Release Notes

Side-by-side editor and AI coaching

Open Build Instructions from any agent's page to load its live instructions into a side-by-side editor and AI chat.

Grounded in the agent's own knowledge

Pick a Knowledge Source project so the assistant knows what the agent is trained on, and optionally turn on Use Human Feedback (via the + options menu) so recurring complaints become concrete fixes.

Accept/Deny suggestion cards with versioned saves

The AI coach proposes one change at a time as Accept/Deny suggestion cards — nothing touches your draft without an explicit accept, and every save creates a versioned snapshot with revert.

Grade and Publish

Grade scores the draft 0–100 across role clarity, guardrails, tone, and more, and Publish pushes the result straight to the agent.

Manual editing, preview, and download

Edit or paste instructions manually, preview the rendered Markdown, copy, or download the draft at any time.

Lock an organization to specific email domains

Organization owners and admins can set an allowed-domain list in organization settings — members outside it lose access, and new invites are checked against it.

Owners and platform admins stay exempt

Owners, the organization's creator, and platform admins are always exempt so no one can lock themselves out.

Impact preview before saving

An impact preview shows exactly who would be affected before the policy is saved, and every change is written to the audit log.

Two-factor check when granting admin roles

Promoting a member to Organization Admin/Owner or Project Admin now asks for your authenticator or email code at the moment of promotion.

Enforced server-side

The role change is rejected without a recent verification, so the requirement applies no matter how the request is made.

Other role changes are unchanged

Demotions and non-admin role changes work as before, with no extra step.

Sync health per connector

The Connectors page now shows a health indicator for each connected connector based on its recent syncs, so failing connections surface immediately instead of hiding in error logs.

More reliable organization filtering

Filtering to organizations with connections is more reliable — shared connections now count correctly.

Specific skip reasons

Sync results include specific skip reasons, so "14 items skipped" tells you why.

Several MCP bridges per Salesforce connection

One Salesforce connection can now power multiple MCP bridges, each with its own token, tool allow-list, and write switch — revoke one agent without affecting the others.

SharePoint sub-site links resolve correctly

Pasting a SharePoint sub-site link now resolves and previews the correct site before import, with optional folder-structure recreation and deletion mirroring.

Salesforce activity logging and contact lookup

Salesforce activity logging sets the proper call subtype, and searches can look records up by contact.

Mixpanel setup guide

New external MCP setup guide for Mixpanel.

Chat in 20+ languages

Live Chat now ships with full localization, including Bulgarian, Czech, Danish, German, Greek, Spanish, Estonian, Finnish, French, and Russian, among others. Translations load with the widget so visitors see the interface in their own language from the first message.

Language switcher in the widget

Visitors can change the chat language themselves from a switcher in the widget, without reloading the page or starting a new conversation.

Set the language from your page

A new SET_LANGUAGE command in the launcher and iframe SDK lets your site set or change the chat language programmatically at runtime — useful when the visitor picks a locale elsewhere on your site.

Voice-to-text in the chat input

Visitors can record a voice message and have it transcribed directly into the chat input, then edit it before sending.

Per-agent transcription control

Audio transcription is a new on/off setting in the Chat skill configuration, so you can enable voice input only for the agents where it makes sense.

Clear unavailable state

When a Live Chat skill is archived or inactive, the chat and its settings pages now show a proper unavailable state instead of appearing functional. Visitors get a clear message rather than a chat that silently fails.

Automatic configuration on setup

Adding the Live Chat skill to an agent now initializes its configuration automatically, so the widget is ready to configure immediately.

Edit MCP integrations after creation

Agent MCP integrations are now fully editable — change the connection, reselect which tools are exposed to the agent, and validate the connection without deleting and re-adding the integration.

Slack skill now available

The Slack skill is enabled across the platform, so agents can be connected to Slack alongside existing channels, with message delivery improvements included.

raia Command in the sidebar

Admins now have a direct link to raia Command from the main navigation.

Multi-turn function calling

Response handling has been reworked so agents can carry function calls across multiple turns, allowing more complex, multi-step tool use within a single conversation.

Dedicated subagent conversation handling

Subagent execution now uses its own conversation management, improving attribution and traceability when a supervisor agent routes work to a specialist.

Non-blocking conversation webhooks

Conversation-start webhooks now run in the background rather than on the request path, so a slow or failing endpoint no longer delays the conversation. Webhook errors are handled independently of the conversation itself.

Scraping resources enable their skill automatically

Creating a scraping resource through the external API now enables the Scraping skill on the agent automatically.

Functions assign their skill automatically

Creating a function through the external API now assigns the Function skill to the agent, so newly created functions are usable right away.

Reasoning tokens in voice cost accounting

Reasoning tokens are now included in cost accounting for voice conversations, giving a complete picture of what a call actually consumed.

