It makes things up
A general-purpose model asked about your delivery terms will produce something plausible. Plausible is worse than silence: the customer believes it, and your manager finds out at the refund stage.
Mimir answers strictly from your knowledge base — and when there is no answer in it, says so and calls a human. It runs on your own server, in one Docker command, and everything it learns stays there.
Do you deliver to Gyumri on Saturday?
Yes — Saturday orders placed before 2pm arrive the same day.
Can I pay in crypto?
There is nothing about crypto payments in my knowledge base, so I won't invent an answer. I'm bringing in a manager — right here in this chat.
The chat on a website is the first employee a customer meets. Three things usually go wrong with it, and all three cost more than the subscription.
A general-purpose model asked about your delivery terms will produce something plausible. Plausible is worse than silence: the customer believes it, and your manager finds out at the refund stage.
In a cloud chat service every dialogue, every phone number and every complaint lives on a third party's servers, under their terms, in their jurisdiction — and you are the one who signed the data-processing paperwork.
The moment a conversation matters — a real order, an angry customer — the bot keeps talking. What is needed there is one button that puts a manager in the same window, with the history already in front of them.
There is no programming anywhere in this list. Everything after the first step happens in the admin panel.
On your server or your VPS. Nothing else has to be installed — the database, the search model and the admin panel are already inside the container.
Topics in a question-and-answer form: delivery, returns, prices, how to order. This is the only thing the bot is allowed to say — which is exactly why it can be trusted with customers.
GigaChat, YandexGPT, or a local model on your own hardware. Then the persona: the name, the greeting, how formally it speaks. The safety rules are built in and are not up for configuration.
The widget picks up your colours, name and greeting by itself. There are ready instructions for Tilda, WordPress, Shopify, React, Vue, Next.js and Nuxt — and a live preview while you pick the colours.
A chatbot is only half the product. The other half is the desk your team works at every morning.
Semantic search finds the relevant topics, the model writes the answer from them. No relevant topic means an honest «I don't know» and an escalation — not an improvisation.
Open a conversation and the panel already shows a summary of what the customer wants, the knowledge-base topics that apply, and three ready replies to send or edit.
The manager joins the same conversation and the bot goes quiet. The customer never sees a handover, a ticket number or a second widget — it is the same window.
Contacts left in the chat become a list with a status, not a stream of messages. The dashboard opens on what is actually waiting today: open leads, escalations, dialogues since morning.
Accent, background, text, position, size, avatar — with a live preview. If that is not enough, there is a field for your own CSS, and the widget carries no framework to fight with.
The model provider's API key is stored encrypted and is decrypted only in memory at call time. It is masked in the logs, and it never appears in the interface after it is saved.
This is the question that decides whether a chatbot is allowed near real customers. The answer here is short and checkable.
The knowledge base in full, the history of every dialogue, customer contacts, the bot's settings and the widget's appearance — all of it in a volume on your own server.
The provider you chose receives the question, the few excerpts found for it, and the answer. Not the base, not the history, not the contacts. Choose a local model and nothing leaves at all.
The semantic search model runs locally on the CPU. Your knowledge base is never uploaded anywhere to be indexed — there is no vendor holding an embedding of your business.
Self-hosted means we have no access to personal data at all, and no agreement is needed with us about it. Your contract is with the model provider you picked, on their terms.
A licence and support, not a subscription per conversation. You bring your own model key, so the more the bot talks the less each answer costs you — the opposite of how SaaS chat is priced.
For one business with one site and a support inbox that is already too busy.
For a company with several brands — and for the one who has to keep them all answering the same way.
For studios and integrators who want to sell this under their own name.
Mimir has been answering on real sites since spring 2026. Alpha is the moment we open it as a product with a licence, a price and someone answering the phone.
Working on live sites: knowledge base, widget, manager desk, escalations, lead capture.
Closed alpha: first licensed installations, deployment done by us, weekly calls with every early client.
Open alpha: self-service installer, importing a knowledge base from an existing site, multilingual answers.
Beta: analytics on what customers actually ask, gaps in the knowledge base found automatically, CRM integrations.
Yes. Point it at a local model running on your own hardware and the loop closes entirely inside your network: search is local by default, generation becomes local too.
It says so plainly and offers to bring in a person, and the conversation appears in the panel as an escalation. It is not allowed to answer from outside the knowledge base, so there is no third option.
The installation is minutes. The honest answer is that the knowledge base takes the time: a usable first version is an afternoon of writing down what your support already repeats every day.
Colours, position, size and avatar are settings with a live preview, and there is a field for your own CSS underneath. The widget has no framework of its own, so nothing fights your styles.
Whichever you already have a contract with. Mimir speaks to GigaChat, YandexGPT, anything OpenAI-compatible, and local runtimes — switching provider is a select box, not a migration.
Send us the five questions your support answers most often. We will show you Mimir answering them from your own material before you decide anything.