Skip to content

Use the assistant

Ask questions about your workspace, get taken to the right screen, or let the assistant take an action — after you approve it. It works with your permissions, so two people can ask the same question and get different answers.

Where to find it:

  • Header → Ask AI (starts with a question about the page you’re on)
  • Keyboard shortcut ⌘⇧A (Ctrl+Shift+A on Windows and Linux) — toggles the side panel
  • Sidebar → AI Assistant — same panel toggle; open /assistant for the full-page view

Panel vs full page

Surface Best for
Side panel Keeping the current page visible while you ask (resizable dock on wide screens; sheet on smaller ones)
Full page (/assistant) Longer sessions when you want the conversation front and center

Both share the same conversations, models, and approvals.

Ask it something

Type a question and send it. Useful prompts:

  • “How do I connect Slack?”
  • “What does this toolbox share, and with whom?”
  • “Take me to my MCP servers”
  • “Why can’t I see the governance section?”

Mention toolboxes, connections, or connectors with @ when you want to point at a specific resource. Attach a file or screenshot when showing is easier than describing.

Slash commands

Type / in the composer for starters:

Command What it does
/new Start a new chat
/rename Rename the current chat
/explain Explain this page
/diagnose Diagnose connection issues
/review Review governance and access
/compose Compose or improve a toolbox
/compact Summarize older turns to free context

Approve what it does

When the assistant wants to change something, it shows an Approval needed card with exactly what it’s about to do. Nothing runs until you choose:

Assistant approval loop 
flowchart TD
  Ask["You ask the assistant"] --> Plan["Assistant proposes an action"]
  Plan --> Gate{"Needs approval?"}
  Gate -->|No| Run["Action runs"]
  Gate -->|Yes| Card["Approval needed card"]
  Card --> Once["Allow once"]
  Card --> Always["Always allow<br/>non-destructive only"]
  Card --> Deny["Deny"]
  Once --> Run
  Always --> Run
  Deny --> Stop["Action stopped"]
  Run --> Audit["Audited"]
Choice Effect
Allow once Run this action this time only
Always allow Remember approval for this kind of non-destructive action (only offered when the server marks it rememberable)
Deny Stop the action

Always allow is never offered for destructive operations. Remembered approvals are enforced server-side and audited. You still need the Approve assistant actions permission (assistant:approve) to clear gated writes; without it, those actions stay blocked.

Models and organization keys

If your organization has configured providers under Settings → Assistant, you can pick a model from the allowlisted catalog (grouped by provider). Providers without an organization key remain visible but disabled.

API keys are organization bring-your-own-key (BYOK) — administrators paste Anthropic, OpenRouter, Fireworks, or OpenAI-compatible gateway credentials in settings. Chat does not use personal keys in the browser. Availability depends on that configuration; see Configure the assistant.

A context meter shows how full the conversation is (green / amber / red). Use /compact or start a new chat when it fills up.

Permissions

Permission Label What it allows
assistant:use Use assistant Chat and run approved actions
assistant:approve Approve assistant actions Allow or deny write / sensitive actions
assistant:manage Manage assistant Configure providers, models, and assistant settings

Built-in Viewer can use the assistant; Member and above can approve. Managing providers needs Manage assistant (typically Org Admin / Owner).

Manage conversations

  • New chat when you change subject — context is per conversation
  • Conversation switcher for history (searchable, grouped by recency)
  • Rename or delete threads you no longer need
  • Edit the last message to truncate and retry from there when the UI offers it

Good to know

  • The assistant respects roles, shares, and restrictions the same way the rest of the app does.
  • After successful changes it may navigate, refresh lists, or open the right dialog — still within allowlisted UI commands.
  • Model choice and whether the assistant is available depend on Assistant settings.