Workspace & Tasks

Everything you do in Aura Workshop is a task: you type what you want, the agent picks the right tools, and you watch it work in a live, chronological transcript. There is no separate "chat mode" or "agent mode" — a single composer is the entry point for every interaction. This page covers the dashboard, the composer, the four task modes, the message thread, the context panel, smart routing, and voice.

Just installed? Start with Getting Started. When one agent isn't enough, reach for Multi-Agent Teams or Autonomous Agents.

Dashboard

The Dashboard is the home screen that greets you when you open Aura Workshop. It gives you a high-level view of your activity and quick pathways into new work. A randomized greeting sits at the top and rotates on each visit. If a required dependency is missing (Node.js, Python, Docker, and so on), a banner appears with a description and a direct link to Settings to resolve it; the banner disappears once the dependency is satisfied.

Quick action cards

Below the greeting, a grid of quick-action cards is grouped into categories (Software Dev, Marketing, Finance, Legal, Research, Design, Operations, Education). Each card carries a short description and a one-click prompt; clicking a card creates a new task pre-filled with that starting instruction and begins running it immediately.

Recent tasks

The body of the Dashboard shows a grid of recent task cards. Each card shows:

Title — the first line or an auto-generated summary of the prompt.
Status badge — a color-coded state: Executing (blue), Completed (green), Failed (red), Interrupted (yellow, resumable), Waiting (gray, queued), and Planning (purple, generating a plan before running).
Time ago — a relative timestamp of creation or last update.
Quick actions (on hover/tap) — Run/Resume to re-run a completed task or resume an interrupted one, Delete to remove it and its messages (with confirmation), and Fork to duplicate it with the same context and history so you can try a different approach.

When a task finishes while you are elsewhere in the app, a toast slides in with the task title and final status; it auto-dismisses after a few seconds, or you can click it to jump straight to the completed task. Multiple tasks can run at once, each with its own model, cancel token, and event stream — switching your default model does not affect tasks that are already running.

Task composer

A persistent composer bar sits at the bottom of every screen — the single entry point for all interactions with the agent. Its controls:

Text input — describe your task, ask a question, or send a follow-up. Enter sends; Shift+Enter inserts a newline.
Files (paperclip) — attach documents and images as context. Supported image formats become vision inputs for multimodal models; other files are read as text.
Folder — mount a project directory as the agent's working path; file reads, writes, and edits are scoped there.
Tools — the tool picker, grouped by MCP server, to enable or disable specific capabilities for this task.
Voice — records a spoken prompt (a pulsing dot indicates recording), transcribes it, and inserts the text.
Mode button — cycles Execute → Plan → Goal → Auto (covered below).
Role picker — assign a role skill (with search, category filter, and a "Default (no role)" option); the role's system prompt and tool permissions apply to the task.
Project selector — associate the new task with a project so its project-scoped memory and skills layer on.
Thinking level — when the selected model supports extended reasoning, choose Off / Low / Medium / High.
Model selector — pick from your configured models, grouped by provider; the choice persists as the default for subsequent tasks. See Models & Providers.
Send — the arrow/paper-plane button (or press Enter).

Type / in the composer for a slash-command typeahead (/run, /workflow, /schedule, and your custom commands) — see Automation & Workflows.

Starting a task

Open the Dashboard and type what you want in the composer.
Optionally attach Files or mount a Folder as the working directory.
Optionally open Tools to enable or disable specific tools for this task.
Optionally pick a project, a role, a thinking level, or a different model.
Cycle the mode button if you want something other than Execute.
Press Enter or the Send button.

The transcript streams in strict chronological order — text, tool calls, tool results, generated files, and charts. Attachments become file references the agent can read; pasted images are saved to disk and handed to the agent's vision tools.

From the daemon viewer or the API

On a headless server, open http://your-daemon:18800/, paste your AURA_WEB_TOKEN, and use the Chat tab — the same agent over a browser transport. Programmatically, POST /api/tasks creates a task and POST /api/tasks/{id}/run streams events over SSE:

curl -X POST https://aura.example.com/api/tasks \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Build a dashboard",
    "description": "Build a KPI dashboard for the marketing team",
    "task_mode": "plan",
    "project_path": "/path/to/project",
    "auto_run": true
  }'

Full surface in the API Reference.

Task header bar

Inside an active or completed task, a header bar shows:

Back to the Dashboard or task list, and the task title.
Status badge (same states as the recent-task cards) and, for tasks that came from a listener, an IM platform badge with the platform icon and channel name.
A running timer while the task executes.
Action buttons when the task is idle — Fork, Export (the conversation as a styled page you can print to PDF), Show/Hide tool-call details, and the context panel toggle.
A Files Created button with a count badge that opens a file browser of everything the task produced (view, download, open).

The four task modes

The composer's mode button cycles Execute → Plan → Goal → Auto. The chosen mode applies to the next new task only; follow-ups to an existing task keep that task's own mode. What changes between modes is how much autonomy the agent has and when it checks in with you.

Execute (default)

A straight agent loop with the full toolset. The agent works until it is done or hits the turn limit. Use it for well-scoped work you can describe in a sentence or two.

Plan

The agent researches with read-only tools first, asks clarifying questions as plain markdown, then produces a numbered plan for approval — nothing is written until you approve. The questions and plan render as interactive cards inline in the chat; approving unlocks the full toolset and runs the plan. Good for anything ambiguous, risky, or spanning many files.

Goal

You state a target. The agent extracts success criteria + deliverables, runs, then an LLM verifier checks the result against those criteria. If the goal is not achieved and budget remains, the agent gets a corrective nudge and loops again. Tune with goal_max_iterations (default 5, range 1–50), an optional goal_verifier_model override, and goal_auto_approve to skip the criteria review. Use Goal for "make it actually correct" work, e.g. "all callers updated, tests green, zero type errors."

