AI agents that do the work, on your own machine.
Apollia OS is the program that runs them. You hand it a whole job instead of a question, and it goes and does the job: it opens your files, reads through your inbox, works out what needs doing, acts on it, and stops exactly where you told it to stop. The model that does the reasoning installs alongside it, on your own drive.
Unplug the network. The conversation keeps going.
Exactly what the preview covers
0.2.0-preview · macOS, Windows, Linux · No account, no licence key
- You hand over a whole task, not a question. The agent reads your files and your mailbox, drops what is already settled, and leaves the drafts.
- Before a move that commits you, it stops and asks for your approval. A refusal changes what follows.
- The model that makes it reason is on your disk: cutting the network changes nothing. With a remote model you configured, the work waits for the network.
- Nothing is sent without your approval, and every move leaves a signed line in the journal.
- Apache 2.0 or MIT The code is public, under either licence, your pick. Clone it, build it, fork it.
- macOS, Windows, Linux One installer per system, plus nine self-contained archives if you would rather stay in a terminal.
- The model is yours The inference engine ships inside the installer. You pick the model, and it runs off your own drive.
- Nothing to sign up for No account, no licence key, no subscription, and no trial that expires on you in two weeks.
An AI that answers is not an AI that does.
The step from answer to action is the step nobody takes for you.
That is where it stalls. For three specific reasons.
-
The work worth doing touches your most sensitive material.
Nobody wants to ship that somewhere else, and in plenty of trades nobody is allowed to.
-
A task is almost never a single question.
A dozen small moves across three tools. An answer, however good, performs none of them.
-
And you need to keep hold of the leash.
A program acting on your behalf has to stop where you decided it stops, ask before the moves that matter, and leave a record of what it did.
You: chase this month’s unpaid invoices.
An AI that answers
Here is a reminder template you can adapt: “Hello, unless we are mistaken, invoice…”
The work is still whole: you open the spreadsheet, find who owes what, write it twelve times.
An agent
- reads invoices.xlsx
- reads this month’s correspondence
- drops two invoices already settled
- writes twelve drafts
- pings you
Done. The drafts are waiting, nothing is sent.
The same sentence, two continuations. An AI that answers hands back a text, and the work is still to do. An agent chains the moves in your tools up to the result, and hands back at the end.
Apollia OS is built for those three. Here is where it starts.
Three beats, and you are running.
From the installer to the first job handed back, the path is three steps long. The second one is the only step that asks anything of you.
- It reads
- It picks a move
- Apollia runs it, or asks you
- It reads the result
It stops: answer handed back, or limit reached.
You install it.
macOS, Windows, LinuxYou say what needs doing.
Conversation or PythonIt works, then hands back.
Your limits, counted by Apollia
Chase this month's unpaid invoices without losing the evening.
This is what a complete piece of business work looks like, from the trigger through to the moment it lands back with you.
Five beats. You only step in on the last one.
-
The trigger
It starts without you, at a set time or when a file lands.
-
Gathering
It reads this month of correspondence and the payment rows, in your own tools.
-
Deciding
It cross-checks the two, drops whatever is settled, keeps the accounts worth chasing.
- Request for approval Leave twelve drafts in your mailbox One message per client. Nothing is sent.
If you approveThe drafts are saved, the spreadsheet is updated, and the journal keeps the record of the move.
If you refuseNothing is written. The agent hands back, and the journal keeps the record of the refusal.
Before a move that commits you, the agent stops on its own, asks, and waits. Two possible continuations, and you are the one who picks.
-
Acting
If you approveIt writes one message per client, leaves each as a draft, updates the spreadsheet.
Nothing is written. The agent hands back and notes your refusal.
-
Handing back
If you approveIt pings you when it is done. Nothing has been sent: the drafts are waiting.
It pings you all the same. No draft, and the journal keeps the record of the refusal.
You set the ceiling
Three limits, counted by Apollia and out of the agent’s reach: the number of steps, the number of tool calls, the duration. When one reaches zero the run stops, finished or not. The figures are the example’s; yours are set in the app.
Code never runs without you: running a shell command or a Python script is put to you every single time.
Eighteen tools in the box, and your accounts on the end of them.
Nothing to install before you start: eighteen tools are already there, and the model knows how to use them from the very first conversation.
-
Your files
Read, write, edit, list a folder, search by name or by content.
-
Running code
A shell command, a Python script, with your go-ahead on every single call.
-
The web
A search, then reading the page it found.
-
Its memory
It finds what it noted down, and only when it needs to.
-
Your rules
What is allowed and what is not, read again on every tool call.
-
You
When it is missing something, it asks you rather than guessing.
-
Your accounts and your team tools
Google and Microsoft in a click, everything else through MCP: eighteen preconfigured entries.
What changes for you, and how to see it for yourself.
Three things change the day the model runs at your place rather than somewhere else. For each one, the page that shows what holds it up in the code.
- Your client files can go through it. Why local
- You already know how to write one. For developers
- You will know what you are getting into before you spend time on it. The exact scope of the preview
The software, plainly.
Exactly what you are installing, and where each line is held up.
- Licence
- Apache 2.0 or MIT, your choice. No restriction on use, redistribution or forking. Both texts, in the repository
- Written in
- Rust for the runtime, Python for your agents. The core and the development kit carry no third-party dependency. The source code
- AI engine
- llama-server, built from llama.cpp sources and packaged inside the installer. Metal on Macs, Vulkan on Windows and Linux. The architecture, in the documentation
- Your data
- In a folder on your computer, in SQLite. The product has no online service to send them to. What stays with you, in detail
- Journal
- Append-only, SHA-256 chained and signed: a line cannot be lifted out of an agent run without it showing. The accountability model
- This page
- No tracker, nothing loaded from any host but this one, 238 KB on first display. The network tab of your browser shows it. What this site collects
- Scope
- A preview, 0.1.0. What it covers and what it does not is written down before you install it. The evaluation page
Here is the grid. Run it against everything you are weighing up.
The questions that actually decide, my answers, and one column left blank. Put the same questions to whatever else you are looking at, and fill that column in yourself.
What I put on the table is the whole codebase, a licence that lets you do as you like with it, and a written answer to every one of these questions. The project started on 5 March 2026 and v0.1.0 is a preview, so you are early enough that what you say about it counts.
The six questions I get asked first.
These are the ones that keep coming back in my inbox, roughly in this order. The answers are the same ones I give in private.
Try it on your real files.
It is the only trial that settles anything: your documents, your machine, your judgement. The download page recognises your system, hands you the matching file, and tells you what is waiting for you on first launch. No account to create, no key to type in, no trial that expires.
0.2.0-preview · macOS, Windows, Linux · Apache 2.0 or MIT
Would you rather have it built for you, around your trade and your files? Have an agent built to order