper OX (self-hosted) endpoints. Configure WHISPER_OX_API_KEY and WHISPER_OX_API_KEY_HEADER to secure the connection. Backward compatible — no auth when unset.
Admin(experimental)Secure HTML Rendering via HTML Store: HTML content generated in chat responses (e.g. financial charts, rendered HTML blocks) is now served through a dedicated HTML Store service with TTL-based memory management. Large HTML content (>2MB) shows a download button instead of rendering inline. See Code Reflector Infrastructure for details.
👉 Why this matters?
The modernized interface reduces cognitive load by hiding secondary tools until they are needed, while the new one-click table copying and integrated source badges significantly speed up report generation. Security is bolstered by requiring confirmation for external links and providing real-time "hallucination" warnings, ensuring you can trust the AI's output. A
dditionally, the new Code Execution tool allows for complex mathematical modeling directly within the chat, and Whisper-powered voice input makes capturing meeting notes faster and more accurate than ever.
2. Better Knowledge Management (RAG)
What's New (Business/End-User Focus)
User(experimental)Email File Ingestion (.eml and .msg): You can now upload and ingest Outlook .msg and standard .eml email files. Email body text is extracted and cleaned (signatures and quoted replies are stripped), and attachments are automatically converted and included.
Admin(experimental)Skip CSV File Ingestion: You can now choose to skip the ingestion of CSV files by setting the new SKIP_EXCEL_INGESTION flag as the uniqueIngestionMode. This gives you greater flexibility to customize your ingestion process and focus only on the file types that matter most to your workflow.
Admin(experimental)Reporting Engine (.uqtemplate): A new reporting engine allows automated report generation from ingested content using customizable .uqtemplate template files. Configure reportTemplates in the ingestion settings to trigger report generation on ingestion. Generated .uqreport outputs are stored as new content items.
Admin(Beta)File Access:
File-based access control is now available for all clients in the Knowledge Base, enabling permissions to be configured at the individual file level rather than only at the folder level. This provides greater flexibility when managing sensitive or team-specific content, ensuring that access can be tailored precisely to business needs without restructuring folders.
General overview of the file-based access control feature
User(experimental)Multi-File Select and Move In Knowledge Base Column View: Multi-file selection capabilities are now available in Column View, making it faster and more intuitive to organize content within the Knowledge Base.
Files can be selected using familiar desktop interaction patterns:
This update reduces repetitive actions and enables faster restructuring of content within large Knowledge Bases.
UserAgentic Metadata Extractor:
More reliable metadata extraction: Extraction jobs now runs more reliably. If a job fails, it automatically retries the metadata extraction from the file, so you'll see fewer files stuck mid-processing.
Re-extract metadata on existing files: You can now trigger metadata extraction on files that are already in a folder, directly from the folder's configuration settings. Two new checkboxes give you control over the scope:
"Extract metadata on existing files in this folder" — re-runs extraction on all files currently in the folder when you save.
"Also extract on files in subfolders" — extends the re-run to nested subfolders as well.
It's useful when you've updated your metadata schema or configuration and want to apply it to files that were ingested before the change.
👉 Why this matters?
These updates allow you to bypass manual "detective work" by ingesting clean email correspondence directly into research workflows and using automated reporting to summarize new data. Moreover, the ability to manage permissions at a file level and re-extract metadata on existing collections ensures that your knowledge base remains secure, accurate, and easily searchable as your business needs evolve.
3. Improvements to Agentic Tables
What's New (Business/End-User Focus)
UserSheet sharing and access permissions: Agentic Table sheets can now be shared with members of the space using two permission levels: Manage and Edit. Manage allows changing the sheet name and managing who has access; Edit allows working with the sheet (e.g. adding sources, running the agent, editing content) but not changing the name or managing permissions. This supports collaboration within the space while keeping control over access and naming.
UserRemove sources from the sheet: Agentic Table sheets now support removing files from the source list. A removed file is excluded from future agent context, while existing references and links in the sheet stay intact. This allows the source set to be corrected or narrowed without affecting prior answers or citations.
Improvements
User Groups in assignee column: The assignee column now shows both individual users and groups that have access; when an owner shares with a group, all members and the group itself appear in the assignee list.
UserProcessing after adding sources to an agentic table: When sources are added to an agentic table, processing is no longer started automatically, which avoids processing large sheets unintentionally and gives users more control over when and what to reprocess. Re-run is available to process all rows or a single row after sources are added.