Provider-qualified model names in usage records

Model usage is now recorded with provider-prefixed names (for example, openai/), making usage and cost reporting clearer when agents run across multiple providers.

Rewrite a document's Markdown with a prompt

Enhance with AI opens a prompt box with presets — clean up formatting, summarize, restructure into sections — and streams the result back into the document view.

Side-by-side review before accepting

Compare the proposed Markdown against the current version, then accept or discard it.

Accepted output saves like a normal edit

Saving goes through the same permission checks as any other edit, and is re-encrypted for organizations with document encryption enabled.

Send to AI without leaving the page

A new Send to AI button on the document view opens Send to AI already scoped to that file.

Direct uploads to agents

Files are now uploaded to agents directly by raia Command instead of going through an external workflow, with duplicate protection built in.

Per-agent document tracking

Every uploaded document is tracked per agent, so re-sending replaces the existing asset instead of duplicating it in the vector store.

Training queue with job status

A training queue shows queued, running, completed, and failed jobs, grouped by organization with a collapsed summary.

Rerun & replace

Re-send a document and clean up the previous copy on the agent in one action.

Final checks before upload

A final PII check runs immediately before upload, and jobs for archived files fail fast instead of hanging.

Clickable score cards and histograms

Quality score cards and histograms now open a searchable table of the files behind them.

One table view per project

Search, sort, and review status columns across all files in a project from a single table.

Open any row in context

Jump from a row into the document view, the detail drawer, or the file's folder.

New managed MCP bridges

Linear, Pinecone (read-only), Okta, DocuSign, Xero, Zapier, Bitbucket, Zendesk, Zoho, and Salesforce are now available as managed bridges.

Bring-your-own OAuth apps

Xero and DocuSign use a bring-your-own OAuth app model, so credentials stay under your control.

New Magentrix connector

Incremental sync, blogs/posts and attachments, MEQL queries, and agent tools.

Zoho Desk import consolidation

Zoho Desk imports are combined into a single Markdown document per batch (up to 10 MB), with knowledge base articles kept separate.

International formats recognized

The phone rule now matches and masks UK, EU, and other international numbering formats, not just North American ones.

Links stay unflagged

URL-embedded matches remain suppressed, so links are not flagged.

Rescan existing documents

Existing documents can be rescanned from the project's Scan page.

Edit without re-uploading

Open a converted document's output and use Edit to load the raw Markdown into an editor, change it, and save.

Permission-checked saves

Saving requires upload rights on the folder the file lives in.

Automatic re-encryption

Edits are re-encrypted automatically for organizations with document encryption enabled.

Cancel discards the draft

Cancel returns to the rendered view without saving changes.

New Agent Report Dashboard

A new dashboard surfaces comprehensive per-agent statistics, including users reached, conversation volume, channel breakdowns, token usage, traffic patterns, and more.

Expanded Agent Management API

The public API now exposes richer endpoints for building and managing agents programmatically, giving teams more control over agent configuration, deployment, and lifecycle operations.

Slack and Microsoft Teams integrations for agents

Agents can now be connected to Slack and Microsoft Teams so end-users can interact with them directly inside those channels alongside existing chat surfaces.

Custom delay setting for Email and SMS

Email and SMS settings now include a configurable delay, so outbound messages can wait a short, randomized interval before sending to better simulate human response timing.

Two-factor verification for downloads — on by default

Downloading files, Markdown, or bulk exports now requires confirming a 6-digit code emailed to you. The gate is enforced on the server, and every download is written to the audit log.

Enabled by default for every organization

New organizations start with it on, and all existing organizations were switched on.

Broad coverage across download paths

Covers original file downloads, bulk signed-URL/source ZIPs, Markdown downloads, project/folder Markdown ZIPs, merged Markdown, and exports.

Time-boxed verification window

Verification opens a time-boxed window (default 15 minutes, configurable 1–120) stored server-side, so the browser cannot skip or forge it.

Code expiry, lockout, and rate limiting

Codes expire after 10 minutes, lock out after 5 failed attempts, and are rate limited per user and per IP.

Previews stay ungated

Single-file in-app previews stay ungated so normal reading keeps working; Markdown reads are gated for downloads or requests pulling more than 5 files.

Expanded audit logging

Server-side audit logging now records signed-URL and bulk ZIP downloads that previously left no trace.

Per-organization control

Owners and admins can turn it off per organization in Settings → Organization.

Optional document (Markdown) encryption per organization

Organizations can now encrypt the stored Markdown of every document with AES-256-GCM, layered on top of the infrastructure-level encryption the database already provides.

Per-organization data keys

Each organization gets its own unique 256-bit data key, generated automatically and stored wrapped by a master key (envelope encryption).

