It would rather say nothing than make something up
Every design decision that followed came from this one. If your content does not cover a question, the honest answer is the one you wrote for exactly that case.
About us
We built the support agent we wanted to be a customer of.
Every team we knew was answering the same dozen questions by hand, and every tool offered to fix it by guessing. So we built one that answers from your own content, admits it when your content does not cover something, and hands that case to a person — in the system your team already opens every morning.
A wrong answer costs more than no answer.
Support chat went through a strange few years. The tools got fluent long before they got honest, and fluency is exactly the wrong thing to be good at first. An agent that invents a refund window does not save your team an hour — it costs them a day, and it costs a customer their trust in whatever they read next.
01 What we will not trade away
Four decisions, not four values.
Not positioning. These are the decisions the product is actually built around, and the ones we would argue with a customer about.
Your content and your customers stay yours
Databases are read over a read-only connection. Personal details are masked before text reaches the model. Files a visitor sends live in a private bucket behind short-lived links.
We are not trying to become your help desk
Tickets go to an endpoint in your project, in your shape, into the tool your team already uses. One fewer tab, not one more.
You should never be surprised by the bill
Every AI call is recorded against the conversation it ran for, with its tokens and cost. Go over your allowance and we write to you rather than billing you.
02 How it is built
Boring, in the way infrastructure should be.
The agent is only as good as what it is allowed to read and where it is allowed to put the result. So most of the work has gone into those two ends rather than into the conversation in the middle.
01
The reading end
6 kinds of source and 24 database engines: crawling, scheduling, re-reading, retiring, masking, and folding related rows into one document a model can actually use.
02
The writing end
A ticket endpoint you own, with your fields, a bearer token you set, and a test button that proves it works before anything is live.
03
The bit in the middle
Deliberately small. Find the passages, answer from them, or stop. 6 widget layouts sit on top of it, and none of them change that rule.
03 The part most tools get wrong
Anything it cannot answer becomes a ticket in your system.
Not in ours. The chat collects what your team needs, shows the visitor a summary to confirm, and posts it to an endpoint in your own project — files and all.
See the whole pathTicket #218409:41
Expedited refund request
- Order
- A-10294
- Transcript
- the whole conversation, attached
- Posted to
- yoursite.com/api/support
Accepted by your endpoint · 200 OK
04 Who we are
Made by Micronstars.
We are a small software team in India. We build the product, we run the servers it sits on, and when you write to us it is one of us who writes back — not a queue, and not an agent answering about the agent.
- [email protected]
- +91 63545 40630
The honest way to judge it is to ask it something.
Point it at one page of your own and ask the question your team answers most. That is the whole evaluation.