Why I'm building Nova

I want a Jarvis. Not the cinematic genius butler — the realistic version. A piece of software that lives on my desktop, listens when I want it to, talks back, holds context across a working day, and reaches into the parts of my life that actually matter to me. Calendar. Notes. Code. The market. My own writing.

The frontier model labs all sell something close to this. I have a Claude Max plan; I have a ChatGPT Plus plan; I have a Gemini key. Each chat window is an island. None of them remember last Tuesday. None of them know what tickers I trade or which Cursor project is currently broken. The intelligence is real, but the surface area is hostile.

So I'm building my own. I'm calling it Nova.

What I want it to do

The non-negotiables, in order of priority:

• Voice as a primary interface. Push to talk and hands-free. I want to ask a question while my hands are full and get an answer back through the speakers. No typing tax for the small stuff.
• A presence orb. Visual feedback that the thing is alive. Idle, listening, thinking, speaking, error. If I can't tell at a glance which state it's in, the interface has failed.
• Pluggable brains. Anthropic, Gemini, Ollama, OpenRouter — I should be able to flip the model dropdown mid-conversation and the only thing that changes is the answer. The orchestration is mine.
• A skills system. I don't want to build "an agent that does X." I want a framework where each capability is a plugin: calendar skill, vault skill, screener skill, portfolio skill. New tools should snap in.
• Memory that's mine. An Obsidian-backed markdown vault, watched live. Notes I take in Obsidian show up in chat. Notes Nova writes show up in Obsidian. The data is on disk, in plain text, in a folder I control.

What I'm not building

Not a SaaS. Not a startup. No team. No login, no billing, no analytics. This is single-user software that happens to be ambitious — closer to a tricked-out IDE than a product. If it ever ships to anyone else, it'll ship as an installer with my opinions baked in. But that's a problem for later.

I'm also not building a foundation model. The interesting work in 2026 isn't training a smarter brain — the brains are already smart. The interesting work is wiring the brain into the rest of your life without a hundred copy-pastes.

Why now

Three things converged. Anthropic's tool-use story finally became something I'd trust with my real data. Vercel's AI SDK matured to the point where swapping providers is a one-line change. And Electron — for all its sins — is still the fastest way to ship a polished desktop app on Windows that talks to native APIs.

The model labs aren't going to build me what I want. They're going to build the 90% case for everyone, and the 10% that's actually mine is going to keep getting punted. Fine. I'll build the 10%.

This blog

I'm going to write up the build as I go. Not a tutorial — more of a logbook. Decisions I made, fights I lost, things I learned about Electron's child-process model that I never wanted to know. If you're building something similar, hopefully a few of these save you a Saturday.

Next up: the stack picks. Why Electron + Next.js + Vercel AI SDK, why pnpm workspaces, and why I started with a monorepo on day one even though it was overkill.