Enablement and acknowledgement

Enable it from Settings → Organization → 'Document (markdown) encryption' — visible to Owners, Admins, and platform admins only, with a required acknowledgement that losing the key means losing the content.

Backfill existing Markdown

New writes are encrypted immediately; use 'Backfill existing markdown' to re-encrypt everything already stored in resumable batches with live progress.

Server-side decryption keeps workflows intact

Decryption happens server-side only, so previews, exports, downloads, PII scanning, agent training, and vector-store uploads all keep working.

Trade-off: keyword search on document bodies

SQL keyword search can't match encrypted document bodies — file name, metadata, and label search are unaffected.

Account security hardening

A round of platform-wide account protections covering passwords, admin MFA, brute-force attempts, and who can register.

Stronger password requirement

Password minimum raised to 12 characters.

MFA required for Owners and Admins

Multi-factor authentication is now required for organization Owners and Admins.

Sign-in protections

Repeated failed sign-ins temporarily lock the account, and sign-up/sign-in are protected by CAPTCHA.

Business email domains only

Personal/consumer email domains can no longer register or be invited.

Expanded audit trail

Expanded security audit trail covering blocked login and OTP attempts.

Automatic PII & secret masking on every file

Every file converted to markdown — whether uploaded directly or ingested through a connector — is now automatically scanned for sensitive data and masked in place before it is served to agents, previews, or downloads.

Runs on every conversion path

Scans run automatically as soon as markdown is produced by any conversion path (uploads, connectors, reconvert, Pro processing).

Destructive rules mask in place

Destructive rules (SSN, credit card, API keys, email, phone) rewrite the stored markdown with masked values (XXXXXXX). The original binary in storage is untouched.

Heuristic rules report without rewriting

Heuristic rules (generic secret, high-entropy strings, digit-only patterns) record findings without altering text so you can review before enforcing.

Global and per-project rules

Global rules apply org-wide by default; each project can add its own rules or override the defaults from the project's Security Scan page.

Scan History and Scan Runs dashboard

New per-file Scan History with the last 25 runs, redacted markdown preview, and a project-level Scan Runs dashboard with aggregate finding counts.

Backfill historical files

Backfill historical files any time — from the project's Security Scan page (Rescan all / Rescan flagged) or, for platform admins, from the org drawer in Admin.

Masked markdown everywhere downstream

Downloads, agent training, and MCP responses all read from the masked markdown once a destructive rule has matched, so raw PII never leaves the platform.

Original files auto-delete after 30 days

Once a file has been converted to markdown and passed its PII scan, the original binary is removed from storage after your organization's retention window. Markdown, findings, notes, and AI outputs stay forever.

New 30-day default

Originals are deleted 30 days after upload. All existing organizations were seeded with this 30-day setting.

Configurable per organization

Set any number of days from Settings → Organization, choose 0 to delete as soon as markdown exists, or toggle Keep forever to opt out.

Conservative hourly sweep

An hourly sweep only deletes files that (a) have markdown, (b) have completed a PII scan (clean or flagged), and (c) are older than the retention window. Nothing else is touched.

Clear state in the file drawer

The file detail drawer clearly shows when an original has been swept and disables the Download original button — everything else (markdown preview, exports, agent access) keeps working.

Accurate storage reporting

Storage totals continue to reflect the file's original size so usage reporting stays accurate after cleanup.

Admin visibility and on-demand runs

Platform admins can view sweep history and trigger an on-demand run per org from the Admin → Org detail drawer.

Instant browser-side conversion

Text-based files such as DOCX, TXT, and source code convert in the browser as you upload. A 99-page DOCX that used to take about 14 minutes in the old queue now converts in roughly 1–2 seconds.

Library-based extraction for common formats

PDF, XLSX, PPTX, and other supported formats use specialized extraction libraries for near-instant text output.

AI quality scoring on every file

Each Markdown output receives a training-quality score, and every organization can tune the scoring prompt to match its own standards.

Pro Processing for scanned files and images

Files that score 0 — typically scanned PDFs and images — can be reprocessed with Gemini 3 Flash in one click. The option only appears for formats that genuinely need OCR.

Parallel processing for large PDFs

Large PDFs are split into chunks, processed concurrently, and reassembled in one atomic step, with live progress in the Admin Pro-Processing Queue.

YAML frontmatter on every file

Converted files now carry a structured metadata header for cleaner downstream use.

Historical backfill

Administrators can re-run the new conversion logic across previously uploaded files for any organization.

Drag-and-drop upload opens the upload drawer

Dropping a file now opens the upload drawer with that file preselected, so AI classification is never skipped by accident.

Search inside documents

A dedicated search view lists matching documents and highlights the search term inside each preview.

