Product updates from Redbelt AI

Personal tasks can use a saved agent to work through a task list automatically. Reuse and edit agent settings across tasks, workflows, and Site refreshes.

New workspaces keep their setup progress if checkout is cancelled or interrupted. You can return to the same workspace, retry payment, and see when access is ready.

Workflow progress stays current as work finishes. Replies, questions, and generated workflow cards stay in order during a conversation and after refreshing the page.

Expanded workflow steps keep the completed run visible, and work that finishes during a refresh updates without leaving the conversation stuck in progress.

Copying a full response or writing draft works while the content loads, including in browsers that require the copy action to begin with your click.

Self-hosted installations gained end-to-end lifecycle and integration workflows, along with expanded plugin and model-provider configuration, so an isolated deployment can be operated without direct database access.

Knowledge previews render faster and connection setup walks through fewer dead ends. Database discovery streams results as they are found, and permission failures explain what is missing instead of failing quietly.

Automatically triggered file and email workflows include a clickable source attachment in the conversation, so you can open the original document while reviewing the results.

Microsoft search and calendar results indicate when more matches are available, so a limited result list is easier to interpret.

Workspace sandboxes recover cleanly from interrupted activation, and release recovery handles partial deployments without manual intervention.

Search across knowledge now runs name-first with an instant typeahead, and filters moved out of a menu into pills you apply after searching. Results have a real empty state, rows stack into two lines at narrow widths, and each row carries an Info option for provenance.

Tasks can be organized into folders, with tables that stay readable from a phone to a wide monitor.

The sandbox runtime that executes code and tools moved onto the self-hosted architecture, with stricter isolation between workspaces and faster image builds.

Entitlements now resolve from a single authoritative source, and invoicing correctly preserves discount and credit lines.

Sites turns a thread into a custom interactive dashboard or website, published from the workspace and refreshed from the same knowledge the agent uses.

A watchOS app records voice notes on the wrist, hands them to the phone through a durable inbox, transcribes them, and uploads them as notes. Recordings survive a crash or a cold launch instead of being lost.

Auto-recorded meetings produce a house-style summary from the stored transcript, and each meeting reports its own usage so notetaker time can be attributed to the right team.

When several people join the same call, Redbelt attaches to the meeting bot already there instead of sending a second one, shows a live indicator per meeting, and fans the finished transcript out to every participant who opted in.

Members and admins are notified as they approach their monthly limits, and account deactivation and deletion now purge cleanly across the workspace.

Redbelt can join a meeting, stream the transcript live, and store the finished record as both Markdown and structured JSON in a meeting folder named for the call and its local time.

A live pane turns the running conversation into grounded cards — topics, summaries, and answers to questions actually asked — and any card can be handed off into a full thread.

Connect a Google or Microsoft calendar, review upcoming meetings, and toggle recording per meeting. Transcripts fan out to every opted-in participant, and canceled events unschedule their bots automatically.

An iOS app shipped with a native navigation drawer, iPad support, and a native composer carrying a model picker, a thinking-effort slider, and on-device speech input.

OneNote notebooks and pages sync into knowledge over OAuth, alongside the existing Microsoft connectors.

Personas can be generated, refined, and previewed from a dedicated editor, and members can change their own password or delete their account.

A convert step turns Markdown into Word or PDF and JSON into Excel or CSV mid-workflow. Converted files become artifacts that later steps — including save to folder — can consume directly.

Support cases can be opened from inside the product, tracked under My Support, and raised directly from a thread or a single response so the context travels with the case.

The workspace received a full visual pass covering navigation, write mode, canvas, and the workflow editor, including panel resizing and a mobile workflows page.

An Atlassian MCP suite connects Jira and Confluence to agents, with a guided flow for adding further MCP servers.

A three-phase redesign reworked layout, navigation, and density across the product, carrying monitoring and alerting into the same surface.

Threads can be filed into folders, tagged, and pinned, with a restructured dock, a knowledge pane, per-thread settings, and drag-and-drop on desktop and mobile.

Agents gained skills — packaged capabilities that sit alongside tools and can be selected per agent.

An Outlook add-in and a browser web clipper bring email and web pages into knowledge without leaving the source application.

Administrators can manage which file types are accepted, with validation applied at upload time, alongside security hardening across the API.

A document table query tool runs SQL directly over tables extracted from documents, with generation and execution steps available to workflows and blocked-procedure checks on every statement.

Local file sync ships as a signed macOS app, so a folder on a laptop can stay in step with knowledge.

