Aselius Workspace
Part of the Aselius Platform · Created by Beunec Technologies Inc.
Version: 1.0.0
Aselius Workspace is where researchers, engineers, founders, and startup teams run real-world tasks: research, drafting, coding prototypes, creative assets, and governed collaboration, with outputs you can preview, edit, download, and share. It is built for software engineering teams, research organizations, consulting firms, startup teams, university post-graduates who need production-grade deliverables, not just chat agent replies.
Access it from the Aselius Platform sidebar under Workspace.
What Workspace Is
Workspace is not a single chat box. It is an agentic task environment built around:
| Area | What it does |
|---|---|
| Taskspace | Your primary work surface, conversation, attachments, engines, previews |
| Governance | Workspace-level AI behavior and content rules |
| Connector | Secure OAuth connections to external apps |
| Credits | Prepaid balance for agentic engine usage |
| Shared with me | Accept invitations and open workspaces shared with you |
At the center is Taskspace (internally ChatSpace): one conversation thread per task, combining a conversation area on the left, your prompts, agent responses, and clickable snapshots of completed work, with a right-side Preview Panel for assets, attachments, sandbox iframe, documents, and HTML.
High-Level Architecture
How to read this: You work in Taskspace. Your message is enriched with hidden context (date, engine choice, agent notes, attachments). The Rivine orchestration runtime routes and executes the task. Deliverables land in Assets and as snapshots in the conversation. The Preview Panel lets you inspect and edit without leaving the browser.
Quick summary: how a turn works
You can pin an engine from the Space UI (skip automatic routing) or let the Mediator choose. You can also set a specialty (e.g. PDF vs spreadsheet for Research, or a legal domain for Cicero) to steer output without re-explaining context every turn.
11. Taskspace
Taskspace is where tasks are executed, previewed, and continued over time.
What you can do
- Send natural-language requests to Rivine agentic engines
- Attach files, paste text, parse URLs, capture screenshots
- Pick an engine and specialty (sub-mode) for the kind of work you need
- Use Agent Notes to jump directly to a pre-configured engine + specialty
- Stream AI responses in real time, including tool progress whenever the engine uses external data or sandbox actions
- Open deliverables from snapshot cards in the conversation
- Approve or reject high-risk actions via Human-in-the-Loop (HITL) when the engine pauses for your decision
- Search across messages, taskspaces, assets, and workspaces from the workspace header
Conversation layout (accessibility)
Older messages use flip cards to reduce visual clutter:
- Your previous requests collapse to "Click to view your request"
- Older AI responses (beyond the latest two) collapse similarly until you expand them
- Attachment and file snapshots stay visible as tappable cards on expanded messages
This flip-card behavior applies to both sides of the exchange, your original prompt and the AI response, so you can review long agent output without losing your original ask.
Snapshots
When an engine finishes meaningful work, a snapshot card appears in the conversation. Click it to open the deliverable in the Preview Panel.
| Output type | Snapshot behavior |
|---|---|
| Markdown document | Opens in the built-in editor (rendered + editable) |
| HTML report / deck | Renders in the HTML preview; export when available |
| Image / video (Creative) | Opens media preview from stored asset URL |
| Code (Rivine Code) | Opens sandbox preview, see Rivine Code section below |
Roles and permissions
| Role | Typical access |
|---|---|
| Owner | Full control, share, transfer ownership |
| Operator | Edit governance, share, manage taskspaces |
| Member | Create and use taskspaces |
| Guest | View-only, cannot edit documents or governance |
22. Rivine Agentic Engines
Rivine is Beunec's proprietary agentic orchestration runtime. Each Rivine engine is a runtime construct, orchestrator, tools, verification, and sandbox policy, not a single generic model call. The runtime handles planning, routing, tool use, verification, and delivery; language models operate inside that harness.
We do not disclose specific third-party model IDs, routing tables, or internal parsing strategies. What matters for users: engines are specialized, governed, and produce inspectable outputs.
