The first-run wizard
The first time an administrator opens the dashboard, a full-screen setup wizard replaces the app. It walks through the essentials in order: setting your login password, setting the vault passphrase that protects your stored credentials, confirming how the deployment gets its AI models, creating a first agent, connecting a messaging app, and switching on web search. Every step deep-links to wherever the credential actually comes from, so you are not hunting for a settings page halfway through.
Each step has Skip for now, and on most deployments the whole tour can be skipped in one click - everything it configures is also reachable later from Providers, Vault, Agents and Connectors. The exception is a managed deployment provisioned with temporary credentials: there, the password and vault steps are mandatory, because those two secrets were generated for you and should not stay that way. Completion is recorded for the whole deployment, so no other administrator sees the tour again.
The wizard is shown to administrators only. If your account is a member rather than an admin, you land straight in the dashboard.
The sidebar, section by section
The left sidebar is split into two groups. A Primary group is always visible; a collapsible More group holds everything else. Items are drag-and-drop sortable and can be dragged between the two groups, and your arrangement is saved to your account rather than to the browser, so it follows you across devices. Admin is pinned at the bottom and is not draggable.
The Primary group
| Section | What it is for |
|---|---|
| Overview | The home screen. Agent and channel counts, system start/stop/restart controls, the Notifications list (a connector that could not start, a locked vault, disk pressure), the Approvals inbox, a daily highlights digest, a collapsible agents list, and a recent-activity feed. |
| Insights | Analytics over a 1, 7, 30 or 90 day window: model calls, tool calls, agent-to-agent traffic, spend, and a per-day work journal with an optional written summary of what actually got done. |
| Chat | Where you talk to your agents. Named conversations persist, several can run in parallel, and an in-flight reply keeps streaming while you look at another section. |
| Scheduler | Recurring and one-time tasks, built with a plain-language schedule picker rather than cron syntax. Shows run history per task. |
| Calendar | The same scheduled tasks as a day, week or month calendar - the visual companion to the Scheduler list. |
| Cortex | The improvement engine: its settings, its phases and their schedules, and the ideas, brainstorms, prototypes, projects and proposals it produces. |
| Files | A file browser over the selected agent's workspace, as an expandable tree. Preview, edit inline, upload, rename, move by drag and drop, and delete. |
Inside the More group
Agents is the one you will use most: a card per agent, and a full detail page behind each card. Connections holds one card per external service you sign into with OAuth. Knowledge manages the document and URL collections your agents search. Memory is a direct view of an agent's long-term memory file. Notes browses the agent's linked markdown wiki. Issues is a tracker where humans and agents file work for each other, deliberately shaped like GitHub issues so nobody has to learn a new tool. Agent to Agent shows conversations between your agents. Org Chart draws the whole operation - humans on top, agents below, with reporting and messaging links.
The rest are reference and operations surfaces: Tools browses every tool the platform can offer and whether its credentials are set, Providers lists AI model providers and their key status, Vault stores credentials, Revisions keeps automatic version history of your configuration and workspace markdown, Terminal gives shell access where your deployment allows it, Logs streams runtime log lines by level, and Config holds deployment-wide settings. Playground - an animated view of your agents at work - is off by default and appears only once an administrator turns it on. Admin appears for administrators when team accounts are enabled.
Press ⌘ K for the search palette. It is a global search, not just a nav filter - it matches agents and scheduled tasks by name and jumps straight to them.
Create your first agent
- Open Agents and press the + button. You are offered a source: Custom (start from nothing), Team or Expert (install a pre-built archetype), or Integration (an agent wired to one external service). The catalog carries 19 starter teams and 180+ specialist agents; if the lists look short, the archetype pack is still syncing.
- Fill in the identity. Name, description, an accent colour, and a model. On a managed deployment the model is pre-selected for you and billed in credits, so you can leave it alone on the first pass.
- Write the system prompt - or generate it. This is the agent's standing brief. The Generate with AI button drafts one from the name, description and tools you have chosen, which is a good starting point to edit rather than a finished answer.
- Save. The agent boots immediately and comes online. The sidebar agent picker and the agent list refresh on their own, and every open dashboard gets a toast announcing the new teammate.
You now have an agent that can hold a conversation and use its default tools. Everything else - connecting it to WhatsApp, giving it your documents, putting it on a schedule - is layered on afterwards, and each has its own chapter.
Some edits apply live and some need a restart. Prompt text, tools and subagents are picked up by a running agent within moments. Structural changes - the model, the environment, memory settings - take effect on the next Restart. Leaving an agent's page after editing a running agent pops a reminder offering to restart for you.
Your first conversation
- Pick the agent in the sidebar picker. It is shared across views, so the Files, Memory and Notes sections all follow your choice. Running agents are listed first; stopped ones sit in a collapsed Offline group.
- Open Chat and type. Type / for a grouped menu of slash commands and the agent's own skills. /help lists everything, /new starts a fresh conversation, /tools shows the agent's effective toolset.
- Watch the work, not just the answer. A plan renders as a live checklist that ticks itself off. Tool calls appear as activity lines. The detail chip in the header cycles Quiet, Normal and Verbose if you want less or more of that.
- Answer any approval cards. If the agent reaches for a tool that needs a human, an inline card appears with a risk badge and Accept once, Accept all and Reject. Nothing happens until you choose.
The header chips are worth learning early. Mode cycles Auto, Plan, Manual and Act - a per-conversation override of the agent's standing oversight setting. The model picker runs this one conversation on a different model. A context-window bar shows how full the conversation is, and cost chips show what it has cost so far. Attachments work by button, paste or drag-and-drop, and there is a microphone for dictating rather than typing.
Conversations are durable. The rail on the left lists them newest-first with a readable label, a live dot while one is streaming, and an unread badge for replies that arrived while you were elsewhere. Hover a row to rename or delete it. Reopening an older conversation restores its history and its running totals.
What to do next
Get one agent doing one job properly before adding a second. In practice that means: sharpen its instructions and tools (chapter 2), give it the documents it needs (chapter 3), and put it where the work arrives - a messaging channel or a schedule (chapter 5 and chapter 6). Only then is a second agent worth the coordination it costs.