Why we are building a companion with a computer of its own, not a chatbot that talks about it

Ask most assistants to file a VAT return and you get a good description of how to file a VAT return. The knowledge is real, but the work stays where it was — with you, at the keyboard, doing the clicking. The gap between "here is how" and "it is done" is the whole job, and it is the gap we are trying to close.

Doing the task, not describing it

The companion we are building operates a computer directly. It opens applications, moves files, fills in forms, and carries a task from the first step to the last. Not a plugin for one tool and not an API integration that only works with software that happens to expose one — the same surface a person uses, driven the same way a person drives it.

That matters most for the work nobody has bothered to automate, because automating it properly was never worth a project of its own: the monthly reconciliation, the supplier chase, the batch of invoices that needs pulling out of a folder and typing into a portal.

The computer it works on is its own

This is the part most people assume backwards, so it is worth being plain about: the companion is not primarily a thing that takes over your laptop. It has a machine of its own — its own desktop, its own browser — and by default that is where the work happens.

The consequence is the bit that changes how it feels to use. Because the task is not running on your computer, it does not need your computer to be awake, unlocked, or even open. Hand it a job, close the app, go to a meeting. It carries on without you and it is done when you come back. An assistant you have to sit and watch is just a slower way of doing the work yourself.

It also means a long job or a risky one never has to touch your files to get done.

And your own computer, when you ask

Some work genuinely has to happen where you are — the file is on your desktop, the system only exists on your network, the login lives in your browser and is staying there. So you can ask it to work on your machine instead, and it will: same abilities, your computer, because you said so rather than because that is the only mode it has.

A memory that outlives the conversation

Context that dies when you close the window is not memory, it is a scratchpad. Tell an assistant once that month-end here is the 25th and not the last working day, and you should not have to say it again next month.

So the companion keeps a long-term memory vault: things you have told it, kept deliberately rather than incidentally. You can read what is in there, edit it, and wipe it. A memory you cannot inspect is not something a business can reasonably be asked to trust.

The languages people here actually speak

This is the part that made us want to build it ourselves. South Africa has twelve official languages, and almost every assistant on the market treats English as the real one and the rest as a translation problem to be solved at the edges. That is not the same as understanding, and anyone who has watched a parent give up on a voice interface knows the difference.

We are building for the languages people here use with each other — including switching between them mid-sentence, which is how most of us actually talk. And it works out loud, in a continuous voice mode: you stay in the conversation instead of pressing and holding between every turn.

Where it is

It is in development. We are not putting a date on it, and we are not going to describe capabilities it does not have yet — everything above is something we can stand behind, written so it stays true on the day it launches.

It also does not have a public name. That is deliberate: we would rather name it when it is ready to be used than pick something now and spend the intervening months defending it.

If the work you would hand it is the kind described here, we would like to hear what that work is — tell us what you would want it to do, and we will come back to you when there is something to show.