Better AI classification

Classification now reads the first several kilobytes of converted Markdown to suggest a more accurate title, summary, and tags.

Vector store cleanup on archive

Archiving a file also removes it from any attached agent's vector store, with a confirmation modal when the file is in use by an agent.

Orphaned file cleanup

A cleanup pass identified and archived 104 files that no longer had a backing storage object.

Pro badge on processed files

Files completed through Pro Processing now show a visible Pro badge in the Data Room.

Agent API key monitoring

A daily check flags agent API keys that are invalid, expired, or set for the wrong region, and surfaces banners so you can fix them before an import fails.

Upload without picking a folder

You can upload first and choose a destination afterwards, or let AI pick the folder for you.

More than 25 connectors across seven categories

Code repositories (GitHub, Bitbucket, Azure DevOps Repos), cloud storage (Google Drive, SharePoint, OneDrive), ticketing and project management (Jira, Zoho, Autotask, Azure DevOps Work Items), CRM and support (Salesforce, Zendesk, ServiceNow, HelpJuice), web scraping (Basic Website Scrape, Firecrawl, Adobe RoboHelp), Microsoft Teams channels, plus HTTP Request and SOAP endpoints.

Fully asynchronous imports

Connector jobs run entirely in the background with sync tracking, so you can close the browser tab while an import finishes.

Bitbucket multi-repository imports

Select multiple repositories and branches within a single import configuration.

Jira filter builder with live JQL preview

Build multiple filters with operators and enum dropdowns, filter by version, and preview the generated JQL as you go.

Zendesk background imports with output choice

Imports run in the background with a live progress counter, and you can produce one combined Markdown file or one file per article.

Basic Website Scrape up to 500 pages

Scrape up to 500 pages with a 500 ms delay between requests, choose combined or per-page Markdown output, and watch progress live.

Firecrawl output and path controls

Scrape failures are fixed, and there are new output options, header/footer controls, ignored paths, and YAML headers.

HTTP Request and SOAP endpoints

Pull from any URL with a GET request and custom headers — including key-protected sources such as sales playbooks — or connect legacy enterprise systems over SOAP.

Adobe RoboHelp nested table of contents

RoboHelp imports now recursively parse toc.js files to discover article paths across nested folder structures.

Live tool access without importing data first

Command now hosts or documents MCP servers so agents can work with external platforms in real time, instead of importing that platform's data first.

Hosted servers

Salesforce (OAuth), Gmail with search_emails, get_thread, send_email and create_draft (OAuth), Zendesk (OAuth or legacy API key), Zoho (12 tools, OAuth), HubSpot (15 tools, personal access token), Azure DevOps (24 tools across wikis, repositories and work items, OAuth), and Microsoft Teams (5 tools, OAuth).

Documented external servers

Guidance for connecting the official GitHub, Zapier, and Bitbucket servers, plus Figma read tools (write operations aren't available through Figma's REST API).

Scope probing exposes only permitted tools

When Command generates an MCP URL and token it probes the credential's authorization scopes and exposes only the tools those credentials actually permit.

Reporting reorganized into Activity, Quality, and Tokens

Reporting is split into three tabs so operational, quality, and usage analysis are easier to find.

New quality metrics

Added conversation counts by channel, average score, and counts within each score range.

Escalation response time

Tracks average and median time from escalation to assignment to a team member.

Average user rating detail page

A conversation-rating detail page provides rating analysis with download support.

Jump from escalation to Copilot

The escalation card links straight into Copilot for faster review of the underlying conversation.

Full conversation counts

The old 1,000-conversation cap is gone — reports now show complete counts and charts update dynamically for the selected period.

External app custom reports

A Salesforce Cases integration pulls ticket data and calculates metrics such as Ticket Deflection.

Admin token usage reporting

A dedicated tab tracks tokens consumed by in-app AI actions, broken down by organization, project, provider, and model.

Multi-Agent Orchestrator

Chain multiple specialized agents together to create complex, intent-driven workflows. A "Supervisor" agent evaluates the user's request and intelligently routes it to the most appropriate subagent. Subagents execute their task and return the result — the supervisor can then refine the response before delivering it to the user.

Keyword-Based Routing

Configure specific keywords on each subagent to enable deterministic routing. When a user's message contains a configured keyword, the message is routed directly to the matched subagent — bypassing the main agent's LLM entirely for faster, more predictable responses.

Direct Subagent Selection in Copilot

End-users can now manually select which subagent should handle their message using a selection interface in the Copilot input area. This gives users explicit control when they already know which specialist they need.

Subagent Response Attribution

Messages handled by a subagent are now clearly marked in the conversation UI with a distinct icon and the subagent's name, making it immediately clear which agent generated each response.

