The panel alongside
To the right of the conversation is the agent status panel. You don't fill anything in there: everything you see is the result of what was said in the conversation, and almost every button leads back into the conversation, not into a form. There are three exceptions — the bin next to the name, reissuing a key, and going to the Plan section: these buttons act immediately, without a conversation. You can also just tell the builder — it will ask for your consent and do it.
The panel answers one question: did my sentence actually do anything.
Why the numbers change by themselves
After each applied change the panel refreshes itself — without waiting for the end of the turn. So the number of knowledge sources changes exactly when the builder finishes reading your website, not half a minute later.
A row whose value has changed is highlighted. Without that, you couldn't tell whether your sentence did anything at all.
Panel rows
Name and status
At the top is the agent's name, and below it a label showing where it actually answers:
| Label | What it means |
|---|---|
| answering | people can write to the agent: there's an enabled channel or an allowed website with a chat window |
| silent: no credits | credits have run out, and the agent won't answer customers until they're topped up |
| channel not set up | a channel has been added but doesn't work: a token is missing, messages won't get through |
| not answering anywhere yet | no channels and no allowed websites |
This is worked out from where people can actually write, not from whether the agent exists. A green "live" dot in a dashboard where nobody can write to the agent would be a lie. The reverse is a lie too: an agent already talking to customers on your website being labelled silent.
The chat window on your site counts too
It's turned on not by a channel but by allowing a domain — but from the panel's point of view it's the same "somewhere to write". Allowed websites are listed by name in the Channels row. See Where the agent answers.
Deleting the agent
Next to the name is a bin. It does not lead into the conversation — it asks you itself and deletes right away. "Delete this agent", said to the builder, does the same — after your "yes", and with the date until which the assistant can be restored.
Before deleting, the dashboard names the consequence instead of asking "are you sure?": the chat window on your website will stop answering, connected channels will go quiet, the schedule will stop.
You can undo this — but not forever
The data stays around for a while, and the dashboard tells you the exact deadline right after deletion. Until then the assistant can be restored with the same ID — you won't have to change the widget code on your site. After the deadline, it can't.
Several agents
If you have more than one agent, chips with their names appear under the name. The selected chip is the agent you're currently talking about and whose status the panel shows. The builder works on the same agent.
Knowledge
How many knowledge sources the agent has. Clicking the row opens the list; for pages read from your website it shows when they were read, when they'll be re-read and whether they still exist on the site — see Knowledge from your site stays fresh.
| What's shown | What it means |
|---|---|
| empty so far | no sources; the agent will answer only from what's written in its personality |
| a number | that many sources |
| 500 or more | a great many sources; the exact number wasn't counted |
| — | not loaded yet |
One page read from your site is one source. So after "read my website" you'll see not one here, but ten or so.
Actions
How many outside systems the agent can call: check an order, submit a request, add a row to a spreadsheet. 0 means the agent can only talk for now. See What the agent can do.
Channels
Where people write to the agent from — all together: allowed websites by name (rombik.app) and connected channels (Telegram, via API). nowhere yet means neither.
Websites come first: for most owners the chat window on their own page is the first — and for a long time the only — place where the assistant talks to anyone at all.
Under a bot in Telegram, Viber, WhatsApp, Discord or Slack you can see who is linked (the owner or a trusted person) and whether the bot answers only them; the buttons "Link my chat", "Disconnect", "Linked only" and "Reply to everyone" are there too. See Personal bot.
Under a mailbox the platform reads itself, you can see what it does with emails: replies to every new email or read only, and whether emails in your name go out without asking or after you approve; the buttons "Read only" / "Reply to emails" and "Send without asking" / "Ask before each email" are there too. See Your email: read only. A read-only mailbox doesn't count as a place where the agent answers: an agent with just one such mailbox is "not answering anywhere yet".
Website code
Open the Channels row — below the list is the line of code that puts the chat window on a page: already containing this agent's key, with a "Copy" button. The "WordPress" switch shows the three plugin fields instead. The same block is in the Chat window sheet, under the preview. Details: The chat window on your site.
Access keys
Open the Channels row — below the list are two levers. You'll need them if your own system talks to the assistant, or if your website signs the people who are logged in (including the WordPress plugin — see the conversation on phone and computer).
- New API key — for when the old one has leaked somewhere. The value is shown only once, so copy it straight away.
- New visitor signing key — also shown once, together with the agent ID and a ready-made "For the WordPress plugin" line. After the change, everyone whose identity your website confirmed becomes anonymous to the bot until the site starts signing with the new key.
Both take effect instantly: the old key stops working that very second, and so does everything signed with it.
Memory
What the bot keeps between conversations. remembers visitors — a recognised customer doesn't have to repeat anything: the bot recalls their name, orders and what was agreed last time by itself. off — every conversation starts from a clean slate.
Tapping the row opens both switches:
- Visitor memory — with memory off, nothing is stored about customers; whatever was already stored is no longer shown or used.
