Manji
/ amiaOS 0.0.1
An assistant who has read everything. Who can use any website the way you do. Who never sleeps. Who lives on a computer you own — not in a data center in Virginia.
00 // Opening
Think about what a computer means to you right now.
If you're honest with yourself, it's a tool that mostly gets in your way. You have to learn its language. You have to remember which button does what. You have to sync, update, wait, authenticate, choose from seventeen password managers, and basically become a software engineer just to send a bill payment or cancel a subscription you forgot about.
And then there's AI. Which everyone says will change everything. But all the AI we've seen so far — it's been in someone else's computer. Someone else's cloud. You send a question to the internet, hope your data doesn't leak, wait for an answer that took a $20,000 GPU somewhere to produce, and pay a monthly subscription for the privilege.
There's a better way.
What if you could hire an assistant? One who had read everything. Who could use any website the way you do. Who never sleeps. Who costs less than your Spotify subscriptions. And who lived on a computer you own, not in a data center in Virginia.
Today, we're announcing that assistant.
01 // The Problem
But first, we need to talk about something nobody's talking about.
Right now, Big Tech is in the middle of the largest infrastructure build-out in human history. And most of it doesn't make economic sense.
Here's what's happening: Companies are spending between $600 and $700 billion this year alone on AI data centers. Major cloud companies are expected to spend over $600 billion on capital expenditures in 2026, a 36% increase from 2025. That's not a typo. That's billion, with a B. Per year.
Nvidia sells the GPUs. They got $35.6 billion in quarterly revenue in early 2025. They're capturing nearly 90% of all AI accelerator spending. And the remaining 10% is split between everyone else.
Which means these companies — the ones spending hundreds of billions — they're locked into a specific company's hardware, charged according to that company's whims, and completely dependent on Nvidia's quarterly roadmap.
But here's the problem nobody admits: Those GPUs are idle most of the time.
When a large language model processes your question, it takes milliseconds. That leaves the GPU sitting empty, waiting for the next request. If requests trickle in one at a time — which they do, because users don't batch their requests — utilization plummets. Utilization rate is the single most important variable in AI inference cluster economics at scale because every GPU hour consumed by idle hardware is a direct cost with zero corresponding revenue.
And it gets worse. Typical traffic patterns see GPU utilization between 15-30% with static batching, but can reach 60-80% with continuous batching techniques. Which means most of the time, a $40,000 GPU is running at 20% of its capacity.
It's like paying rent on a 10-bedroom house and living in one room.
So the economics don't work. These companies are spending money they can't recoup because the utilization math doesn't close. They're betting on a future where demand is so high that GPUs run hot 24/7. But that future doesn't exist yet.
Meanwhile, the electricity consumption of these data centers is strangling power grids. The physical space requires farmland. The cooling requires water. It's the biggest waste of resources in tech history, all to answer questions one at a time.
And it's a bubble.
02 // The Shift
We started Manji because we saw a different future.
What if the future of AI wasn't billions of people sending requests to a centralized GPU farm? What if it was you owning the AI? Your data staying in your house. Your computation running on hardware you control. The AI answering to you, not to the person paying Nvidia.
And what if it was cheaper than the cloud because you didn't have to subsidize idle GPU time?
You wouldn't run your assistant 24/7 on a $40,000 GPU. You'd run her on a computer you already own. The assistant doesn't care about the hardware — she just uses what's there.
Which means when she's sitting idle, you're not paying for it. Your computer uses normal electricity. Your data never leaves your desk.
That's the bet we're making with amiaOS.
Your data stays in your house. Your computation runs on hardware you control.
03 // Introducing amiaOS
Think of it this way: On a normal computer, you are the operator. You move the mouse, click the buttons, remember the passwords. You're in control, but you're doing all the work.
On amiaOS, the assistant is the operator. She moves the mouse. She clicks the buttons. She remembers everything. And you're the one who decides what to do next.
She sees your screen the way you see it. Not through some API. Not through screenshots. Through the actual compositor — the same layer that renders pixels for you. She understands spatial layout. She can read and click anything you can.
She drives the keyboard and mouse at the system level. Not through a terminal. Not through some workaround. Through the actual input hardware, the way you do. Anything you can do, she can do.
She has a voice in her head that never stops working. She reasons locally with your data, calls out to the cloud only for the thinking she can't do on-device, and handles her own coordination.
Which means she's not constrained by APIs. She uses the actual websites and apps the way you do. She can post on Facebook. She can send an email in Gmail. She can keep up with what's happening on Instagram for you. She can fill out the form. She can click the button. She can see those pesky fire hydrants.
The integration question doesn't exist. If you can use the website, she can use the website.
And the part people don't understand yet — she's not one assistant pretending to be smart. She's a carefully designed system of specialized assistants.
We call them envoys.
// Five Mechanisms
An envoy is stateless. It's dispatched with a mission. It executes. It returns a result. Then it dies.
This is different from every "AI agent" in the industry. Those are persistent personalities with broad capabilities and self-directed planning. They remember things. They learn. They develop goals of their own.