AdminUse RfpAgent instead of Due Diligence when creating Agentic Table in Space UI: Creating a new Agentic Table from the Space UI now uses the RfpAgent module instead of the deprecated Due Diligence module, so users always work with the current Agentic Table implementation.
UserUI Polish for Agentic Table: Consolidated UI refinements including a refreshed rerun button, improved dropdowns on the sheet overview, a new loading preview for the agentic table, and updated inline reference badges for consistent styling and clearer readability.
UserCollapsible Content Support in Agentic Table Cells: Cells now support HTML details/summary for accessible, styled expand/collapse with proper overflow handling.
UserRfPAgent Subagent: Persistent Chat Context: Subagents now reuse a persistent chat_id via ShortTermMemory, enabling them to understand and leverage broader conversation and table context – not just single lines – so follow-up questions are answered more reliably.
👉 Why it matters
These updates transform Agentic Tables into a robust collaboration hub by introducing multi-user permissions (Manage vs. Edit) and deeper "contextual memory," ensuring your team works from a single source of truth. By decoupling source ingestion from automatic processing, users gain critical control over computational costs and time management for large datasets. Furthermore, moving to the RfpAgent module and introducing UI enhancements like collapsible cells and persistent chat context ensures that complex analyses remain readable, reliable, and easy to maintain over time.
4. Unique AI Spaces
What's New (Business/End-User Focus)
User(experimental)Customize end-user space- A user can create, edit and delete multiple end user spaces. End-user spaces are customized spaces for a particular assistants (The instructions from the end user doesn’t alienate the assistant instructions, but appends to the assistant instruction).
Admin(experimental)Enable End-User Spaces: An Admin can now enable end user spaces on an assistant, which allows a user to create customized spaces. See here for enabling the end user space on space 2.0, and here for enabling on Custom spaces.
User Sub-Agents Now Inherit Uploaded Files from Orchestrator: Sub-agents now automatically have access to files uploaded to the main agent (or orchestrator). Any documents provided at the orchestrator level are shared with downstream sub-agents, ensuring consistent context and seamless task execution across the agent hierarchy.
Improvements
UserEnable Sub-Agents to Access Files from Main Agent Chat: Sub-agents can now access files uploaded in the main agent chat. This improves usability, and ensures consistent file availability across parent and sub-agent interactions.
UserSupport Rendering of Base64-Encoded Images from MCP Tool Responses: Unique AI now supports rendering base64-encoded images returned by MCP Server tool calls, fully aligned with the MCP Image Content specification. The base64 data is automatically converted (e.g., to PNG format) and can be directly accessed and displayed within agent responses.
👉 Why it matters
By allowing end-users to create and customize their own spaces, teams can tailor assistant behaviors to specific niche tasks without overwriting core organizational instructions. This update also eliminates manual file-sharing friction; sub-agents now "inherit" context from the main agent, ensuring complex, multi-step tasks are executed with full visibility into all uploaded documents. Finally, the support for base64 image rendering from MCP tools enables more sophisticated visual outputs, making the platform a more versatile tool for data visualization and technical analysis.
5. Next-Generation AI Models
What’s New (Business/End-User Focus)
AdminModel-Agnostic Token Counting: Token counting has been refactored to be model-agnostic, removing the tight dependency on the OpenAI tokenizer and enabling accurate accounting across supported LLMs. The tokenizer can now be configured directly within the Language Model settings. Currently supported tokenizers include OpenAI encoders as well as Qwen and DeepSeek, with additional tokenizers available on request. When switching models, ensure the tokenizer is correctly adjusted in the Language Model Info configuration to maintain accurate token tracking.
New Model Deployments:
Anthropic
AdminClaude Opus 4.6 is available through LiteLLM using the following enum:
litellm:anthropic-claude-opus-4-6
xAI
AdminGrok 4.1 models are available through LiteLLM using the following enum:
litellm:grok-4-1-fast-non-reasoning
litellm:grok-4-1-fast-reasoning
Gemini
AdminGemini 3.5 Pro Preview model is available through LiteLLM using the following enum:
litellm:gemini-3-1-pro-preview
⚠️ Upcoming Module Retirement
The following AI modules will be retired in 2026.12:
DueDiligence
ImageSummarizer
InvestmentResearchDocuments
InvestmentResearchTable
LunchSearchV4
QueryTable
QuestionPredictor
TranscriptInteraction
WhatsappWriter
ContextMemorySearch
If you are currently using any of these modules, please reach out to your Customer Success contact to plan the replacement.