Available engines (today)
| Engine | Best for | Deliverable | Minimum credits to select |
|---|---|---|---|
| Auto | Platform routes to the best engine | - | None |
| Rivine Chat | Professional, action-oriented dialogue | Markdown or inline HTML | $2.00 |
| Rivine Research | Source-grounded research and analysis | Markdown → PDF or spreadsheet HTML | $10.00 |
| Rivine Code | Software engineering, debugging, security review | Sandbox code + live preview | $20.00 |
| Rivine Cicero | Legal drafting, review, and policy work | Markdown legal memo | $5.00 |
| Rivine Creative | Image, video, and audio creation/editing | PNG / video media | $5.00 |
| Rivine FIDA | Financial research and quantitative analysis | Not available yet | Coming soon |
Not available yet: Rivine FIDA, Rivine Agentic Engines as an API, and several connector/import integrations (see below).
Which engine should I pick?
Specialty selections
Most engines require a specialty before you send a message. Specialties activate a focused sub-agent mode with its own operating instructions.
| Engine | Specialties |
|---|---|
| Rivine Chat | Chat |
| Rivine Research | PDF, Spreadsheet |
| Rivine Code | NextJS, Python, NextJS + Python |
| Rivine Cicero | Contract Review, Legal Drafting, Immigration Law, International Relations, Corporate Law, Employment Law, Policy and Law Drafting |
| Rivine Creative | Image Generation, Image Editing, Video Generation, Video Editing, Audio Generation |
If a specialty is required and not selected, send is blocked until you choose one. Set specialty in the Space UI or describe it in your prompt, the Mediator can also infer it from context.
Grounded data (when enabled)
For Chat, Research, Creative (briefing), and Cicero, you can enable grounded sources, finance, health literature, SEC filings, maps, web search, and more. Agents are instructed to cite sources rather than rely on memory alone. Source toggles are per-workspace; disabled sources are never called.
Agent Notes
Agent Notes are pre-written harness instructions for each specialty, they tell the runtime how to execute (evidence rules, escalation triggers, output shape), not just what to answer.
- In Auto mode, Agent Notes list every specialty across engines; picking one auto-selects the matching engine + specialty.
- On send, the selected note is injected invisibly into your message (you do not see it in the chat bubble).
- Notes enforce behaviors like: cite sources, never invent spreadsheet cells, escalate legal risk to HITL, block unsafe creative requests, etc.
Separately, Agent Notes can also surface after a run as transparency metadata in the conversation: the routing decision, which engine handled the task, verification or escalation flags, and citation summaries. This is not a separate agent, it exists to help you audit what happened, not to replace the deliverable.
How engine execution works (simplified)
Rivine Code is designed as an autonomous coding runtime build, debug, audit, and preview software without opening a local IDE. For non-technical users, that means seeing a working preview in the browser. For engineers, it means a quick prototype you can download and carry into production tooling.
Current behavior in Workspace (transparent):
- Engine outputs appear as snapshots in the conversation (markdown, HTML blocks, generated images).
- Markdown and HTML open in the Preview Panel (editable document editor or HTML live preview).
- Most deliverables are saved to Assets.
- Rivine Code project file trees run in an isolated sandbox with preview capability; those sandbox filesystems are not persisted to Assets today (planned for a future release), only the snapshot/preview bridges the two in the conversation.
- The HTML tab provides a sandboxed iframe preview for HTML content (code editor + live render).
Human-in-the-Loop (HITL)
When the runtime hits verification limits, low confidence, or high-risk actions, it can pause and show an approval request in Taskspace. You can approve or reject before execution continues. This is especially relevant for legal, security, and data-sensitive workflows.
33. Mediator Modes
Mediators shape how your intent is interpreted before routing.
| Mode | Purpose | When to use | Plan availability |
|---|---|---|---|
| Alpha | Optimizes intent for global regulatory compliance | Fast routing; good for clear, single-intent requests | All plans |
| Bravo | Applies the Agentic Annotation Protocol | Default; more careful intent analysis before routing | Pro, Essential |
| Charlie | Optimizes intent and awaits approval before routing | Revision pass, useful when routing confidence is low or context is ambiguous | Essential |
Transparency note: Mediator selection is available in the Taskspace UI and gated by subscription plan. Full end-to-end wiring from UI selection into every stream request should be verified against your deployment; the frontend currently stores mediator choice in the interface while the primary stream payload focuses on message, attachments, sources, and file context.
If you select an engine explicitly in the Space, routing can skip the Mediator entirely and go straight to that engine.
44. Preview Panel (right side)
The Preview Panel is your inspect-and-edit surface without leaving Workspace.
