atmon

Beta Built 2026-08-07

About atmon

Models got good at deciding.
Doing stayed hard.

The gap this closes is not intelligence. Your assistant can already work out what needs doing in your accounts. What it cannot do is act in them safely, at the size real work comes in, and leave you something you can check afterwards. That part is engineering, and this is the engineering.

A paper slip trailing from a finished action, pressed with a terracotta stamp: every call leaves a record.

Why this exists

Everyone has now had the same afternoon. You ask an assistant for something that would actually save you a day, it writes back something clever, and then the two of you sit there.

It cannot open the drive, cannot read the invoice, cannot send the email. The reasoning was never the bottleneck. Reaching the accounts was.

You can hand an assistant the keys to your accounts and hope. Plenty of tools do. But hope does not survive the first time a piece of software holding your login does something at scale that nobody asked for.

It does not survive the person in your company whose job is to ask where the passwords are kept, either.

A finished plan pinned above a workbench and a terracotta arrow leaving it and stopping at the shuttered cabinet below: the thinking is done and the tools are behind a door.

So the work is a layer underneath. It holds the keys, takes the plain-words request, checks what you allowed before it opens anything, and writes down every action. The assistant gets the ability to act. You keep the accounts, the limits, and the record.

What we believe

Three things this is built on

The limit is not how clever the model is

Ask an assistant to do a real piece of work and it stops at the first place it would have to act. A better model stops at the same place, faster. What is missing is not reasoning.

Permission belongs outside the model

A rule written into a prompt is advice, and advice is something a model can be talked out of. The same rule checked underneath, before the key comes out of the store, is a limit. Every limit here sits where the model cannot argue with it.

A record beats a promise

Anything can claim it finished. We would rather hand you the list of what it read, what it changed, what it skipped and what it is still waiting on, and let you check it yourself.

Where this stands

The honest boundary, in full

atmon is in beta. Here is the whole of what that means, in the same words we would use on a call.

How we build

The team

atmon is built by a small team, and two habits explain most of what you have read on this site.

We build against real work

Not against a feature list. Tell us what your week actually looks like and which apps it happens in, and that shapes what gets built next.

The fastest way to change this product right now is to describe a job you would hand over if you trusted it.

We draw rather than photograph

Everything on this site is drawn, and every drawing is written as code, so the page loads nothing from anywhere else and every reader sees the same thing.

We would rather show you a drawing of something that works than a screenshot of something that does not exist yet.