Unique AI Release Notes: February 2026
The latest two releases (2026.08 & 2026.10) introduce significant improvements across the chat experience, knowledge management, agent collaboration, and AI model infrastructure.
The chat interface has been modernized with a cleaner layout, improved prompt and tool access, voice input powered by Whisper, and new safeguards such as confirmation for external prompts and a floating hallucination indicator.
On the analytics and automation side, admins gain new capabilities including NPS export and an optional code execution tool.
Knowledge management has also advanced with enhanced RAG features such as email file ingestion, automated report generation, file-level access control, and improved metadata extraction. Collaboration within Agentic Tables is strengthened through granular permissions, persistent agent context, and more controlled source processing.
Additionally, the platform now supports customizable end-user spaces, better file sharing across agents and sub-agents, model-agnostic token counting, and several new AI models. Together, these updates focus on improving usability, security, collaboration, and analytical power for complex enterprise workflows.
1. Better Chat Experience
What's New (Business/End-User Focus)
User External Prompt Links Require Confirmation: Prompts opened via external links will be pre-filled but won’t auto-submit; users must explicitly confirm before sending, reducing unintended execution and aligning with modern security best practices.
Admin (experimental) Voice Input with Whisper Speech-to-Text: A new Whisper-based voice input mode is now available in the chat input. It features real-time audio waveform visualization during recording and a processing indicator while transcription is in progress.
Admin (experimental) Code Execution Tool: Space admins can now enable it by toggling the tool on in Space Admin settings.
The code execution tool will be automatically disabled if an admin selects an incompatible model. Code execution currently only supports models in the OpenAI responses API via Azure or LiteLLM. Click here to view the complete list.
Required Feature Flags
For code execution to be available after this release, the following three feature flags must be enabled: FEATURE_FLAG_USE_OPENAI_V1_13819, FEATURE_FLAG_ENABLE_CODE_EXECUTION_UN_17498, SOURCE_AVAILABLE_TOOLS_REMOTELY
The Code Execution tool can easily be toggled on and off
The code execution tool presents with an error when an invalid model is selected
Admin Add NPS Rating to Analytics: Space Admins can now download the Net Promoter Score rating from the chat analytics page.
UI / UX Refresh of Chat Interface
User (Beta) Chat Welcome View:
User (Beta) Buttons inside the chat box: Chat input actions (upload, voice, tools, send) are moved inside the chat input area for a cleaner, more consistent look. Upload and voice are de‑emphasized so they’re not confused with the primary send action, while still easy to find when needed.
User (Beta) Simpler prompt menu: My Prompts and Suggested Prompts are in a compact bar below the input with clear labels, so you can open and pick prompts faster.
User (Beta) Easier tool selection: Tool selection is available via a “+” menu, with grouped tools (MCP tools) in dropdown groups for easier separation of concerns.
User (experimental) Cleaner answer layout: User messages use a bubble style on the right; assistant messages retain a flat style on the left for a clear question-and-answer distinction. Assistant message avatars are no longer shown, making threads easier to read, especially on mobile.
User (experimental) Sources and content with the answer: Files and sources linked to an answer are shown as clickable badges directly under that answer’s text instead of as separate blocks between messages.
User (experimental) Unified message actions: Copy, feedback (thumbs up/down), references, and debug sit in a single action row for assistant messages. References and debug open in a side panel so you can keep reading the thread while viewing references or debug info. On user messages, actions appear on hover to keep the default view minimal.
User (experimental) Tables with copy: Tables in assistant messages use an updated design with a copy button in the lower-right of the table (visible on hover on desktop, always visible on mobile) so you can copy table data quickly.
User (experimental) Updated suggested follow-up questions: Follow-up questions use a clearer design with a list-style layout so suggested next questions are easier to scan and tap.
User (experimental) Floating hallucination indicator: The previous Answer Assessment button is removed. If no hallucination is detected, nothing is shown. When the model flags potential inaccuracies (moderate or severe hallucination), a floating button appears in the lower right (yellow for moderate, red for severe). Hover over the button to view the reason for the hallucination. It stays visible while you scroll so you can still see and act on the warning.
Requesting Approval in Chat
The new elicitation behavior (in-chat approval and query review) is experimental and off by default. “Query elicitation for Web Search” can be enabled for selected organizations or environments through advanced configurations in Space Admin. Full documentation here.
