AI Assistant

The AI Assistant is the chat panel you open from the floating button in the bottom corner of the app. You can ask it questions about your data and, in bases where it is enabled, ask it to make changes for you — create a table, add records, update a workflow, and so on.

Attaching files

You can give the assistant images or PDFs to look at — a screenshot, a photo, a scanned invoice — and ask questions about them.

Attach with the paperclip. Click the paperclip button next to the message box to pick one or more files, or simply drag and drop them anywhere onto the chat — a "Drop files to attach" overlay appears while you drag. The files you stage show up as small chips (images as thumbnails) above the message box; click the ✕ on a chip to remove one before sending.
What you can attach. Images (PNG, JPG, WebP, HEIC) and PDF files, up to 10 files per message and 15 MB each. Anything else — or a file that is too big — is skipped with a short note telling you why.
Send with or without text. You can send attachments on their own or together with a question. After the message sends, the staged files clear automatically.
On reload. When you reopen an earlier conversation, messages that had attachments show the file names, but the file contents themselves are not re-sent to the assistant.

What the assistant remembers

The assistant can remember durable facts you tell it about yourself, so you don't have to repeat them in every conversation. If you say something like "I'm in the Sales team" or "always show me dates as DD.MM.YYYY", the assistant can save that as a memory and use it to give you better answers later.

Only what you tell it. The assistant remembers facts and preferences from your own words. It does not mine your emails, records, or web pages for things to remember.
Private to you. Memories are stored per user. No one else can see them, and they are never shared across accounts.
Reference only, never commands. Memories are treated as background information about you. They can never change what the assistant is allowed to do or bypass the confirmation step for actions.
You are in control. Open the Memories panel from the brain icon at the top of the chat to see everything the assistant has remembered. You can edit the wording of any memory or delete ones you no longer want. A memory can apply everywhere or be scoped to a single base.

To ask the assistant to remember something, just tell it — for example, "Remember that I prefer short answers." To make it forget, delete the memory from the Memories panel.

Skills — reusable instructions

A skill is a reusable snippet of instructions you save once and activate whenever you want it to shape how the assistant works — a preferred tone, a specific output format, or a step-by-step way of doing a task. Open the Skills panel (the sparkle icon in the chat header) to manage them.

Create a skill. Click New skill, give it a short name, an optional description, and the instructions (the guidance itself — for example, "Reply in Lithuanian and keep answers under three sentences, always ending with a summary table."). You can also ask the assistant to create one for you — because saving a skill changes what it will do later, it shows you the full instructions to approve first.
Personal or shared. By default a skill is personal — private to you. Turn on Share with this base to make it available to everyone in the base; sharing (and editing or deleting a shared skill) requires base-admin rights. You can only edit or delete your own personal skills.
Activate a skill. Activate one from the panel, or just ask ("use my Terse skill"). Active skills appear as chips above the message box; the assistant follows their guidance until you remove them. Click the ✕ on a chip to deactivate it.
What a skill can and can't do. A skill only guides how the assistant approaches a task. It is applied as reference guidance and can never give the assistant new tools, widen your permissions, or let it skip a confirmation — those rules are fixed by the system and are unaffected by skill text.

Actions that need your confirmation

For safety, some actions the assistant wants to take do not run automatically. Instead of changing your data straight away, the assistant pauses and shows you a confirmation card in the chat, then waits for your decision. This means a create, update, or delete never happens behind your back — you are always the one who says "go".

What the card shows

Each confirmation card describes exactly one proposed action:

Action name — a friendly label for what will happen (for example, "Create table" or "Delete record").
Summary — a short, plain-language description of the change.
Details — the specific values the assistant plans to use, listed as a compact set of fields (for example, the table, the record, or the name being set).
Status — a badge showing where the action stands: Needs approval, Executed, Rejected, Failed, or Expired.

Approving or rejecting

While an action is waiting, the card shows two buttons:

Approve — runs the action now. The assistant then continues the conversation and reports the result. Once it finishes, the card updates to Executed and shows what happened.
Reject — cancels the action so it is never run. You can add an optional reason; the assistant notes your decision in the conversation and the card updates to Rejected.