AI-Optimized Subagent Descriptions

An integrated "Optimize with AI" button rewrites your subagent descriptions into 75–150 word prompts specifically formatted for optimal LLM tool selection accuracy.

Visual Orchestrator Graph

A built-in React Flow visualizer displays the hierarchical relationship between your agents. Active subagents are outlined in green, archived ones in red. Click any node to navigate directly to that agent's configuration.

MCP Integrations (Client Mode)

Connect your agent to external MCP (Model Context Protocol) servers and use their tools natively. Your agent can now call tools hosted on any MCP-compatible server — CRMs, databases, internal APIs, or third-party platforms — directly within conversations.

Tool Discovery & Selection

When adding an MCP integration, click "Get Available Tools" to probe the remote server and browse all available tools. Select which tools to expose to your agent — only selected tools are passed to the LLM, keeping context lean and routing accurate.

Streamable HTTP & SSE Transport

MCP integrations support both Streamable HTTP (default) and Server-Sent Events (SSE) transport protocols, ensuring compatibility with a wide range of MCP server implementations.

RAIA MCP Server (Server Mode)

Expose your raia agent as a tool that other AI platforms can query. External systems can send prompts to your agent via a standardized MCP endpoint, and your agent responds using its full AI pipeline — instructions, vector store, and linked packs included.

OAuth & Bearer Token Support

Connect MCP servers using either Bearer token or OAuth authorization, giving you flexibility to match whatever your MCP provider expects.

Configurable Retention Periods

Control how long conversations are stored before they are anonymized. Set a retention period between 1 and 3,650 days (up to 10 years). Conversations older than the configured period are automatically processed.

Conversation Anonymization

When the retention period expires, conversation data is permanently overwritten — message content is replaced with [ANONYMIZED], and associated metadata (user names, emails, phone numbers, IP addresses) is cleared. This is a destructive operation designed for compliance — data is truly gone, not soft-deleted.

User Anonymization

Optionally anonymize the conversation users associated with expired conversations. User names, emails, phone numbers, and external IDs are cleared independently from message content, giving you granular control over what gets retained.

Audit Logging

Every anonymization run is logged to the Action Logger with counts of conversations and users processed, providing a full compliance audit trail.

Smart Cutoff Logic

The retention period is calculated from the most recent message in a conversation, not the conversation creation date. This ensures active conversations are never anonymized prematurely.

Daily Inbound Message Limits

Set a daily message limit per user to prevent abuse and control costs. Messages from email, SMS, live chat, copilot, and other channels all count toward one shared daily limit.

Cross-Channel Tracking

Usage is tracked per-channel (email, SMS, live chat, etc.) with a unified total. Admins can see exactly how many messages a user sent on each channel within the current day.

Per-User Custom Limits

Override the global default for specific users. Grant VIP customers a higher allowance, or restrict problematic users to a lower cap — all without affecting other users.

Temporary Unlimited Access

Temporarily grant a user unlimited messages (e.g., during onboarding or escalation resolution) without permanently changing their configured limit.

Automatic Daily Reset

All counters automatically reset at the end of each UTC day. No manual intervention required.

Quota Exceeded Alerts

When a user hits their daily limit, the system generates an alert visible in the Message Quota Skill's Alerts tab. Admins can review, adjust limits, or grant temporary unlimited access directly from the alert.

Zero-Token Blocking

Quota checks run before the message reaches the LLM. Blocked messages consume zero tokens, ensuring costs are fully controlled.

Organization Default Limits

Set baseline monthly caps for tokens, SMS, emails, and call minutes at the organization level. These defaults apply automatically to any agent or user that does not have a custom limit configured.

Agent-Level Usage Limits

Override organization defaults with per-agent monthly caps. Each agent can have its own token, SMS, email, and call minute budget — preventing any single agent from consuming disproportionate resources.

User-Level Usage Limits

Control resource consumption per end-user. Set monthly caps on how many tokens, SMS messages, emails, and call minutes each user can consume across all their conversations.

Real-Time Usage Dashboard

Monitor consumption with live progress bars on the Agent Usage tab. The dashboard auto-refreshes every 60 seconds, showing current usage against configured limits for all four resource types.

Subscription Ceiling Enforcement

Agent and user limits cannot exceed the organization's subscription maximum. The UI displays the maximum allowed value below each input field and warns admins if a subscription downgrade makes existing limits invalid.

Graceful Fallback Chain

Limits resolve in order: Custom Agent/User Limit → Organization Default → Subscription Limit. Deleting a custom limit immediately reverts to the next level in the chain.

Zero-Cost Enforcement

All quota checks occur before the action is executed. If a limit is exceeded, the action is blocked before reaching the LLM or external service — no resources are wasted on blocked requests.

