My Sovereign AI

Sovereign AI. You own every layer.

Privacy

Developer Tools

Artificial Intelligence

GitHub

Visit WebsiteSee on Product HuntGithub ⧉

Hey Product Hunt 👋 I kept noticing the same thing: the AI tools I relied on were getting better at knowing me, and that was exactly the problem. Years of conversations, preferences, the shape of how I work — all of it living on servers I don't control, in formats I can't take with me. Not because anyone is evil. Because that's the business model. The moat is the memory. Then I looked at where the rest of "me" actually lives: email, bank exports, health records, receipts, calendars, a dozen AI chat histories. You are the most fragmented database on earth. And the joins that would be genuinely useful — subscriptions against usage, renewals against your calendar — are only safe if you hold both sides. Nobody can assemble that for you safely. You have to assemble it yourself, on hardware you own. So SovereignAI runs as one file on your machine. Zero runtime dependencies — the code you audit is the code that runs. Local open weights by default; rent a frontier model when you want it, with the disclosure on screen when you do. Three constraints I refused to break: • Every memory shows its provenance — which conversation it came from, which model wrote it. If we can't say where a memory came from, we don't keep it. • Deleting deletes, down to the bytes. • The exit is a feature, not a promise: one checksummed export, encrypted with a passphrase only you hold, verifiable without importing it anywhere. The format is documented well enough to outlive us. How it evolved: the first version was a really good chat window, and that was the mistake. Chat made the data plane invisible — you couldn't see what it knew about you or where that came from. So the default view became the mind itself: memories with receipts, what arrived from where, which doors are open. The chat is now the least interesting part. Two honest notes. We're in gated early access — a person reads every request. And the core is fair source (FSL-1.1-MIT): self-host it, fork it, use it internally with no restrictions; each release becomes plain MIT two years after it ships. That's not "open source," and I'd rather say so plainly than blur it. The question I'd genuinely like answered: which layer would you need to own before you'd trust an AI with your actual life? Ask me anything.