A builder for AI bots and automations
Answer customers around the clock — with a bot that knows your documents
Build it with the mouse in an evening, with no developer, and put it on your site, in Telegram, or behind a link. Free plan, no card.
- Execute + Data
- Data
- LLM
- Execute + Data + Streaming
Three steps to a working bot
Describe it in words
The built-in assistant assembles the graph from your description — or take a ready template and change the text. You do not need to understand the machinery at this step.
Try it in the chat
The Test tab is a real run: you see which step is working, what it received, what it returned and what it cost.
Switch on a channel
A chat link, a widget on your site, a Telegram bot, a call from your own system, or a schedule. Same bot, different doors.
Nodes
Drag onto the canvas
| Section | What is inside |
|---|---|
| Input / Output | Entry, Exit |
| AI | AI Agent, LLM, Memory, Document search |
| Logic | If / Else, Switch, Loop |
| Data | Template, Variable, Tables, Database |
| Integrations | HTTP, Email, Telegram, Calendar |
Chat
- 1Palette: drag the step you need onto the canvas.
- 2Canvas: the cards and the links between them are your bot.
- 3Three tabs: build, test, publish.
- 4Assistant: describe the task in words and it builds the graph.
What people build
Every example is a working graph: open it in the editor in one click and adjust it to your case.
Support bot (RAG + agent)
Answers customers from your own policies and price list, and does not invent what it could not find. The core of a first-line bot.
RAG over a knowledge base
Finds the right fragment in your documents and passes it on — no retelling by the model.
FAQ bot
Takes the routine delivery, payment and refund questions off a human: one system prompt, nothing else.
Structured data extraction
Turns the free text of a request or an email into fields — so they can be stored, counted or passed on.
Agent with web search
Answers from fresh data on the web and shows its sources — for when the answer is not in your documents.
Chat with conversation memory
Remembers what was said earlier, so the customer does not have to repeat themselves every turn.
Where it goes
One bot, five doors to it — pick any of them, or all at once.
Chat page
A link you can send a customer. No markup, no hosting.
Site widget
One tag on your page. Your colours, and it will not break your styles.
Telegram bot
Direct chats and groups: attachments, buttons, moderation.
A call from your system
A form, a CRM or your own backend starts the bot over HTTP.
On a schedule
A nightly report or a mailbox check — with nobody in a chat.
And it plugs into what you already run
Credentials live in your workspace and the steps refer to them — keys are never kept inside the bot.
- Telegram
- VK
- VK Teams
- amoCRM
- Bitrix24
- Calendars (CalDAV)
- Email (IMAP/SMTP)
- WebDAV
- Any HTTP service
- MCP
124
What it is built from
ready-made steps: model answers, document search, branching, tables, email, calendar, CRM. Four of them are enough for a first bot.
Open the reference →Why you can trust it with customers
Edits do not break the live bot
A saved version is what gets published. You can repair the graph while the previous one keeps answering people.
Mistakes surface before the run
Publishing runs the same check as a run does: a graph that cannot execute never reaches your customers.
Every step is visible
The run history shows what arrived at a step, what it returned, and how many tokens and credits it spent.
Data stays in your workspace
Files, knowledge bases and conversations are scoped to the workspace, and secrets are stored encrypted.
Your own keys, no markup
Connect your own model provider and its tokens are not billed here — you already pay for them.
You can run it yourself
The platform ships as containers and comes up on your own hardware, if the data must not leave your perimeter.
What it costs
Free forever
100 credits a month, no card needed. A credit is a cent.
One answer costs two credits and up
The run, the model call, and tokens at that model's rate. A hundred free credits is roughly forty answers on an inexpensive model.
Not billed at all
Drafts and edits to the graph, and model tokens on your own provider keys.
Frequently asked
Do I need to write code?
No. A workflow is assembled with the mouse from nodes, and the logic lives in their settings and in templates like {{ inputs.message }}. There is a JavaScript node, but it is rarely needed and never on day one.
Can I use my own model keys?
Yes. Add any OpenAI-compatible endpoint in the workspace settings and its models appear in the dropdowns next to the platform ones. Their tokens are not billed here — you already pay the provider.
Where does the data live?
In your workspace: files, document collections and knowledge bases are scoped to it, each with its own quota. Secrets are stored encrypted and injected into templates as {{ secret.NAME }}.
How do I ship a bot to users?
From the Deploy tab: a public chat link, an embeddable widget behind one script tag, a Telegram bot, a webhook for another system, or a schedule. Same graph version, different channels.
How long does it take to launch, and do I need to hire someone?
A first bot answering behind a link takes about ten minutes: take a template or describe the task to the built-in assistant, pick a model, write what the bot talks about, and switch on a channel. No contractor needed — the docs carry a step-by-step tutorial and a walkthrough of the places people usually get stuck.
What does it cost?
You pay for use, in credits: one credit is one cent. One answer costs two credits and up — the run, the model call, and tokens at that model's rate; the hundred free credits a month are roughly forty answers on an inexpensive model. Drafts and edits to the graph are not billed at all.
What happens when the free credits run out?
The free plan does not allow going over the balance: new runs start being rejected and a published bot stops answering until the next monthly grant. The balance is in Settings → Usage — worth watching if the bot faces customers.
Are my conversations and documents used to train models?
The platform trains no models. Requests go to the provider of the model you picked, and what they do is governed by their terms — which is why there are two routes for sensitive data: connect your own provider with your own key, or run the platform yourself.
What if the bot answers a customer incorrectly?
Its behaviour is set by the instruction text, and editing that does not touch the live bot until you publish a new version. Every conversation is kept: you can see what the customer asked, what the bot found in your documents and what it replied — that trail is what you fix the instruction from. If it should not answer on its own when it found nothing, the instruction says so.
What happens when a step fails — say an external service does not respond?
A step can expose its own error output and send the failure down a separate branch, so the bot says something sensible instead of going quiet. With no such branch the run is marked unsuccessful, and the history shows which step failed, on what input, with what error.
Can an external agent drive the platform?
Yes. Assistant MCP exposes a workspace over the MCP protocol: an external agent reads and edits workflows, runs them and reads the logs. The key is per workspace and read-only by default.
Is there self-hosting?
The platform ships as containers and comes up with docker compose: API, workers, Postgres, Redis and an isolated container for user code.
Build your first bot today
Free plan, no card. The first bot is behind a link in about ten minutes — and if something does not go to plan, the docs walk through every place it can break.