Aya, the AI assistant

Aya is AY Inbox's AI assistant: a chat panel in the dashboard that reads across your contacts, campaigns, mailboxes, unified inbox, CRM, team, and settings, and takes almost any action you could take yourself. Answers stream word by word and each tool step appears as it starts.

Open Aya from its face in the bottom-right corner of the overview pages (see Home and navigation) or Cmd/Ctrl + I anywhere. The face shows when Aya is working or waiting for your approval, and its badge pulses when it has an answer you have not read. The panel follows you between pages.

The same assistant also answers in Slack, when you mention @ayinbox or message it. See Slack.

Aya needs a paid plan; a free trial has no credit allowance and the AI surfaces stay locked. Beyond that, access is the Use AI role permission (on for admins and managers, off for viewers); without it Aya and AI drafting are hidden. Admins can cap per-member spend by day, week, or month. See AI credits and Team and roles.

Conversations are private to you by default. An admin can enable Shared history under Settings > Workspace > Aya, after which every member with Use AI sees, opens, continues, and deletes any conversation, attributed by owner. Turning it on exposes existing conversations, so decide deliberately. Delete individually from the history rail, or Clear history to remove every conversation you own.

In a teammate's conversation, the results of tools your own role cannot use are withheld, from you and from Aya when you continue it; you see that a step ran, not what it returned. Secrets are never kept in any conversation: an invitation token or a webhook signing secret is shown once, to the member who asked for it, in the dashboard step that produced it. Aya itself never sees them, and Slack never shows them.

Conversations run as tabs: open several with +, and a run keeps streaming in its own tab while you read another. Expanding gives a full workspace with searchable history grouped by day.

Docked, the panel resizes by dragging its inner edge, which also takes the keyboard once focused: arrow keys nudge it (hold Shift for a bigger step), Home and End go to the narrowest and widest the window allows, and Enter or a double-click restores the default. On desktop Aya opens as a floating window you can drag by its header and resize from any edge; it remembers position and size. Dock it to either screen edge with the dock button or Alt + P, and pop it out again the same way or by dragging the header away.

Work continues without the panel. Minimizing docks it to a status bar showing live state: working with how long the run has taken, waiting on approval, or holding an unread response. Closing it entirely also keeps runs going, and Aya's face in the bottom-right corner of the overview pages shows the state: it thinks while Aya works, asks for attention when an approval waits, celebrates when a run finishes, and its badge pulses while a response is unread.

Aya reads what the Advisor measures in your workspace and suggests fixes for it, never generic prompts. An empty conversation lists them under Suggested for you, one line per kind of problem ("3 new mailboxes are sending at full volume"). Choosing one asks Aya to fix it: it checks the current state, proposes the smallest change, and shows each one as an approval card first.

Now and then Aya also offers one of these in a small bubble that rises above its face, with Fix with Aya and Not now. Tips are deliberately rare so they stay worth reading:

• only problems the Advisor rates Needs attention or Critical
• at most two a day, at least 30 minutes apart, never in your first moments on a page
• only once you pause: not while you type or scroll, not with a dialog open, and never while the Aya panel is open
• each kind of problem once, unless it gets worse, and one on the page you are on comes first
• dismissing two in a row quiets them for three days, and the bell in the bubble pauses them for a week

These preferences are yours alone, kept in your browser for each workspace.