Teams Channel & DM Support

Your raia agent can now respond to direct messages and channel mentions in Microsoft Teams. Users can interact with the agent naturally within their existing Teams workspace.

Scope-Based Access Control

Choose who can message your agent: allow all users in the tenant, or restrict access to specific users and channels using Microsoft Graph IDs.

Multi-Agent Picker

When multiple agents are configured for the same Teams tenant, the bot presents an Adaptive Card allowing the user to select which agent they want to interact with.

Pause vs. Disconnect

Pause an account to temporarily disable a specific agent's Teams connection without affecting other agents. Disconnect to fully remove the Microsoft OAuth connection for your entire organization.

File Upload Support

Users can attach up to 5 files (20MB each) directly in Teams messages. Supported formats include documents, images, and spreadsheets.

File Attachments on the Prompt API

The Prompt API now supports file attachments in addition to plain text. The workflow mirrors message attachments in conversations: (1) call the API to obtain a signed upload URL, (2) upload the document to Google Cloud Storage using the signed URL, and (3) attach the uploaded file to the prompt when creating or sending it by passing the returned storagePath in body.files.

Improved Country Selection

The country dropdown in Live Chat → Geographic Restrictions is now sorted A–Z with keyboard jump support, making regional blocking easier to configure.

Geographic Restrictions Behavior Fixes

Live Chat now shows a clear warning when a restricted user tries to send a message, and the chat no longer appears active while blocked. Copilot is unaffected — internal conversations are created normally regardless of region.

Auto-Assignment for Escalation

The Escalation Skill supports a default assignee — escalated conversations can be automatically assigned to a specific admin instead of left unassigned.

Escalation Webhook Payload

Escalation webhooks now include the full conversation payload, not just the form submission, so downstream systems receive complete context.

New Alerts Tab for Escalation

An Alerts tab has been added to the Escalation Skill to improve visibility into triggered events, aligning the UI with the Auditor Skill.

Notification Metadata Fix

Metadata collected from the Live Chat form is now properly passed in admin notifications, ensuring admins receive complete context on every alert.

Legacy Notifications Removed

The 'Notify on Auto-Archive' option and legacy email notifications have been removed from the Notification Skill to simplify configuration and reduce alert fatigue.

Auditor Skill Enhancements

Continued refinements to the PII and Hack Auditor engines improve detection accuracy and reduce false positives.

GPT-5.6 Model Family & Reasoning Mode

Added support for the GPT-5.6 model variants across the API, UI, and chat widget. Agents can now be configured with an explicit reasoning-effort setting alongside the model selection.

Slash Commands in Copilot

Copilot users can trigger predefined actions and shortcuts by typing `/` in the message input, streamlining common workflows.

Used Functions Display in Copilot

Copilot conversations now show which functions the agent invoked to produce each response, giving admins full visibility into agent decision-making.

Copilot File Handling Upgrades

Paste copied files directly into chat, paste very long text and have it auto-converted into a file, drag-and-drop files into the conversation, and attach up to 20 files per message (previously 4).

Conversation User Deletion & Anonymization

Admins can permanently delete a conversation user through a confirmation dialog. All associated conversations are anonymized before deletion, and anonymized conversations are excluded from live chat lookups.

Orchestrator Safety Guard

Function calls are now automatically disabled after a configurable max iteration count to prevent runaway orchestration loops between subagents.

MCP Integration CRUD & Configuration

Full lifecycle management for MCP integrations. Configure tools, headers, and authorization per integration, and wire them into an agent's AI config from a dedicated settings surface.

SSE and Streamable HTTP Transports

MCP integrations support both Server-Sent Events and Streamable HTTP transport protocols to work with any MCP-compatible server.

Per-Integration Tool Filtering

Probe a remote MCP server and select which tools to expose to the agent. Only selected tools are passed to the LLM, keeping context lean and routing accurate.

Custom Headers & Authorization

MCP integrations accept custom request headers and Bearer/OAuth authorization to work with any MCP-compatible provider.

MCP Tool-Call Status in Streaming Responses

The UI now displays live tool-call status while the agent invokes MCP tools during a streaming response.

Conversation Data Retention Setting

Configure a retention period after which conversations are automatically anonymized. Anonymization is logged for full compliance audit trails.

isAnonymized Fields Across Models

User, conversation, and message models now carry an isAnonymized flag so downstream systems can consistently filter anonymized records.

Comprehensive User Cleanup

Removing a user from an organization now cascades a full cleanup of associated records, ensuring no orphaned personal data remains.

Microsoft Teams Skill Launch

New Teams skill and integration shipped across API and UI. Agents can respond to DMs and channel mentions inside Microsoft Teams.

Teams Message Moderation

Added message moderation to satisfy Microsoft marketplace review requirements, followed by iterative updates from the review process.

