Welcome to ShogunAI

Every assistant you use starts each conversation from zero. You paste the context back in. You re-explain the project, the people, the decision you made on Tuesday. Then you copy the answer out and do the work yourself.

That is not an assistant. That is a very articulate blank page.

ShogunAI is the operating system for the AI-native individual. It has two layers, and the second one is the point.

The memory layer

ShogunAI builds a model of your work while you work. Not a folder of notes you have to maintain — a running picture of what you read, wrote, and decided, assembled passively in the background and stored on your Mac.

It captures text through the accessibility layer macOS already exposes to screen readers. It is not a screen recorder: there is no rolling video of your day, no frame taken every few seconds, no audio file. What gets stored is text and where it came from.

There is one narrow exception, and it is off until you turn it on. Some windows hand over no text at all — a chart in a shared slide, a dashboard drawn on a canvas, a PDF that is really a scan. With visual recall enabled, those windows — and only those — produce a compressed screenshot, which stays encrypted on your Mac for a retention period you choose and is deleted automatically when it reaches that age.

That distinction is the whole design, and we wrote more about it in Passive memory, not screenshots and Visual recall, and the exception we argued about.

The execution layer

Memory that only answers questions is a search box with better manners. The reason to know what happened in your day is to finish what it left open.

So ShogunAI reads its own memory and does the work: drafts the follow-up you said you would send, prepares the brief for the call you have in an hour, files the thing you forgot to file. It reaches your tools through their own APIs, and anything that leaves your machine — a sent email, a posted message, a created event — stops for your explicit approval first.

We wrote about where those gates sit in The execution layer.

Why both, why now

Three things became true at roughly the same time. Models got good enough to act on messy real-world context instead of just summarizing it. Tools grew a common protocol layer, so an agent can reach twenty of them without twenty bespoke integrations. And laptops got fast enough to run capture, indexing, and search locally without you noticing.

In 2007 the iPhone folded the phone, the camera, the browser, and the music player into one device. Memory, execution, and data sovereignty are having that moment now. They arrive as separate products today. They belong in one system.

What we will not do

Some constraints are worth stating before we ship anything, because they are the parts that get quietly traded away later:

• Text and provenance, not a recording of your day. Capture reads text rather than pixels, and no audio file is written at any point. Where a window yields no text at all, visual recall fills the gap — off by default, opt-in, never uploaded, and deleted at a retention age you choose. We wrote out every rule around it rather than leave it in a footnote.
• Your machine is the default. Memory lives locally. What goes out goes out for a named reason, and the log tells you which reason.
• Nothing sends itself. Automatic actions are the reversible, local ones. Anything another human will see waits for you.
• No free tier funded by your data. Every plan is a paid plan, starting with a seven-day full trial. You are the customer, which is the only arrangement that keeps the above three honest.

ShogunAI is in private beta on macOS. If your day is full of context that nobody but you is holding, it was built for you.