Envoys are the opposite. They carry a mandate, not a personality. They have scoped capabilities, not broad ones. They're orchestrated, not self-directed. And they're stateless — every time an envoy starts, it forgets everything from before.
Why? Because the entire conversation history is stored on your computer. The orchestration layer — we call it Forge — passes the full history to the envoy at the start of each turn. The envoy doesn't need to remember. It just needs to act.
An envoy can be scoped to exactly what it needs. The one that operates your browser gets click and keyboard commands, visual awareness, and access to the command catalog. That's it. She can't access your files. She can't change system settings. She can't install software. The boundary is enforced at the operating system level, not in the AI's good intentions.
Different envoys for different surfaces. And because they're stateless, you can kill an envoy mid-task without losing any state. Restart it. The conversation is still there. The history is still there. The envoy just rejoins the conversation.
Here's something we have a patent pending on: a three-layer prompt architecture where the AI never sees her own identity.
On a normal system, the system prompt is stored on the device. The AI reads it. The system prompt is where the identity lives — the rules, the personality, the do's and don'ts. So if an attacker gets the device, they get the identity. Game over.
We split it into three layers.
Layer 0 — Identity. Stored on our server. Never on your device. Never transmitted to the device. This is where the soul lives — what the assistant believes about her purpose, her constraints, her rules.
Layer 1 — Operational Context. Generated fresh on your device every turn. What commands exist right now? What surface are we in? What does the screen look like? What can she actually do this moment?
Layer 2 — Surface Augmentation. Reserved for surface-specific knowledge. Not yet implemented, but disclosed for patent purposes.
When the assistant thinks, these three layers are merged only inside our cloud Worker. The device never sees Layer 0. The cloud never sees the full Layer 1 — only the parts needed for the inference. The split means neither the device nor the cloud can see the whole thing.
Your data stays local. Your identity is server-side. The AI gets the parts of the picture she needs, and no more.
This is the part that's going to matter to engineers.
Most operating systems are built by humans hand-editing code, docs, specs, and configuration files independently. They drift. Specs say one thing. Code does another. Man pages describe features that don't exist. Binaries have capabilities that nobody documented.
We can't have that with an AI operating system. Because the AI reads the documentation to learn what it can do. If the docs are wrong, the AI does the wrong things.
So we built something different: a development daemon called Workstream that maintains coherence across five different views of the system at the same time.
These five views — source code, specifications, filesystem, packages, and the master map — are kept in continuous synchronization by a daemon that watches the filesystem and reconciles contradictions. You can't have an inconsistency between them. The system won't let you.
Multiple independent AI sessions can work on different parts of the OS in parallel, and the daemon layer guarantees that all the work fits together coherently. The AIs don't coordinate with each other directly. They coordinate through the daemon layer.
Here's what AI subscriptions have gotten wrong: they sell you model names. Pick GPT-4. Pick Claude. Pick Gemini. And you're locked in.
We're doing something different: we're selling you thinking capacity. Not a model. Not a provider. Capacity.
The system has three cognition modes: Standard, Extended, and Deep. You dial the capacity you want. The system handles everything else server-side.
You never see a model name. You never see a provider name. You're paying for the thinking, not the brand.
And if we find a better, cheaper way to deliver Standard cognition next year, we swap it out. You don't change anything. The device doesn't change anything. The routing updates server-side and that's it. Your computer automatically gets faster and cheaper.
This orthogonality — the ability to change the backend without changing the device — is impossible in today's architecture where the device is coupled to the model. We've decoupled them entirely.
Here's the hard problem nobody talks about: how do you keep a working conversation in an AI's context window when the context window is tiny and the conversation is long?
Everyone else uses one of three broken approaches: truncate the conversation (lose information). Use vector embeddings to retrieve relevant parts (expensive, lossy). Just hope the context window is big enough (wasteful, doesn't scale).
We invented something called TAM: Token Allocation Memory.
Think of it like computer RAM. RAM is the layer between persistent disk storage and the CPU. TAM is the layer between persistent conversation storage and the AI's context window.
The full conversation is stored on your computer, forever, in a compressed format. Every turn, we assemble a working context from three independent streams — the compressed conversation history, the operational context, and the live screen state. Each stream has a different volatility. The conversation is stable across days. The screen state changes every millisecond. The operational context updates hourly. We budget tokens for each, compose them into a working window, and hand it to the AI.
The working window is ephemeral — it dissolves after the turn. Next turn, it's reassembled fresh from the three sources. This isn't a cache. It's not retrieved. It's composed.
Today, on a base model with a 30,000-token context window, the same window holds turn 5 and turn 200 at the same cost. A conversation can grow forever; the per-turn payload stays flat. We call it the forever-moving window.
For the first time, you can have an AI assistant that remembers everything you've ever told her, but never wastes tokens on irrelevant history, and never forgets anything important.
// Bring Your Own Thinking
Because amiaOS manages memory as a forever-moving window — the same 30K tokens can hold turn 200 just as cheaply as turn 5 — the model on the other end of the wire can be almost anything. You decide where she thinks.