Tabs
| Tab | Purpose |
|---|---|
| Assets | All taskspace files, filter by user vs AI, type, download, delete, open |
| Attachments | Pending items before you send a message |
| Preview | Markdown documents, images, plain text |
| HTML | HTML code editor + sandboxed live preview |
| Share | Share workflow (when opened from share actions) |
Markdown documents
- Rendered and editable in the built-in document editor
- Table of contents in fullscreen
- Word count
- Export: Markdown, PDF (print), DOCX
- Save updates the asset (guests: read-only)
- Follow-up prompts can request surgical revisions (targeted patches) instead of full regeneration, you don't need to ask the engine to redo the whole document to fix one section
HTML
- Split view: code editor + sandboxed iframe preview
- Export: HTML file, PDF (print)
- AI responses that include HTML code blocks appear as HTML snapshots in the conversation, click to open the HTML tab
- Research PDF-style reports and spreadsheet layouts preview as self-contained HTML, suitable for review, print, or export workflows supported by your plan
Panel behavior
- Resizable on desktop (width remembered)
- Full-screen sheet on mobile
- Optional fullscreen mode for focused editing
55. Assets
Assets are the durable output library for each Taskspace.
| Property | Details |
|---|---|
| Types | Document, code, image, data, other |
| Creator | User or AI |
| Typical flow | Upload/parse → markdown asset, or AI generates → saved asset |
| In AI context | Taskspace assets are included as current_file_context on each message so engines stay grounded in your files |
By source
| Source | Examples |
|---|---|
| You | Uploads (documents, images, references) |
| AI | Generated markdown, HTML, media from Chat, Research, Cicero, Creative |
Assets are versioned. AI-created assets are tagged with engine metadata and citations where applicable.
Snapshots vs Assets
- Snapshots in chat are quick-open cards tied to message content (pasted text, file references, HTML blocks, images).
- Assets are persisted files you manage in the Assets tab.
Rivine Code's sandbox filesystem is separate from Assets for now, only the snapshot/preview bridges the two in the conversation.
66. Attachments & file parsing
Supported upload types
Uploaded files are converted to a markdown-first representation for AI grounding (exact parser engines are proprietary).
| Category | Extensions |
|---|---|
| Images | jpg, jpeg, png, gif, bmp, svg, webp |
| Documents | pdf, docx, xlsx, pptx, csv, txt, rtf, md |
| Audio | mp3, wav, m4a, flac, ogg, mpeg |
| Video | mp4, mov, avi, mkv, webm, mpeg |
| Code | py, js, jsx, ts, tsx, html, css, json, xml, sql, java, tex, vue, svelte, go, rs, rb, php, c, cpp, h, cs, swift, kt, yaml, yml, scss, less, mdx |
Limits
| Limit | Value |
|---|---|
| Max file size | 200 MB |
| Max files per message | 10 |
| Max files per batch upload | 10 |
| Daily uploads | 3 (Base) / 10 (Pro) / 20 (Essential) |
Not supported
- Archives (.zip, .rar, etc.)
- Executables and other blocked extensions for security
Import methods
| Method | Status |
|---|---|
| File upload | Available |
| Paste text | Available (no server parse) |
| Paste / clipboard image | Available |
| Web page or YouTube URL parse | Available |
| Screenshot capture | Available |
| Import from Manage Files | Available |
| Google Drive / OneDrive / Dropbox attach | Coming soon |
What is AI-powered here
- Parsing runs on Beunec's server-side pipeline (not a chat completion you type into).
- Parsed content becomes markdown assets the engines can reason over.
77. Hidden context injection (agentic awareness)
To activate the full Rivine harness without cluttering your chat, Workspace injects structured tags into outbound messages. These are hidden from the visible conversation but used by the runtime for classification, routing, and output quality.
| Tag | Purpose |
|---|---|
<context_identity> | Your local date, time, timezone, and location (if browser permission granted) |
<agentic_engine> | Explicit engine override when not using Auto |
<user_propritize> | Specialty / sub-agent mode selection |
<agent_note> | Full specialty harness instructions |
<pasted_text> | Pasted content attachments |
<file_attachment> | References to parsed asset files |
| Language hint | Response language preference (non-English) |
This is how Workspace gives agents situational awareness (when/where you are working) and execution discipline (which engine, which specialty, which rules) without you managing prompts manually.