Auto (autonomous loop)

The agent runs one cycle, decides its own next-wake time (within guardrails), and either stops itself (goal kind) or keeps watching (monitor kind). This is the composer's quick entry point to autonomous agents — Auto tasks show live loop-status cards in the transcript and survive app restarts because each next wake is armed as a scheduled row. The full managed panel, cadence limits, and runtimes are covered in Autonomous Agents.

Beyond a single agent. To put several models or peers on one problem together — Discussion, Plan, Goal, and parallel Research fan-out — use the Multi-Agent Teams tab.

Stopping and resuming

Stop any task with the Stop button in the chat (or Stop next to Send in the viewer, or POST /api/inference/stop with a task_id). Stop aborts the stream, kills running tool processes, and marks the task interrupted; resume it later with POST /api/tasks/{id}/resume.

Task workspace & message thread

The main area of a task is a scrollable message thread showing the entire conversation in exact chronological order — no grouping, no collapsing of the timeline.

User messages appear as right-aligned bubbles, with file and folder attachments shown as removable chips below the text.
Assistant messages stream progressively as rich Markdown — syntax- highlighted code, LaTeX math (KaTeX), Mermaid diagrams, and tables.
Thinking blocks from extended-reasoning models appear as collapsible sections, collapsed by default.
Tool execution cards — each tool call is a collapsible card showing the tool name, an icon, and a status (spinner while running, OK on success, ERR on failure); expand to see the full input parameters and output.
Role dividers mark handoffs in multi-agent tasks, with a + Memory button to save that role's work.
Routing badges show which model served each turn when smart routing is active, and error blocks surface tool, API, or provider failures inline in red. A Load Earlier button fetches older messages for very long histories.

Activity indicator

While a task executes, a fixed bar just above the composer shows a spinner and live operation text describing exactly what the agent is doing — the tool currently running and its progress, a preview of the agent's reasoning, text-generation progress, or, for fan-out, a count like "4 file(s) created by agents." It updates in real time as the agent moves between operations.

Follow-up suggestions

After a task completes, a collapsible bar offers contextually relevant follow-up chips (with a count badge). Clicking a chip drops its text into the composer, ready to edit or send.

Response actions

Copy Conversation — copies the whole thread as formatted text.
Per-code-block Copy — a copy button on every code block.
Per-diagram Save as PNG — export any Mermaid diagram as an image.
Per-table Copy as CSV — export a rendered table to a spreadsheet.
Export (task header) — open the conversation as a styled page and print to PDF, including text, code, tool summaries, and images.

A Scroll to bottom button appears when you scroll up in a long thread, and the Files Created button opens the file browser of the task's outputs.

Context panel

The context panel is a collapsible right sidebar with detail about the current task. Toggle it with the hamburger icon in the task header; drag its left edge to resize.

Workflow progress — for multi-agent team tasks, a vertical stepper with a progress badge (e.g. "3/6") and per-role status dots: pending (gray), active (pulsing blue), done (green). Roles running on remote deployments show a Remote badge.
Parallel agents — when three or more agents run via fan-out, a section shows a count (e.g. "5/8 completed") and a card per agent with its index, status, and tool-call count.
Thinking content — a collapsible, maximizable view of the model's raw reasoning when extended thinking is on.

Usage stats

The panel footer shows live usage for the current task:

Metric	Description
Context usage	A color-coded bar for how much of the model's context window is consumed (green → yellow at 50% → red at 80%+).
Input / output tokens	Tokens sent and generated, formatted with K/M suffixes.
Cache read	Tokens served from the provider's prompt cache, when available.
Total cost	Cumulative dollar cost for the task, from the model pricing table.
Latency	Time-to-first-token for the most recent API call.
Model / provider	The model and the provider serving it for this task.

The panel's data persists when you navigate away and back. Context management (compaction at the model's window limit) happens automatically as the conversation grows.

Task classification & smart routing

Every message is analyzed by a fast one-shot classifier that decides how to handle it — no manual mode switching required. The classifier also drives which agent tools are made available for the next turn (categories only ever grow within a task, never shrink).

Category	What happens
Single-agent task	One agent runs with the tools it needs — the default for most requests one agent can finish end-to-end.
Team task	Routes to a multi-agent team, selecting a matching team or assembling roles on the fly. See Multi-Agent Teams.
Scheduled task	A recurring phrasing ("every Monday at 9am") creates a schedule automatically. See Automation.
Clarification needed	The agent asks follow-up questions before proceeding.

Smart routing

If you pick Aura Local Routing as your model, Aura scores each request and picks the most cost-effective tier that can handle it, then resolves a concrete model from your routing config; the status bar shows which model it chose. A pinned model is always used exactly as selected — provider or API errors surface in the chat rather than silently switching models. Per-provider spend caps and a fallback order (walked only for routing and cap enforcement) are configured in Settings → Billing. Full details in Models & Providers.

Voice

The Voice tab is a spoken conversation with the agent. The picker lists voices (not chat models); it is multilingual and theme-aware. Each conversation is saved as a task, so its transcript shows up in your history and reopens in the Voice tab. Under the hood it prefers realtime speech-to-speech and falls back to chained transcription + text-to-speech.

You can also speak individual prompts from any composer with the Voice button — a pulsing dot indicates recording, and the audio is transcribed into the text input. Speech-to-text and text-to-speech providers, voices, and speech rate are configured in Settings.

Where to go next

Projects — group tasks, files, and goals into a scoped workspace.
Multi-Agent Teams — many models and peers on one problem.
Autonomous Agents — self-paced loops that stop themselves.
Models & Providers — connect providers and tune routing.
Memory & Knowledge — automatic recall across tasks.