Admins can cap spend per member, share individual subfolders, and sync directory groups instead of managing membership by hand.

Outlook meetings surface live in the workspace, and Slack became available as an MCP integration with output variable hints for workflows.

A new entity extraction endpoint returns graph-ready entities from documents for downstream systems.

Redbelt answers in Teams and Slack through first-party bots, with OAuth setup, threaded replies, and the same knowledge scoping as the web app.

A versioned v1 API shipped with streaming, system instructions, and API-key authentication, plus an SDK with citation utilities for building on top of it.

Calendar events sync with delta updates and distillation, so meeting context is available to agents.

Billing gained selectable plans, on-demand usage, per-organization budgets, and rate limits, with usage visible before the invoice arrives.

Write mode supports dictation while streaming, and members can set a default model, keyboard shortcuts, and export formats. macOS file sync gained a desktop OAuth flow and a download wizard.

Every model call writes to a usage ledger, and an admin dashboard reports spend by member, workspace, and time period at billing-grade accuracy.

A widget embeds Redbelt in any external page, scoped to selected folders, authenticated by public token or API key, with reCAPTCHA support and an inline display mode.

Confluence spaces sync into knowledge with delta sync, custom queries, and selection filters.

Monitored mailboxes route incoming mail into knowledge by rule, with mass upload and native previews for EML, MSG, and OFT files.

Files and folders can be moved, sorted server-side, and paged through at scale, while connector-managed folders are protected read-only so a sync cannot be corrupted from the interface.

Built-in rules redact sensitive material from model output, with the option to reprocess existing documents.

Write Mode pairs a canvas editor with personas, so long-form drafting happens beside the thread rather than inside it.

A Windows file sync connector keeps a local folder in step with knowledge, with sync previews, friendly names, and per-folder frequency control.

Word, Excel, and PowerPoint files preview in the browser without downloading them.

Workflows can be triggered from chat, run against manual templates with loop limits, and followed step by step on a timeline.

Model Context Protocol servers can be attached to agents, extending them with external tools.

Workflows became a first-class, database-backed resource with a visual builder, reworked steps, and a dedicated section in the navigation.

PDFs and text files preview in a drawer over the knowledge tree, with bulk selection and file management alongside it.

Connectors run on a schedule with explicit authorization, including Dropbox Business team accounts, improved SharePoint traversal, and reliable Google Drive token refresh.

Threads can be exported to Markdown or PDF, or saved straight back into a knowledge folder, and they rename themselves in the background.

Connector syncs only move what changed, with completion and error states reported per file.

A public API reference and agent documentation went live, along with the legal section of the site.

Organizations can issue and revoke API keys for programmatic access.

Agents and tools were restructured for clearer customization, with tagging and an additional model provider.

The first version of the Redbelt AI marketing site and documentation shipped.

Resources gained owners, roles, and sharing rules, making Redbelt usable by a team rather than an individual.

Stripe billing, WorkOS single sign-on, and machine-to-machine access arrived together.

Warehouse tables — including BigQuery — can be imported as data models for the SQL chat agent to query.

A theme system replaced hard-coded colors, and the admin surfaces gained mobile and tablet layouts.

The SQL chat agent answers questions against connected databases, guided by curated golden queries attached to the relevant files.

Canvases became downloadable and rateable, with cost and execution-time logs shown per run.

Large audio and video files are supported end to end, and the document distillation pipeline was substantially improved.

A full production deployment pipeline shipped with versioned database migrations.

The first three connectors shipped with recursive folder synchronization and background file processing.

Canvas arrived as a working surface next to the thread, with appearance settings and its own component set.

A code tool with an analyzer joined the toolset, and diagram generation became an agent of its own.

Connectors gained selective sync, so only the folders you choose are ingested.

Agents can generate Mermaid diagrams as part of an answer.

The knowledge interface was rebuilt, including mobile behavior, and new accounts now get a personal organization automatically.

Long-running work reports progress through a notification hub instead of leaving the page silent.

Uploaded files can be viewed in place, with batch upload and retry handling behind them.

Extraction schemas can be generated and saved as reusable templates.

Audio files are transcribed and distilled into knowledge alongside documents.

Agents can crawl and summarize web sources as part of a task.

Server-sent events stream responses end to end, and the first continuous deployment pipeline went live.

Complex Excel workbooks, HTML tables, and scanned text became extractable into structured output.

Responses stream while tools run, with the interface wired to the extraction schema.

The first build brought threads, folders, teams and organizations, a prompt library, and authentication together into a working application.

The document extraction and distillation pipeline that everything else is built on took its first shape.