- Repeat answers — for frequent identical questions ("what are your opening hours?") the bot gives a ready answer: faster and without spending credits. Only worth turning off if answers must differ every time.
You can also switch both in words — "don't remember visitors", "answer fresh every time".
Remembers about you
What the assistant has noted about you yourself in a linked chat: "Doesn't eat meat", "Meetings only after 10:00". The number of notes, and next to it how many of them were noted after reading someone else's text: those are worth reviewing. nothing yet — nothing noted; — — the notes couldn't be loaded, which is not the same as "remembers nothing".
Tapping the row opens the list: you can edit or forget any note, and at the bottom add one yourself. Only dashboard admins see this row. More: What it remembers about you.
Agent files
What the assistant has saved in its persistent folder between conversations: a log, notes, spreadsheets. The number of files and how much of the 100 MB is used. nothing yet — it hasn't saved anything; out of space — the folder is full and the assistant can only read it; unavailable — there's no persistent folder right now, files last for one reply; — — the files couldn't be loaded, which is not the same as "there's nothing".
Tapping the row opens the list: you can download or delete any file, and at the bottom delete all files. Deleted files don't come back, and "undo" in the conversation doesn't restore them. Only dashboard admins see this row: the folder exists only when it's you yourself writing to the assistant.
Skills
Instructions for the assistant for specific tasks: the number of active skills, and next to it how many drafts suggested by the assistant itself are awaiting approval. none yet — not a single one; — — the skills couldn't be loaded, which is not the same as "none".
Tapping the row opens the list. At the top is "Wants to remember": drafts with "Save" and "No" buttons; a draft written after reading someone else's text is marked, and it takes a second tap to save it. Below are the active skills: expand one to read its text, or remove it. At the bottom is a box for a link to a ready-made skill: first the text is shown on screen, then "Add skill". Every dashboard member sees the list; admins change it.
Calendar
Whether your calendar is connected and what the assistant can do with it: not connected, sees busy times only (secret address), books with your permission, books on its own — or can't be read. — — the calendar status couldn't be loaded, which is not the same as "not connected".
Tapping the row shows the connected calendar, your working hours, when it was last read, and the coming week's events as the assistant sees them; "Check now" and "Disconnect" are there too. Only dashboard admins see this row.
Language
Appears only when a language has been set. Shown in words, not as a code:
- detects automatically — the agent answers in the language of the person it's talking to;
- Ukrainian, English — the agent always answers in that language.
Plan and agents
Two rows at the bottom of the panel answer a question you'd otherwise only see as a refusal.
| Label | What it means |
|---|---|
| Plan | which plan is active right now, by name: Free, Starter, Pro, Business |
| Agents | how many agents you already have and how many are allowed: 2 of 3 |
| — in the plan row | the name couldn't be loaded; the limits next to it are still real |
The Plan row opens — it leads to the section with prices and where plans and credit packs are bought. See How much it costs.
If you already have as many agents as your plan allows, the "one more agent" button isn't offered. The Agents row explains why: without it, the limit would only show up as a missing button.
Credits
A bar showing what's left, and a percentage. The bar empties: you're watching how much you still have.
- green — plenty left;
- yellow — a fifth or less left;
- red — a tenth or less left.
Below the bar is roughly how many conversations that will last. The number is worked out from your own spending, not the platform average: one person's conversation costs 60 credits, another's 300. Until there have been any conversations, this line isn't shown — someone else's average is worse than saying nothing.
If your plan has no ceiling, the bar is replaced by unlimited. Details: How much it costs.
What to do next
At the bottom the panel may show the nearest gap — and a button right there that continues the conversation with the right sentence:
| Label | Button next to it |
|---|---|
| No agent yet | Suggest where to start |
| Out of credits — the assistant is silent | Top up |
| Credits running low | Top up |
| Knows nothing about your business | Learn from your site |
| Not answering anywhere yet | Connect a channel |
Most of these buttons don't do anything by themselves — they put a sentence in the box for you to finish and send. Once the basics are in place, this block disappears.
Top up is the exception: money isn't handled by the builder, so this button doesn't write a sentence but takes you to the Plan section. If you're not the workspace owner, the button won't be there — instead there's a line saying who can buy more credits.
Customer payments
The Customer payments row says whether payment in the conversation via LiqPay is connected and how much has already been paid. It's shown even when payment isn't connected: that's where connecting it starts. Inside: the ceiling for a single link, how many unpaid links one conversation can get, the price list of services, and the latest payments with their status. If the payment details couldn't be loaded, the row shows a dash, not "not connected".
Access for other agents
The Access for other agents row says how many keys have been issued to other programs — Claude, Cursor, someone else's bot — that call the assistant as a tool, and whether it's visible to other agents. It's shown even at zero: that's where you start. If the keys couldn't be loaded, the row shows a dash, not "no keys issued".