User (experimental) In-chat approval form: Tools and assistants can now ask for your approval or input in the middle of a conversation. When they do, the form appears in the same message thread with a clear “Requesting Approval” label. You can Confirm or Cancel, so you stay in control before the tool continues. Your response is sent back to the tool so it only proceeds with what you approved.
User (experimental) Query review for Web Search: When using Web Search with query elicitation enabled, you can review and edit the search queries before they run. A form in the chat shows the proposed queries; you can add, remove, or change them, then confirm or cancel. This gives you a chance to correct or refine searches before they execute.
User (experimental) Consistent form experience: The approval form supports different field types (standard MCP Elicitation schema) and validation, so behavior is consistent whether the request comes from an MCP tool or an internal feature like Web Search query review.
Improvements
Admin Whisper OX API key auth: The speech service now supports optional API key authentication for Whisper OX (self-hosted) endpoints. Configure WHISPER_OX_API_KEY and WHISPER_OX_API_KEY_HEADER to secure the connection. Backward compatible — no auth when unset.
Admin (experimental) Secure HTML Rendering via HTML Store: HTML content generated in chat responses (e.g. financial charts, rendered HTML blocks) is now served through a dedicated HTML Store service with TTL-based memory management. Large HTML content (>2MB) shows a download button instead of rendering inline. See Code Reflector Infrastructure for details.
👉 Why this matters?
The modernized interface reduces cognitive load by hiding secondary tools until they are needed, while the new one-click table copying and integrated source badges significantly speed up report generation. Security is bolstered by requiring confirmation for external links and providing real-time "hallucination" warnings, ensuring you can trust the AI's output. A
dditionally, the new Code Execution tool allows for complex mathematical modeling directly within the chat, and Whisper-powered voice input makes capturing meeting notes faster and more accurate than ever.
2. Better Knowledge Management (RAG)
What's New (Business/End-User Focus)
User (experimental) Email File Ingestion (.eml and .msg): You can now upload and ingest Outlook .msg and standard .eml email files. Email body text is extracted and cleaned (signatures and quoted replies are stripped), and attachments are automatically converted and included.
Admin (experimental) Skip CSV File Ingestion: You can now choose to skip the ingestion of CSV files by setting the new SKIP_EXCEL_INGESTION flag as the uniqueIngestionMode. This gives you greater flexibility to customize your ingestion process and focus only on the file types that matter most to your workflow.
Admin (experimental) Reporting Engine (.uqtemplate): A new reporting engine allows automated report generation from ingested content using customizable .uqtemplate template files. Configure reportTemplates in the ingestion settings to trigger report generation on ingestion. Generated .uqreport outputs are stored as new content items.
Admin (Beta) File Access:
File-based access control is now available for all clients in the Knowledge Base, enabling permissions to be configured at the individual file level rather than only at the folder level. This provides greater flexibility when managing sensitive or team-specific content, ensuring that access can be tailored precisely to business needs without restructuring folders.
General overview of the file-based access control feature
Detailed technical description of the file-based access control feature:
User (experimental) Multi-File Select and Move In Knowledge Base Column View:
Multi-file selection capabilities are now available in Column View, making it faster and more intuitive to organize content within the Knowledge Base.
Files can be selected using familiar desktop interaction patterns:
This update reduces repetitive actions and enables faster restructuring of content within large Knowledge Bases.
User Agentic Metadata Extractor:
It's useful when you've updated your metadata schema or configuration and want to apply it to files that were ingested before the change.
👉 Why this matters?
These updates allow you to bypass manual "detective work" by ingesting clean email correspondence directly into research workflows and using automated reporting to summarize new data. Moreover, the ability to manage permissions at a file level and re-extract metadata on existing collections ensures that your knowledge base remains secure, accurate, and easily searchable as your business needs evolve.
3. Improvements to Agentic Tables
What's New (Business/End-User Focus)
User Sheet sharing and access permissions: Agentic Table sheets can now be shared with members of the space using two permission levels: Manage and Edit. Manage allows changing the sheet name and managing who has access; Edit allows working with the sheet (e.g. adding sources, running the agent, editing content) but not changing the name or managing permissions. This supports collaboration within the space while keeping control over access and naming.
User Remove sources from the sheet: Agentic Table sheets now support removing files from the source list. A removed file is excluded from future agent context, while existing references and links in the sheet stay intact. This allows the source set to be corrected or narrowed without affecting prior answers or citations.