User-Level Quota Tracking

Introduced per-user resource quotas alongside existing agent-level limits, giving admins fine-grained control over end-user consumption.

Auto-Archive for SMS and Email

Extended auto-archive configuration to SMS and Email skills, matching behavior previously available on Live Chat.

Simplified Agent Creation

Agent creation is now a single form instead of a multi-step wizard, making it faster to spin up new agents.

Restricted Document Management

Org owners and agent owners can now manage and download restricted documents from within the platform, with permission checks enforced end-to-end.

Query Performance Improvements

Broad database query tuning across conversation, auto-archive, and Copilot workloads reduces latency and load.

Extended Prompt Caching

Prompt caching now retains entries longer, meaningfully reducing repeat model calls for common workflows.

New Live Chat version / migration

A larger Live Chat redesign and migration was discussed as close to release and production deployment.

Live Chat background image customization

Upload and delete background images for the chat widget. Configure background image and background color by light/dark theme. Background image covers and scales within the widget. Color mask/fade uses the configured background color.

Live Chat templates / standard template setting

New template configuration was shown in Live Chat settings.

Live Chat after-hours / unavailable message webhook

Instead of storing deferred messages locally and sending later, after-hours messages are sent to a webhook. Supports sending multiple after-hours messages.

Live Chat limited-date availability

Live Chat availability can now be configured by specific dates, not just times.

Live Chat webhook payload now includes embedded page URL and user IP

Conversation metadata captures where the chat is embedded and the user IP.

Live Chat buttons: improved regex appearance logic

Button visibility conditions can parse assistant responses and manual assistant messages, not just user input.

Escalation / button behavior improvements

Discussion around using escalation logic to trigger buttons. Regex checking both user message and AI response. Proposed improvement: escalation auto-mode should also parse AI answer, not only user question.

Inline / mid-conversation forms

A form can appear inside a conversation based on similar appearance conditions as buttons. Use case: user says they want to register, and a form appears for event registration.

Conversation forms / user information forms

Forms in Live Chat for collecting user or conversation info. Added validation for maximum string length.

500-character validation on custom form fields

Frontend and backend validation added to prevent very long input strings in user and conversation forms.

Live Chat CSV/table export

Tables can be copied or downloaded as CSV in Live Chat.

Widget fixes

Issues found in widgets have been fixed.

Chat SDK async command support

SDK commands such as opening/closing chat, setting users/context, and sending messages can now be awaited using native JavaScript promises.

Prompting / enhanced prompting skill

New Prompting skill. Clients can activate the skill and override default prompts at the agent level. User message is transformed and enhanced before being sent.

Prompting toggle in Copilot

Copilot input includes a toggle/control for enhanced prompting.

Force prompting setting

Admins can force prompting to always be used.

Hide prompting toggle setting

Admins can hide the prompting control from users while still using prompting behind the scenes.

View enhanced/transformed prompt

Users and admins can inspect the transformed prompt that was actually sent. Discussed UI options: icon, modal, status pill, dropdown, or subtle prompt-enhanced indicator.

Prompting state persistence via OpenAI background processing

Prompt API improved to use OpenAI background processing and response IDs, storing state in Redis so Kubernetes pod restarts do not lose long-running prompt state.

GPT-5.5 model added

Latest GPT-5.5 model was added to the application.

GPT-5.4 Nano for Optimize with AI

Optimize with AI and completion API updated to GPT-5.4 Nano.

Formatted Copilot message downloads

Downloaded Copilot messages can render Markdown as formatted documents rather than raw Markdown.

Word/PDF/other document export improvements

Markdown now exports into cleaner Word/PDF-style documents.

Copy/download tables as CSV

Tables in Copilot can be copied and downloaded as CSV.

Increase Copilot attachment limit

Copilot file attachments increased from 4 files to 20 files.

Sticky save button for dirty forms

UI improvement so save button remains visible when a form has unsaved changes.

Copilot reskin / Launchpad core product reskin

Ongoing work to reskin Copilot and Launchpad.

New Prompting Skill

Prompting is now a first-class skill and appears in the skill list, with agent-level configuration.

Function Skill Context

Function skill invocations now include the conversation ID in their metadata payload for improved traceability.

Complete Live Chat Redesign

A full redesign of the Live Chat widget shipped across the API, UI, and widget repos. A design-token layer was introduced to unify styling across themes.

Widget Command API with Promise Acknowledgment

SDK commands (OPEN_PAGE, UPDATE_THEME, DESTROY, PATCH_SKILL_CONFIG, reset-conversation, and more) return promises that resolve when the widget confirms the action.

Widget Config from CDN JSON

Widget configuration is now delivered as static JSON from a CDN, dramatically improving initial load performance and cacheability.