Tapping the row opens the list: the key's name, its first characters and when it was last used. Here you can also issue a key (the value is shown once, together with a ready-made setup for the program) and revoke one. Below is the assistant's address for other programs and how other agents see it. Every dashboard member sees the list; admins issue and revoke keys. More: The agent as a tool.
Products
The Products N row appears as soon as the assistant has been given a link to your shop's products: how many there are, when prices were last checked, and a search — the same one the assistant uses. When the shop stops serving the file, the row is dimmed: "prices aren't updating".
If the feed has products with sizes and colours, it also says how many products the shopper sees (six T-shirt sizes are one product) and how many of them have variants; in the search, such a product shows "from 450 ₴ · 4 variants, 3 in stock".
Below that is Searched, not found: shopper searches that don't match a single product in the catalogue, with the number of times and the date. It's a direct hint about what to add to your range. Added the product? Click "Remove"; the row comes back by itself if shoppers search again and find nothing. Details: Store products.
Chat window
The Chat window row says whether it's the default one or already has its own look, and opens a preview: the real chat window with your settings on a sample product page. An agent that works on its own and doesn't talk to anyone doesn't have this row. Details: The chat window on your site.
Is it selling
The Is it selling row sits next to the chat window: how many conversations there were in the chat window this week and how many of them reached the cart — "140 conversations this week · 12 added to cart". It opens into a funnel for the last day, 7 or 30 days: opened the chat window → asked a question → clicked "add to cart", went to a product, went to a messenger, continued with the assistant in the messenger, left a contact, called a person — as numbers, with percentages between the steps.
While there are fewer than thirty conversations, there are no percentages, and the row says "not enough data": "2 of 3 added to cart" isn't 67 % — it's two people. What exactly is counted and what isn't: The chat window on your site.
Waiting on you
Rows that appear only when there's something to say — and they sit next to problems, not among the numbers. Each of them means the work won't happen without you, or won't happen at all any more.
| Row | When it appears | What to do with it |
|---|---|---|
| N actions waiting for your permission | the agent prepared an action that needs confirmation and stopped | click it — you'll see exactly what will be sent and to whom |
| N connections without consent | a link to someone else's account is set up, but nobody has given consent yet | actions on their behalf won't work until they do |
| Asking N | the agent stopped and is waiting for your decision | you write your answer right there; after it, the agent continues in a separate turn. Without an answer the work stands still, and after a week the question expires |
| Task N | the agent is doing a long job pass by pass — or it's paused | status and progress as numbers; you can pause, resume or cancel — Long task |
| Collected N | the agent has gathered cards through its work: properties, venues, contacts | sets with numbers, cards inside; wrong ones can be corrected or thrown out — What the agent collected |
| Postponed N | the agent moved a one-off job to later | you can see exactly what it will do and when; you can cancel it |
| Missed N | a job was missed, or the turn running it failed | next to it is a "Retry" button — the job is set up again in the agent's own words |
| Do not contact N | there's someone the agent no longer writes to | a list with reasons; a mistaken entry can be struck off |
| Asking to resume N | a person on the stop list changed their mind and asked the agent to write to them again | their words are shown verbatim; the decision is yours — the "Restore" button |
There's deliberately never a zero here: "Postponed 0" is a row about nothing that takes up your attention every day. And when there is a number, it means work that will wake up by itself one day and spend your credits.
What changed
A separate row opens the log: everything the builder did with the agent, as a set per conversation turn — who, when and exactly what. Each set is reverted as a whole: half an idea is worse than either of its complete versions. Irreversible sets don't offer the button and say exactly why. See Undoing changes.
Problems
If something has broken for the agent — a key to an outside service has expired, the service isn't responding, credits have run out — it's shown above everything else in the panel, in the same words used in the notification. Fixed problems don't take up space in the panel: it's about what's happening now. See When something breaks.
If a row shows a dash
The panel is assembled from several requests. If one of them didn't get through, its row stays a dash, and the rest show the truth. A panel that goes blank entirely because of one problem would show less than a panel with one dash.
If the status couldn't be loaded at all, a text warning appears at the bottom.
On a phone
The panel collapses into a single line above the input box. It shows:
- the agent's name;
- what the builder has just changed — or, if nothing has changed yet, the status label.
The last one matters more than the first two: it's what answers the question "did my sentence do anything?". The Show agent status button expands the full panel.
Ask instead of searching
The panel shows numbers, but it's easier to ask for details — the builder reads them from the same place the dashboard gets them:
How is it set up right now?
Purpose, role, tone, language, greeting, refusal text, contact form, who it can hand work to, which servers are connected and where problem notifications go.
How many conversations have I had?
How many conversations and messages, what people have been asking about lately, and how that's spread across the days of the week. The same numbers the assistant uses for its daily summary.
How do customers rate the answers?
How many good ratings, how many bad — and which exact questions got an answer people didn't like. That last part turns the number into work: the list of questions is the list of what's worth adding to knowledge.
Next
- What the agent knows — so the "Knowledge" row stops being empty.
- Where the agent answers — so the label becomes "answering".