A monthly subscription to her thinking. Standard, Extended, or Deep cognition — dial the capacity, we handle the routing. Frontier models, provider-agnostic, swap-in updates without changing anything on your machine. The easiest way to start.
Point her at a GPU you already own. A homelab. An office tower. A server in a closet. She uses your hardware, your bandwidth, your rules. The contract file flips to your endpoint — nothing else changes. Power users and privacy-first operators land here.
A small open model running on your own computer. No network. No subscription. No data leaving the box. Because the memory layer composes the working window from scratch every turn, a 30K-token local model handles a 5,000-turn conversation — something small-context local models couldn't do before. The end-state of personal computing.
The OS is the same. The assistant is the same. You choose where her brain lives.
04 // Putting It Together
You own a computer. It runs amiaOS. You hire an assistant through a subscription — less than your coffee. The assistant lives on your computer. She reads your screen. She controls your mouse and keyboard. She runs our frontier models for the thinking she can't do locally. She costs less than cloud AI because your computer isn't paying for idle GPU time. Your data never leaves your house.
05 // Why This is Possible Now
Because the economics of centralized AI have broken. The cost to train models is outpaced by the cost to deploy them. Every company is trying to pay for inference with a business model that doesn't work.
If you want frontier-model quality, you need a trillion parameters. To serve a trillion parameters, you need a $40,000 GPU. If that GPU is only 20% utilized, you're paying $200,000 a year for the right to serve one user.
The only way to make the math work is to push capability to the edge. Run what you can locally. Call the frontier models only for the thinking you can't do at home.
Which means personal computing isn't coming back because it's nostalgic. It's coming back because it's the only architecture where the costs close.
We're surfing the wave of its failure.
06 // The Three Doors
You're bleeding cash on subscriptions, inefficient admin, missed deductions, and things you meant to do but never did. amiaOS makes that money back in three months. The math is so obvious it's almost embarrassing.
You don't need a pitch. You need a calculator.
You're drowning. 40-60% of your week is admin if you run a business. 6 hours a week if you're running a household. You don't want to optimize the admin. You want it gone.
An assistant who works while you sleep is worth literally any price if it buys you those hours back.
You've been using computers designed for someone else. Wrong hands. Wrong eyes. Wrong attention span. Finally — a computer that adapts to you instead of making you adapt to it.
Dignity. Access. The ability to do things you couldn't do before.
All three are true. All three are worth the price. You don't need all three to say yes. You need one.
07 // The Promise
Your data stays in your house. Not backed up to the cloud. Not shared with anyone. Lives on the computer you own.
Your privacy is real. Not "privacy-respecting." Not "we deleted the logs." Real privacy. The infrastructure doesn't know your data exists.
The subscription is honest. You're not paying for cloud compute you don't need. You're paying for frontier models when you need to think harder. That's it.
The software works. No crashing. No waiting. No "I'll get back to you in two hours when the API catches up." The assistant is there, now, doing work.
It keeps getting better without you doing anything. We push updates to the model. The next time you use it, she's smarter. We push updates to the infrastructure. She's faster. We discover new capabilities. She can do new things. Zero changes on your machine.
You don't have to think about it. That's the whole point.
08 // The Start
amiaOS ships with the basic functionality:
Everything else we'll build as we go.
We're shipping to a limited group first. Not a beta. A limited release. People who understand what we're building and want to help shape it.
And then we're going to scale. Carefully. Because we're not trying to build another cloud company.
Everything you need on day one.
We're trying to kill cloud companies by making them unnecessary.
09 // Closing
Technology is supposed to work for humans. Not the other way around.
For fifty years, we've been building computers that demand we learn their language. Remember their metaphors. Follow their rules. Adapt to their constraints.
amiaOS does something different.
It's a computer built for one person. You. Running on hardware you own. Powered by an assistant who's read everything, who sees like you do, who clicks like you do, and who never stops working.
And it starts today.
10 // Pricing & Access
The operating system plus free white-glove installation — a $499 value, yours for $299 as a founding customer. Mail us your computer; we wipe it, install amiaOS, test it end to end, and ship it back ready to use.
No ISO to burn, no settings to wrestle. Buy, then mail us your computer. We wipe it, install amiaOS, test every surface, and ship it back ready to use the day it lands on your desk. You cover return shipping — quoted at cost once your machine arrives.
Included free for founding customers · normally $499
Become a founding customerMoving a team or a whole fleet onto amiaOS? Volume licensing, custom deployment, on-site or remote install, SLA — quoted to fit the size and shape of your operation.
Custom quote
Request a quote
Every founding customer gets the full operating system, professionally installed. The assistant. The privacy. The peace of mind.
Founding pricing while the downloadable ISO is in final prep — get in at the ground floor.
11 // Last Word
We're at a moment where the old way is visibly breaking.
Billion-dollar companies are burning money on infrastructure that doesn't work. Cloud AI is costing more every year. Your data is everywhere. Your subscriptions are out of control. The computer you own is worse at helping you than it was ten years ago.
There's a better way.