Cross-Tab User Syncing

User identity now syncs across browser tabs so the same visitor sees a consistent conversation regardless of which tab is active.

GPT-5.4-mini and GPT-5.4-nano

Added support for the GPT-5.4-mini and GPT-5.4-nano models across the API, UI, and widget. The default model was updated to GPT-5.4 with low reasoning effort; the internal utility model moved to GPT-5.4-nano.

GPT-5.5 Support

GPT-5.5 became available as a selectable model for agents.

Background Response Processing

Long-running model responses are now processed in the background with a status-polling flow, so in-flight completions are preserved across service restarts.

Async Vector Store Uploads

Vector store file uploads run asynchronously with tracked upload status and improved error handling. The UI now detects and surfaces stale document processing.

Structured Response Data

Agent responses now persist structured responseItems and responseContext fields for downstream analytics and replay.

Agent-Level Usage Limits Foundation

First release of agent-level monthly caps for tokens, SMS, emails, and call minutes, with usage tracking surfaced on a new Agent Usage tab.

Enhanced Prompting Skill

A new Prompting skill expands and improves user prompts before they reach the agent. The expanded message is streamed and can be inspected in Copilot, and is tracked as its own resource-usage source.

Portable Storage Layer

A unified cloud-storage abstraction lays the groundwork for portable and self-hosted deployments across providers.

Enhanced API Key Management for External APIs

Refactored API key management to support multiple keys at both the organization and API skill levels. Added expiration options (TTL or fixed date) and key status handling (Active, Expired, Revoked). Existing keys remain fully supported to ensure backward compatibility.

Training Documents Access Restriction

Introduced a setting allowing training document downloads to be restricted to the agent owner only, giving admins finer control over the Vector Store contents.

GPT-5.4 Model Integration

The GPT-5.4 model has been integrated into the platform. Analytical capabilities have been expanded, allowing the model to be selected and utilized within AI agents for complex reasoning and improved response accuracy.

Manual Vector Store ID Configuration

The ability to edit the Vector Store ID for individual agents has been implemented. This feature allows multiple agents to be pointed to the same knowledge base, enabling shared data access while streamlining the management of vector storage across different agent configurations.

Expanded File Format Support

Support for the most commonly used file formats, including PDF, DOCX, XLSX, and CSV, has been implemented across Copilot, Live Chat, and Email Skill. A new conversion pipeline has been introduced to process these documents into AI-friendly formats before they are stored in the Vector Store.

Archived Conversations Visibility and Notifications Cleanup

A Show Archived filter has been added to the Copilot conversation list, allowing archived sessions to be viewed alongside active ones.

Expanded System Event Messages in Conversation UI

Support for muted system messages has been extended within conversation transcripts for both Live Chat and Copilot. By incorporating a broader range of system events into the interface, the tracking of user actions and the restoration of activity logs have been made more transparent and efficient for administrative monitoring.

Optimize UI of Live Chat Settings Panel

The Live Chat settings interface has been redesigned to improve navigation and clarity. Existing configurations have been reorganized into clearly defined tabs and application-specific sections, complemented by the introduction of accordion UI components to manage high-density settings and reduce visual clutter.

Added Auto-Assignment Configuration for Escalation Skill

Escalation Skill now supports configuring a default assignee for each escalation type. When an escalation is triggered, the conversation can automatically be assigned to the specified admin instead of remaining unassigned.

Added Option to Disable Live Chat Access

A new configuration option allows Live Chat access to be disabled while keeping the agent available through Copilot. This is useful when an agent is intended to be used only within Copilot rather than exposed on a public Live Chat link.

Added Control for File Links in Copilot Annotations

A new Copilot setting allows administrators to enable or disable file links (hyperlinks) in document annotations, giving organizations control over whether underlying training files are downloadable from annotations.

Standardized Email Templates for Consistent Branding

Introduced a unified structure and styling standard for all system-generated email templates to ensure a consistent user experience and alignment with organization branding. This update standardizes elements such as headers, footers, greetings, layout spacing, and the use of organization logos or names. It also ensures that email templates properly reflect organization branding settings where applicable.

Fixed Validation Logic for Optional Fields in Live Chat Pre-Chat Form

Corrected validation behavior in the Live Chat pre-chat form where fields configured as Not Required still prevented users from starting a chat if left empty. The Start Chat button is now enabled as long as all required fields are completed, ensuring the form behavior correctly matches the admin configuration for optional fields.

Fixed Inconsistent Authors in Copilot Activity Log

Resolved an issue where activity entries in the Copilot Activity Log could display inconsistent or incorrect authors for certain actions. The log now consistently records and displays the correct author for each activity, improving traceability and audit accuracy.