Improvements
User Groups in assignee column: The assignee column now shows both individual users and groups that have access; when an owner shares with a group, all members and the group itself appear in the assignee list.
User Processing after adding sources to an agentic table: When sources are added to an agentic table, processing is no longer started automatically, which avoids processing large sheets unintentionally and gives users more control over when and what to reprocess. Re-run is available to process all rows or a single row after sources are added.
Admin Use RfpAgent instead of Due Diligence when creating Agentic Table in Space UI: Creating a new Agentic Table from the Space UI now uses the RfpAgent module instead of the deprecated Due Diligence module, so users always work with the current Agentic Table implementation.
User UI Polish for Agentic Table: Consolidated UI refinements including a refreshed rerun button, improved dropdowns on the sheet overview, a new loading preview for the agentic table, and updated inline reference badges for consistent styling and clearer readability.
User Collapsible Content Support in Agentic Table Cells: Cells now support HTML details/summary for accessible, styled expand/collapse with proper overflow handling.
User RfPAgent Subagent: Persistent Chat Context: Subagents now reuse a persistent chat_id via ShortTermMemory, enabling them to understand and leverage broader conversation and table context – not just single lines – so follow-up questions are answered more reliably.
👉 Why it matters
These updates transform Agentic Tables into a robust collaboration hub by introducing multi-user permissions (Manage vs. Edit) and deeper "contextual memory," ensuring your team works from a single source of truth. By decoupling source ingestion from automatic processing, users gain critical control over computational costs and time management for large datasets. Furthermore, moving to the RfpAgent module and introducing UI enhancements like collapsible cells and persistent chat context ensures that complex analyses remain readable, reliable, and easy to maintain over time.
4. Unique AI Spaces
What's New (Business/End-User Focus)
User (experimental) Customize end-user space - A user can create, edit and delete multiple end user spaces. End-user spaces are customized spaces for a particular assistants (The instructions from the end user doesn’t alienate the assistant instructions, but appends to the assistant instruction).
Admin (experimental) Enable End-User Spaces: An Admin can now enable end user spaces on an assistant, which allows a user to create customized spaces. See here for enabling the end user space on space 2.0, and here for enabling on Custom spaces.
User Sub-Agents Now Inherit Uploaded Files from Orchestrator: Sub-agents now automatically have access to files uploaded to the main agent (or orchestrator). Any documents provided at the orchestrator level are shared with downstream sub-agents, ensuring consistent context and seamless task execution across the agent hierarchy.
Improvements
User Enable Sub-Agents to Access Files from Main Agent Chat: Sub-agents can now access files uploaded in the main agent chat. This improves usability, and ensures consistent file availability across parent and sub-agent interactions.
User Support Rendering of Base64-Encoded Images from MCP Tool Responses: Unique AI now supports rendering base64-encoded images returned by MCP Server tool calls, fully aligned with the MCP Image Content specification. The base64 data is automatically converted (e.g., to PNG format) and can be directly accessed and displayed within agent responses.
👉 Why it matters
By allowing end-users to create and customize their own spaces, teams can tailor assistant behaviors to specific niche tasks without overwriting core organizational instructions. This update also eliminates manual file-sharing friction; sub-agents now "inherit" context from the main agent, ensuring complex, multi-step tasks are executed with full visibility into all uploaded documents. Finally, the support for base64 image rendering from MCP tools enables more sophisticated visual outputs, making the platform a more versatile tool for data visualization and technical analysis.
5. Next-Generation AI Models
What’s New (Business/End-User Focus)
Admin Model-Agnostic Token Counting: Token counting has been refactored to be model-agnostic, removing the tight dependency on the OpenAI tokenizer and enabling accurate accounting across supported LLMs. The tokenizer can now be configured directly within the Language Model settings. Currently supported tokenizers include OpenAI encoders as well as Qwen and DeepSeek, with additional tokenizers available on request. When switching models, ensure the tokenizer is correctly adjusted in the Language Model Info configuration to maintain accurate token tracking.
New Model Deployments:
Anthropic
Admin Claude Opus 4.6 is available through LiteLLM using the following enum:
xAI
Admin Grok 4.1 models are available through LiteLLM using the following enum:
Gemini
Admin Gemini 3.5 Pro Preview model is available through LiteLLM using the following enum:
⚠️ Upcoming Module Retirement
The following AI modules will be retired in 2026.12:
If you are currently using any of these modules, please reach out to your Customer Success contact to plan the replacement.