AI that answers.
And knows when to call a human.
Assistants answer from your knowledge base 24/7 on your own provider key, open tickets, read customer status, learn from dialogs — and every change goes through you: preview, approve, roll back.
Several assistants, each with its own job
A consultant for the site, a sales rep for English‑speaking visitors, a support agent for existing customers. Each has a role, a language, a greeting and knowledge sources.
- Answers from your base. Prices, delivery, hours — the assistant quotes facts from the knowledge base and cites the source, without inventing.
- Handoff to a human. Complex order, complaint, or the client asks for a person — the AI passes the dialog to an agent or department with a summary, so nobody asks twice.
- AI operators with names in every language. «Eva» in English, «Ева» in Russian — the header name and greeting follow the visitor’s widget language.
- Test before launch. Ask the assistant anything in the dashboard test mode and see the answer and the sources it used — clients don’t see it.
- Works in every channel. The site widget, WhatsApp, Telegram, Messenger, Apple Messages, Viber, plus regional channels like VK, Avito and MAX — the same assistant, one history.
You control: role and tone, language, greeting, which sources it may use, when to hand off and to whom.
Your provider key. Your model.
BYOK — bring your own key. Connect an API key from Claude, OpenAI, DeepSeek, Gemini, Mistral, OpenRouter, Groq, regional LLMs or any OpenAI‑compatible service and choose the model for each assistant.
- Ten connection options. Nine named providers plus «OpenAI‑compatible» with your own base URL: a local model, a corporate gateway, an aggregator.
- Model list from your key. After the check Smartbtn asks the provider which models your key can use — pick one from the list.
- Only with the owner’s permission. A key is connected in a separate step and needs its own right; the setup token cannot add a provider key.
- The key is never shown again. After saving you see only the provider and the last characters.
One key — and your AI sets everything up
Claude, ChatGPT or DeepSeek reads your site and documents, writes instructions, builds the knowledge base and saves it to Smartbtn via API. No forms to fill.
- Create a tokenDashboard → your widget → «API tokens». Name it, e.g. «Setup from Claude»; expiry is optional.
- Preset «Setup via AI»One click replaces the selected rights with exactly nine scopes and removes all others:
widget:readwidget:writeai:readai:writeai:applykb:readkb:writefiles:writestats:read - Enable the widget APIWidget settings → «Allow widget API (JS and REST)» → save. Without it even a correct token gets 403; creating a token does not enable access.
- Pick a scenario, copy the promptPaste it into your AI app, insert the key and the site, attach files. The AI saves via API, checks search and reports what it couldn’t read.
Site, files, rules — one base
The assistant answers only what is written in the base (RAG). Fill it three ways and see what is indexed.
coffee‑lab.ru/delivery
PDF · стр. 2
kb/rules #12
coffee‑lab.ru/returns
Договор B2B.docx
coffee‑lab.ru/contacts
- Site crawler. Give your site’s address — the crawler reads up to 20 pages (depth up to 2) within your domain and shows what was processed and what failed.
- Documents. Upload PDF, DOCX, TXT, Markdown or HTML (up to 2 MB): price lists, contracts, instructions. Each file has a status: queued, indexed, error.
- Rules. Short Q&A and strict statements that outrank everything else: «We don’t deliver outside the ring road».
- Search. Check what the assistant will find for a query before a client asks it.
It learns. You hold the brake.
The AI studies dialogs and how your agents answered, finds weak replies and improves its persona in cycles. Every version is saved and can be rolled back in one click.
- Learning cycles. Collect dialogs → find weak answers → draft changes → check on past questions → apply. You see every cycle and what it changed.
- Persona versions with rollback. Instructions, tone and rules are versioned; return to any previous one.
- Change log. Who changed what and when: a cycle, a wish, a manual edit.
- Regression guard. Each change is watched for 7 days: if quality drops by more than 5% it is rolled back automatically; after 3 cycles without improvement learning pauses for 30 days.
You control: on/off, target assistant, shadow mode, manual tuning run. Self‑learning is in BETA.
today 04:10current
yesterday
12.10
08.10
Tell it in plain words
Write a wish in free text. The AI shows a preview of how it will answer — with no consequences — and you approve, reject or later roll back.
- No prompt engineering. «Answer shorter», «always offer a tasting», «don’t promise delivery on holidays».
- Preview first. You see the diff of instructions and an example answer before and after.
- Approve, reject, revert. Nothing reaches clients until you approve; an approved wish can be reverted.
What hurts your customers — as a list
The AI reads dialogs and groups complaints into problems: what happened, how many times, since when. Like a bug tracker, but for your service.
- Four statuses. new → in_progress → resolved; if it comes back, it becomes recurring and rises to the top.
