Contexts
intentions, memory, files, tools, permissions
The end of the app
An operating environment for the age of AI agents. State an intention — SairiOS assembles the workspace, the tools and the team to accomplish it. On your machine. On your account. Open source.
Every other AI gives you a paragraph. This one gives you a workspace.
the core · listening
Intention
Six seconds between an intention and a place to pursue it.
The problem
One intention. Twelve applications. You copy between them, you remember what mattered, you hold it together. The computer did none of that work. You did.
SairiOS deletes the glue.
You say what you want to accomplish. The operating system builds the place to do it. When you are done, that workspace can become permanent. When it proves itself, it becomes something you can hand to anyone else.
A whole computer, with the intelligence already inside
SairiOS ships as a virtual machine image. It builds from scratch and boots to a desktop unattended — and the agents, the permission broker and the context engine are already there when it comes up. There is nothing to install, configure or connect on the first run.
intentions, memory, files, tools, permissions
specialists that persist between sessions
observation · proposal · execution
the image builds from scratch and boots
milestone 0 · experimental prototype · services are loopback-only and unauthenticated
Say it, and watch it build
This is not a loading animation. Every beat is a thing that actually happened, so you can name what the machine is doing at any frame.
working
working
working
working
working
working
0/6 beats
A different workspace every time
This is the proof it is not a chat window wearing a costume. What you asked for decides what gets built — the panels, the layout, the specialists, even the colour of the room.
Intention
Compare what three papers actually claim about scaling
12 documents, ordered by evidence weight
claims linked to the passage that made them
method · dataset · reported delta
who said what, and when
no dock · no app grid · no file browser
A team, not a chatbot
Every context comes with specialists. You see each one's role, what it is doing right now, how far along it is and what it last found — and you can pause it, redirect it, or overrule it.
Ranking 12 documents by evidence weight
last found
Two sources cite the same unverified table.
Normalising units across three papers
last found
Reported deltas use different baselines.
Attacking the emerging conclusion
last found
The trend disappears below 2B parameters.
And they remember. An agent that did good work for you is the same agent next time, carrying what it learned — visible, attributed, and deletable by you at any moment.
You can see it thinking
Early findings are wide, dim, soft-edged, still moving. As the work resolves they tighten, brighten, lift off the surface and gain weight. You can read the state of an investigation from across the room without reading a single word.
Four services still being correlated.
exploring
Three hypotheses, none ranked.
exploring
Nothing to propose yet.
exploring
layout carries the information — nothing else in software does this
Nothing happens without you
Every action arrives as a plain-language request: what it wants to do, why, and what it will actually change. You can take a permission back. Everything that happened is written down.
Request
The log correlator wants to read the last 200 traces for /checkout.
awaiting your decision
It asks. Every time. Because an operating system that acts without asking isn't yours.
Say it out loud
Hold the key, speak, let go — the words are already in the field, still editable, never sent without you. Recognition happens on your own machine; your voice never leaves it. No chime, no personality, no waiting for it to finish a sentence.
on-device recognition · no audio leaves the machine
The computer keeps running without you
Agents keep working while the screen is closed. Today you speak to the machine directly; the roadmap is a message bridge, so an intention sent from a chat app lands in the same context and comes back as a result. That bridge is not built yet, and this page will say so until it is.
Hold the key, speak, let go. Recognition runs on-device; the transcript is editable text.
Services bind to loopback. Reaching the machine from elsewhere is your tunnel, your rules.
Send an intention to the machine and read back what the agents did while you were away.
The same bridge, a second transport. Not built yet — no code for it exists today.
one core · many ways in · nothing sent without you
It's yours
No context is read to improve a model. An experimental prototype, in the open, where you can check every claim on this page against the source.
Contexts, memory and history live locally.
It runs on your model provider account, stored on disk.
There is no middleman and nothing to renew.
Open source. Early. Built in the open.
Get it
make setup && make devThe mock agent needs no credentials, no network and no external process. Add your own provider key when you want real thinking. SairiOS never bills you.