88. Governance
Route: Workspace → Governance tab
AI-powered? The page itself is configuration UI. Saved rules inform AI behavior on the backend, it is not a chat feature.
What you configure
System instructions
- Primary role
- Specified industry
- Behavioral guidelines
- Response format
Content restrictions
- Prohibited topics
- Content filters
Who can edit: Owners and Operators. Guests see read-only fields.
99. Connector
Route: Workspace → Connector (dedicated page)
AI-powered? No, this is credential and permission management so engines can access external systems when authorized.
Connectable today
| App | Access |
|---|---|
| Google (Gmail/Calendar) | Mail/calendar read & send |
| Microsoft 365 | Mail/calendar |
| Supabase | Read/edit |
| Firebase | Read/edit |
| MongoDB | Read/edit |
| Stripe | Read balances/transactions |
| GitHub | Read repositories |
Not yet implemented
- Dropbox
- Box
Connect Apps toggles in the Taskspace input area list additional sources (web, maps, health, etc.) but are currently disabled in the UI, connecting apps is done through the Connector page first.
1010. Credits
Route: /credits (linked from the balance widget in Workspace header)
AI-powered? No, this is prepaid usage billing.
- Credits are consumed when you run agentic engines (usage-based after free monthly allocation).
- Each Rivine engine has a minimum balance required to select it (see engine table above).
- Low balance is highlighted when under $5.00.
- Credits can be shared via CloudHandle (per plan features).
- Pay-per-token usage is metered per call at the runtime level (see "Under the hood" below).
1111. Shared with me
Route: Platform sidebar → Shared with me
AI-powered? No, this is collaboration and access control.
- View pending invitations and active shared workspaces
- Accept or decline invitations
- Open a shared workspace (lands in its first Taskspace)
- Roles: Owner, Operator, Member, Guest
Workspace owners/operators can share from the Preview Panel Share tab (when enabled).
Under the hood (what we share publicly)
We do not publish exact model names, routing tables, or parsing strategies, that is proprietary Beunec infrastructure. Here is what matters for trust and quality.
1. Orchestration runtime (RE-AI graph)
A graph-based runtime coordinates each turn:
- State persists across steps and supports human-in-the-loop pauses (approval, clarification, escalation)
- Long tasks can resume without losing thread context
- Pay-per-token usage is metered per call
2. Reasoning framework (REA / REI / REAI)
Engines use one of three tool-calling reasoning patterns, chosen by task type:
| Pattern | Orientation | Typical engines |
|---|---|---|
| REA | Analysis-first | Research, Cicero |
| REI | Implementation-first | Code |
| REAI | Analysis + implementation | Chat |
Each loop: plan → call tools → verify → terminate with a structured answer. Models are nudged to finish via a final_answer contract rather than trailing prose.
3. Harness engineering ("brain and hands")
Beunec separates decision-making from execution:
- The orchestrator plans and delegates
- Sandboxes and tools run in isolated environments (E2B for code; scoped tool access for research/legal)
- If execution fails, the harness can retry or adapt without wiping conversation memory, similar in spirit to resilient agent harnesses in the industry, tuned for Rivine's multi-engine layout
For Research and legal work, downstream formatters (PDF/Sheets) do not re-run open web research, they work from the lead agent's grounded output, controlling cost and drift.
4. Verification
- Deterministic checks where possible (schemas, tests, linters on the Code cold path)
- Inferential self-checks via agent instructions and escalation tags
- Grounded citations for factual engines
- Repeated failure or high-risk actions can escalate to you (HITL)
5. Agentic Annotation Protocol
Every agent, sub-agent, and tool-facing model receives a standardized instruction layer: capability boundaries, when to call tools, when to escalate, and positive/negative behavioral examples. This is the same family of ideas as the GCAP/ASPS annotation approach, consistent behavior across the stack, not ad-hoc prompts per feature.
