Decurix gives every AI worker a desktop of its own. Brief it like a colleague, teach it like a new hire, and work alongside it. It asks when a call is yours, proves every result before it says done, and gets faster and cheaper with every task it learns.
Windows and Linux · runs on your own machine
A picture of the Decurix app: three worker desktops, one preparing a sales return, one paused for approval before paying suppliers, one finished and checked; and a "Needs you" panel with the approval request.
Software gave people tools. The next software gives them workers: AI with a computer of its own, a responsibility, and the ability to learn. Decurix is the operating system they run on.
Teach a worker the way you'd teach a new hire. Brief it in plain words, answer its questions, or record yourself doing the job once. Every answer becomes memory, and work it has proven becomes its own script.
One conversation per worker, on your laptop or your phone. It works on its own screen while you do yours. Questions and approvals arrive in one place, and anything that sends, pays or publishes waits for your yes.
What a worker learns is yours: its memory, its skills, its record of every run. Each run is graded and kept, so your own work becomes the evidence a worker earns trust on, and the training data for a model you own.
Click New desktop and choose its brain. It gets a full Linux desktop with a browser, office apps and a terminal.
Attach the files and say what you need, in your own words. No workflow builder, no integration first.
It opens apps, reads files, fills forms and writes documents. Every action is recorded with before-and-after screenshots.
When a call is yours it asks a precise question with the evidence, instead of guessing. Your answer is remembered.
Totals, files, paths and recipients are checked in code before any model's opinion. Only a confirmed result is reported done.
A job it has proven runs as its own script or replay, still checked every time. You're needed less with every run.
A worker reconciling a bank statement. It asks one question, remembers the answer, and checks the totals before it reports done.
An animation of a worker matching bank statement lines against a ledger, asking one question about a possible duplicate payment, recording the answer, and confirming the totals match before reporting done.
Most AI agents think through the same job from scratch every time. A Decurix worker keeps what worked. The strongest model handles anything new. Once checked runs agree, the job becomes the worker's own script or replay, and every result is still checked. A miss sends it back to the model, and back to learning.
In that order. A worker may ask, or say it failed. It must never report success for something it did not see.
Facts a program can decide (sums, files, paths, recipients) are checked in code before any model gets a vote.
Send, pay and publish stop for approval every time. There's no "always allow", and no skill can switch the gate off.
A skill stays an apprentice until clean, checked runs prove it. One slip and it's learning again.
A smaller model with your answers should reach the same result as a bigger one alone. A smarter model just asks you less.
Business rules are yours to set in plain language. The worker reads them before every task. No per-app hacks.
Each task shows what was asked, what was done, where a person stepped in, what it cost, and whether the answer held up.
Workers' desktops run on your own machine. Share one folder if you like, and nothing else of your PC is reachable.
Passwords are withheld at the source. They never reach the chat, the task record or a recording.
Every task a worker does is a probe: something to learn from and generalise. We think a worker that learns from each one, and proves what it learned, can in time take on anything a person does at a computer.
Real work on a real computer, with you in the loop for judgement and approvals, and every result checked.
Practice on realistic cases, prove a whole process on evidence you review, then run it unattended within that scope.
Workers that hand work to each other, ask a taught colleague before a person, and own whole processes.
Your workers, your skills and your own model: a workforce that keeps what it learns and keeps getting better.
One app. It sets up the workers' desktops locally, and your files stay on your computer.
The first public builds are on their way. Leave your email below and we'll write when they're ready.
Leave your email for early access and release news. Early teams shape what Decurix learns first.