• Find, read, create, and delete contacts, edit fields, labels, subscription, and notes, and bulk-edit
• Read a contact's activity timeline and the emails you sent them
• List campaigns and stats, and list leads by status, which is how it finds leads that went cold
• Create, edit, or delete campaigns, manage sender mailboxes and tracking domains, read logs, and edit sequence steps
• Start or stop a campaign (always asks)
• Read analytics, list mailboxes, and manage one: sending limits, warmup, tracking domain, warmup-ban appeals, disconnection
• Read the unified inbox, draft a reply, or send a reply or new email (each send shows the full message and asks). A drafted reply is checked against the campaign brain and comes back with any flags
• Read a campaign brain's public facts with get_brain (by brain, campaign or thread), and edit brains, their material and which campaign uses which (internal rules and internal material are read only with get_brain_internal; it and every brain edit except creating one always ask, even when writes are allowed without asking, because they change what drafts tell prospects)
• Snooze threads, set labels, mark read, cancel a queued send
• Run the CRM end to end: deals, tasks, notes, pipelines, stages
• Read and manage automations, including enabling one
• Manage the team, workspace and voice settings, plan and limits, API keys, and webhooks
• List connected integrations, give you a link to connect one, and set up Slack: notification channels, the inbox channel, and where the assistant answers
• Search the web and read a public page
• When the operator turns on Looot research, read a company's website, its recent news and its LinkedIn company posts (see AI contact research)
• Find leads at company domains with a Looot lead job when the instance has them on (always asks), follow its progress, and check one address through the workspace's verifier

It knows what page you are on, so "summarize this campaign" and "draft a reply here" resolve correctly.

When the available tools exceed the OpenAI-compatible provider's 128-tool limit, Aya discovers the tools it needs on demand, including tools from connected MCP servers. Discovery results include each tool's argument schema, so its usage remains available in the conversation after another search. No permitted capability is removed. Discovery is a read-only step; the action itself keeps its normal permissions and approval requirements. Like other model steps, discovery counts toward the run's step budget and AI credits.

Reading runs automatically. Changing always asks, showing an approval card with the action and every argument Aya passed (expand All arguments to read them). You can Approve that one action or Skip it. A member with Manage settings can also choose Always allow, which lets that kind of action run without asking for everyone in the workspace from then on; other members approve one action at a time.

Some actions always ask, and Always allow is never offered for them:

• sending a reply or a new email, and placement tests
• starting or stopping a campaign, enabling an automation, starting warmup, releasing a mailbox's send hold or a lead's hold, and adding leads to a campaign, since each puts mail in motion
• inviting a member, changing a member's role, and creating or editing an API key
• starting a Looot lead job, which spends the instance's Looot balance and can add leads to a campaign

Inviting or removing a member, changing a member's role, creating or editing an API key or a webhook endpoint, and rotating a webhook signing secret also need the same recent confirmation as doing it in Settings: if you have not confirmed it is you in the last five minutes, Aya stops and asks you to confirm in the dashboard first. From Slack, do these in the dashboard.

• tools from connected MCP servers

A send's card shows the sending mailbox, the recipient, the subject, and the plain and HTML bodies, and only that exact message goes out, from that mailbox, to that recipient. If the conversation gets a newer message before you approve a reply, the reply is refused and Aya asks again. It refuses suppressed addresses.

Each approval decides one specific action: approving a card that is no longer waiting (because it was already decided, in another tab or in Slack) does nothing.

Settings > Workspace > Aya lists every action that is always allowed, who allowed it and when. A member with Manage settings can Revoke one, after which it asks again.

Drafts stay drafts. Campaigns Aya builds are never started and automations are left disabled; you get a button to open them in the real editor. Copy it writes is grounded in your workspace voice profile and follows the same human-writing rules as the rest of AY Inbox.

Each model turn costs one AI credit, and a run is capped at 20 model turns between approval pauses. A turn can call several tools, so the activity list can show more than 20 tool steps. Aya reserves the last turn for an answer from the evidence already gathered, without calling more tools. If it cannot verify the answer, it is instructed to explain what remains unknown and what to check next. An empty provider answer is reported as an error. Remaining credits show at the panel's bottom, and the stop button ends a run anytime. On a self-hosted install with a free or local model, runs are not charged and a "free model" notice appears.

Completed tool steps stay in the conversation history if a later provider call fails or you stop the run. Aya gives recording those steps and refunding a failed model turn up to five seconds to finish after cancellation. Changes Aya makes (a created task, an updated contact) appear in the shared activity feed like any other edit, so teammates' views stay current.