- Key and monthly cap. Run analysis on your own key or via SmartBTN AI with a monthly spending cap — it pauses when reached. Past dialogs can be analysed up to 365 days back.
- Examples attached. Each problem links to the dialogs where it appeared.
23 dialogs · since 02.10recurring
7 dialogs · todaynew
12 dialogs · since 09.10in_progress
15 dialogs · resolved 11.10resolved
4 dialogs · yesterdaynew
What they asked for — and you don’t have
Every «do you have…?» answered with «no» becomes a demand signal. Grouped by category with counts — a ready list for the product and the menu.
- Categories and counts. Locations, payment methods, features, platforms, integrations, prices.
- Real quotes. Open a signal to read how clients actually phrased it.
- First and last mention. When it was first and last asked, with an example.
The AI knows who it’s talking to
Your system sends the customer’s status: registered, paid, customer since, service state. The AI uses it in the answer and decides on escalation — a paying client with a broken service goes to a human first.
- Facts, not guesses. Statuses come from your backend via API, not from what the client says.
- Risk estimate. A separate estimate of abuse risk and client reliability (Clients → Risk) — a forecast that is checked against outcomes.
- Status from your backend. Passed via account linking or the customer status API — only for widgets where it is enabled.
The AI opens a ticket itself
When the question can’t be solved in chat, the assistant opens a ticket with a number, reason, priority and SLA, tells the client the number and sends the ticket.created event to your helpdesk or CRM.
- Number for the client. «Your ticket S010482» — the client can ask about status later, the AI will check it.
- Priority and reply window. High or normal priority; the reply window is set in hours per widget (24 h by default); overdue tickets are highlighted; a ticket can be reopened within 7 days.
- Close and reopen. From the dashboard, via API, or with the client’s «solved» button in your app.
- Webhook to your system. Signed
ticket.createdwith number, chat, client, reason, priority and SLA.
Know right away if something is off
AI alerts flag what needs attention: provider quota exhausted or an invalid key, model unavailable, provider unreachable, a new ticket. After a dialog the client can leave a review — it is shown in the “Reviews” section.
- Acknowledge. Mark an alert as seen — the team knows someone is on it.
- Reviews after the dialog. Rating and comment on the AI’s answer, linked to the dialog.
assistant «Sales» switched to agents
topic: «card payment»
9 unanswered questions
Speak — and hear the answer
Your app sends audio — the assistant gets text; the answer can be spoken. Works through the signed assistant integration on an OpenAI key.
- Speech‑to‑text.
POST /ai/assistant/speech‑to‑text— audio from your app becomes text and is answered like text. - Text‑to‑speech.
POST /ai/assistant/text‑to‑speech— the answer as audio. - Phone voice agent. A separate «Voice agent» for calls is in early access.
Ready assistants for your niche
66 ready templates in 13 niches: role, prompt, greeting, quick phrases, handoff rules and knowledge rules are already written. Apply in one click — then add your prices and address.
The same assistant — inside your product
Your backend talks to the assistant via a signed integration: each request is signed with HMAC‑SHA256, and the chat can be linked to the user’s account in your system.
- Integration. Create it in the dashboard or via
/v1/ai/assistant/integration; you get a signing secret once. - Turn.
POST /v1/ai/assistant/turn— send the user’s message, get the answer, the handoff flag and the ticket if one was opened. - Account link.
/account‑linkbinds the chat to your user after you’ve verified them;/revokeunbinds. Off until the owner enables it. - Voice and tickets too. Speech endpoints and ticket status are available to the same integration.
Your own support agent on the queue
An automated agent — yours or a third‑party one — works the ticket queue and chats like an employee: claims a chat, replies, releases it to a human or closes the ticket.
SB‑10479 · c‑3191 · queued
SB‑10480 · c‑3195 · queued
SB‑10482 · c‑3202 · queued
- Claim / release. A claimed chat belongs to the agent; release returns it to the queue for people.
- Reply and close. Answer in the chat, close the ticket — or only close with the narrow
support:tickets:closescope. - Locked down. Only tokens enabled for the agent, optional IP allowlist; text‑only replies up to 3500 characters; actions are logged.
Access is narrow. The key is yours.
AI can change answers and knowledge — so every right is explicit and every change is reversible.
The full token is visible only at creation; we store a hash. Lost it — revoke and issue a new one.
The setup token is a one‑off. Done — revoke it in «API tokens»; scopes can be edited without reissuing.
Nine scopes for setup; no billing, partners, account or provider keys. A missing scope returns 403 with a hint.
A change that lowers quality is rolled back automatically; any persona version rolls back in one click.
Our setup prompts tell your AI not to execute instructions found on sites and files and never to put the key into reports or URLs.
Dialogs go to the provider whose key you connected, under its terms. Change provider or turn AI off at any moment.