You do not have to decide immediately. Confirmation cards stay in the conversation, so you can reopen the chat later and the cards will still be there with their current status. A card only ever runs once — approving the same card again does nothing new.

Sending email from chat

When you ask the assistant to draft or send an email, it never sends anything on its own. Instead it opens an editable email composer right in the chat — a confirmation card you review, change, and only then approve.

Reviewing and editing before you approve

The composer is prefilled with what the assistant proposed, and every part is yours to change before sending:

From — pick which of your own connected email identities to send from. Only your mailboxes appear here; you can never send as another person.
To, Cc, Bcc — the recipients, entered as comma-separated addresses. Cc and Bcc are hidden until you add them with the Add Cc / Add Bcc links.
Subject — the email subject line.
Message — the body text. If the assistant wrote an HTML message, a small note tells you so.
Signature — choose one of your own saved signatures to append, or No signature. Your default signature is marked.

Your edits save automatically as you type, so the version that sends is always exactly what you see on the card. When you are ready:

Approve & Send (shown as Save draft for a draft) sends the email — with your edits — from the identity you chose. The card then updates to Executed.
Discard cancels the email so nothing is sent.

Because you approve the final, edited content, the assistant can propose a rough draft and you can polish the recipients, wording, and signature before it ever leaves your mailbox. For safety there is also a daily limit on how many emails the assistant can send on your behalf.

Email signatures

The Signature picker in the composer lists signatures that you own. You manage them in Settings → Signatures.

Create or edit. Give the signature a name, choose which email account it belongs to (or All accounts), pick a language (English or Lithuanian), and enter the signature body. You can type simple HTML for formatting; a live preview shows how it will look. A separate plain-text version is used for plain-text emails, and is filled in automatically from the HTML if you leave it blank. For your safety, the HTML is cleaned when you save — scripts and unsafe attributes are removed.
Default per account. Turn on Use as default to make a signature the default for its account. Each account (and the shared All accounts group) can have only one default, so setting a new default automatically clears the previous one. The composer pre-selects your default and marks it in the list.
Delete. Remove a signature you no longer need; you will be asked to confirm.

Signatures are private to you — no one else can see or use them, and they are never applied to automation or notification emails.

Operations panel

The Operations panel (the activity icon in the chat header) gives you visibility into what the system is doing and how much it has consumed.

Workflows tab

Shows automation workflow executions from the last 24 hours — running, succeeded, failed, or cancelled. Running executions appear at the top with a pulsing indicator; if one has been running for more than 30 minutes it is flagged as Stale. Click the stop button to cancel a running execution.

Tokens tab

Shows how many AI tokens you have used today, this week, and over the last 30 days, broken down by model. This lets you track your AI usage at a glance without needing to check an external dashboard.

Actions tab

Lists all pending (awaiting your approval) and recent (last 24 hours) agent actions across every conversation. Pending actions show a badge count on the Operations icon so you know when something is waiting.

Conversation management

The History sidebar (clock icon) lets you switch between conversations, search across them, and manage them.

Search. Use the search bar at the top to filter conversations by title or message content.
Preview. Each conversation shows its last message, a relative timestamp ("2h", "3d"), and a message count so you can find the right one at a glance.
Streaming indicator. A pulsing blue dot appears on the active conversation while the assistant is generating a response.
Rename. Click the pencil icon on any conversation to give it a meaningful title. Press Enter or click away to save.

Tips & Troubleshooting

Expired cards. For safety, a pending action is only valid for a limited time. If you leave a card unanswered for too long it becomes Expired and its buttons are disabled. Just ask the assistant to try again and it will propose a fresh action.
"You don't have permission." Approval re-checks your access at the moment you click, using your own role in the base. If your permissions changed, or you are a read-only collaborator, the action is refused and the card shows a permission message — it is not run. Ask a base admin for the access you need, then retry.
"Failed." If an action was approved but couldn't complete (for example, a value was no longer valid), the card shows Failed with a short explanation. Nothing partial is left behind for a failed action — you can ask the assistant to try a different approach.
No cards appear. If the assistant makes changes without ever showing a confirmation card, then confirmation is simply not enabled for that base. This is a per-base setting; contact your administrator if you would like changes to require approval.