6. Security posture (summary)
- Workspace-scoped assets and authenticated API access
- Sandbox isolation per engine/thread
- Credentials scoped to connectors you authorize, not globally exposed to agents
- PII handling and output guardrails at the infrastructure layer
What is AI-powered vs not (summary)
| Feature | AI-powered? |
|---|---|
| Taskspace chat & streaming | Yes |
| Rivine engines + specialties | Yes |
| File/URL parsing to markdown | Server pipeline (proprietary; not user-facing LLM chat) |
| Preview Panel editors | No client-side edit/export |
| Assets library CRUD | No standard API |
| Governance | Config only shapes AI via backend |
| Connector OAuth | No |
| Credits & billing | No |
| Shared with me | No |
| Workspace search | Depends on implementation; not generative chat |
| Voice input | Browser speech-to-text → your message text |
| Mediator UI | Intent shaping layer; plan-gated |
Detailed breakdown
AI-powered today
- Mediator routing
- All Rivine agentic engines and their sub-agents
- Grounded research and citation assembly
- Surgical document/code revisions via follow-up prompts
- Creative media generation pipelines
- Rivine Code sandbox execution and cold-path delivery (APST, coder, CI/CD steps where enabled)
Not AI-powered today (may evolve)
- Built-in document editor: Manual edits in the preview panel are your changes. The editor does not silently rewrite content unless you ask an engine to revise it.
- Workspace administration: Members, roles, permissions, billing, and connector OAuth setup are standard platform controls, not agent-driven decisions.
We will disclose when additional surfaces become AI-assisted.
How Aselius Workspace compares
We will not claim features we have not shipped. Here is an honest positioning lens for researchers, engineers, and founders evaluating agentic tools:
| Dimension | Typical copilot / workspace products | Aselius Workspace |
|---|---|---|
| Architecture | Often a chat layer over one model | Rivine runtime with specialized engines, specialties, verification, and HITL |
| Output handling | Copy-paste from chat | Assets + Preview Panel edit markdown/HTML, export, reopen from snapshots |
| Task continuity | Session-based chat | Taskspaces with persistent assets and governed configuration |
| Enterprise control | Tenant policies vary | Governance per workspace + role-based sharing |
| Coding | IDE assistant or codegen | Rivine Code as runtime construct with in-browser preview (not "just autocomplete") |
| Transparency | Opaque routing | Visible engine/specialty choice, agent notes, HITL checkpoints |
We are not claiming to replace your entire IDE, legal counsel, or brokerage stack. We are offering a unified place to run governed agentic work and leave with deliverables, documents, code previews, research outputs, and creative assets, without stitching together five separate tools.
Usage tips
- Pick the right Space engine: or let Bravo Mediator route, then lock the engine if it chose correctly.
- Enable only the grounded sources you need: fewer sources, faster turns, lower cost.
- Use follow-ups for surgical edits: "patch section 3", "fix the chart labels", instead of regenerating entire documents.
- Rivine Code: describe stack (Next.js, dashboard, landing page); use preview to validate before download.
- Rivine Research: say PDF or spreadsheet upfront, or set the specialty toggle.
- Cicero: treat output as legal information, not legal advice: escalation paths exist for high-stakes matters.
Recommended workflows
Researcher
Engineer / founder
Legal / policy operator
Team lead
Roadmap (disclosed only)
Available now: Rivine Chat, Research, Code, Creative, Cicero; Mediator modes; grounded sources; asset previews; Code sandbox preview; pay-per-token metering.
In progress / not yet production:
| Item | Status |
|---|---|
| Rivine FIDA (financial intelligence & backtesting) | Coming soon |
| Rivine Agentic Engines as an API | Coming soon |
| Google Drive / OneDrive / Dropbox attach in Taskspace | Coming soon |
| Dropbox / Box Connector | Not yet implemented |
| Rivine Code sandbox filesystem → Assets persistence | Future |
Beunec will disclose additional capabilities when they are ready for production use.
Legal & product notes
© 2026 Beunec Technologies, Inc. All rights reserved.
- Aselius Platform is created by Beunec Technologies Inc.
- Workspace is one application within the platform, selectable from the sidebar.
- Rivine, Reverse Engineering of Analysis & Implementation, and related orchestrations are proprietary, not open source. Documentation describes capabilities and outcomes, not internal implementation recipes.
- Rivine Cicero and similar engines are not substitutes for licensed attorneys, physicians, or financial advisors.
- Agentic Annotation Protocol principles (capability transparency, refusal of harmful requests, no fabricated real-time data) are embedded in engine harness design.
- AI outputs require human review for high-stakes decisions (legal, medical, financial, security). Cicero and Research outputs are assistive, not substitutes for licensed professionals.