Sovereign AI · fair source · self-hosted
One AI that runs on your machine, thinks with models you choose, remembers with receipts, and proves you can leave. No cloud required. No account. Nothing to trust but code you can read.
⬡ Zero-dependency core · fair-source licensed, MIT after two years · your data never trains anyone's model.
NewFrontier-class open models — gpt-oss-120b, Qwen3.6, Gemma 4 — locally, on one NVIDIA gaming GPU plus the RAM in your box, via FreeToken.What's supported today →Every other "private" AI
Your data lives on their servers, behind their login, under their terms. "Private" means they promise not to look. Leaving means starting over — your conversations, memory, and tuned behavior stay locked in the account you're trying to close.
⬡ SovereignAI
Runs on your hardware. One file, zero dependencies, readable code. The database sits in a folder you can copy; everything exports into one checksummed file whenever you want it. Ownership isn't a slogan here; it's an exit path we test.
The test · steal it, apply it to anyone
Any product claiming to be private AI can be measured in three questions, no trust required. Watch which one produces the subject change.
Can I read every line that runs?
Ours: yes — the core has zero runtime dependencies, so the audit is finishable rather than a gesture at a dependency tree.
2Can I take everything out, verified, in a documented format?
Ours: yes — one checksummed export in a documented format, and a verify command that checks the archive without importing it.
3When something isn't private, does the product tell me at that moment?
Ours: yes — the exact outgoing bytes are shown before any remote send, and every remote answer carries a receipt of what left.
The full test — how each question gets dodged, and our own scorecard including the answers we can't give yet — is at /three-questions. Quote it freely, with or against us.
Supported today · shelf dated August 2026
What thinks, what it can read, and where it reaches you — as of this month. Every chip below ships in the current build; the ones marked preview or experimental say so on the chip, and nothing on this shelf is a promise.
Curated by job, licenses printed, sized against your RAM — and now your GPU — before you download a byte. The sparse tier (20B–120B total, 3–5B active per token) runs on one NVIDIA gaming GPU plus host RAM through FreeToken.
Chosen per persona, switched in minutes. Local by default; a remote model is disclosed on screen at the moment it's used.
Parsed on your machine by dependency-free parsers; nothing is uploaded to be read. Distill history into memory on your say-so, each fact citing the conversation it came from.
Every channel reads one store only you hold. Nothing syncs through a cloud, because there isn't one.
The whole app is one file — runtime, interface, database engine. Works offline. Reports nothing to anyone.
Audited claim by claim in the Sovereignty Ledger — where an unknown stays unknown.
New this month: the customs declaration — what leaves your machine, shown before it leaves and receipted after; FreeToken as a recognized local engine; the frontier sparse-MoE tier, Qwen3.8-27B and Nemotron 3.5 Lightning on the starter shelf; GPU-aware sizing; and a model's reasoning streamed live but never written down. Preview means built and tested against the provider's API, not yet run on a live bill; experimental means it works and we're not yet proud of it. The shelf carries its date because shelves go stale. Run it now →
Seven days · one repossession
Forty companies hold pieces of you — your chats, your receipts, your decisions, your patterns. The week below is you taking them home, one recovery at a time. The story is an illustration — the receipts are not: every mechanism named under a moment ships today, none of it is roadmap.
One downloaded file is the whole thing — runtime, app, interface. It asks what to call itself and gets to work. There's no account screen, because there's no one between you and it.
⬡ single binary · zero runtime dependencies
Your ChatGPT and Claude exports drop straight in. It reads them on your machine and distills what matters into memory — by the second coffee it greets you with what it learned, citing the conversation that taught it each fact.
⬡ sovereign import-chat --distill
You feed it a mail export. Out come the subscriptions, renewals, and receipts hiding in there — including the ones you forgot you're paying for. Parsed locally, nothing uploaded, and every finding links the email that proves it.
⬡ sovereign import-email · subscription audit · renewals radar
The PDF goes in; you ask what you actually agreed to. Every answer arrives with the source passage attached, so you're checking a claim, not trusting a vibe.
⬡ local knowledge base · sources attached to every answer
A hard problem earns a big rented model for the afternoon — while the only thing allowed to write your memory is a small local model you own. The receipts name which model wrote what.
⬡ the “cognition stays home” switch · the cognition role
Chat, memory, documents, search — all of it runs with the Wi-Fi off, because all of it lives with you. The passenger next to you is arguing with a spinner.
⬡ local open weights · offline BM25 retrieval
The same memory and knowledge answer in VS Code, JetBrains, and your browser — any assistant that speaks MCP reads the one store only you hold, instead of keeping its own copy of you.
⬡ sovereign mcp · VS Code, JetBrains & browser extensions
Not hidden, not flagged deleted in someone's database — overwritten on your own disk, bytes zeroed. Some things shouldn't be remembered, even by you.
⬡ SQLite secure_delete · playable below
New open weights land in the news; minutes later they're serving your chat. Your memory, personas, and documents don't move an inch — and every reply names the exact weights that produced it.
⬡ per-persona models · weight-digest receipts on every reply
One command folds everything into a checksummed file; another proves it intact without even importing it. You're not staying because leaving is hard. You're staying because it isn't.
⬡ sovereign export --encrypt · sovereign verify
That's week one. The years after are the point — memory that compounds somewhere no vendor can reach it. Don't take the story's word either — press the buttons ↓
Operate it, right here
Three of the product's core gestures, replayed on this page exactly as they behave after install. No servers were contacted in the making of these moments.
⬡ Strike a memory
Watch the bytes zero out. In the product, secure_delete does this to the real ones. Deletion is deletion.
⬡ The first five minutes
chatgpt-export.zip ready — parsed on your machine, nothing leaves it.
Import → distill → a greeting grounded only in what was actually learned. Your AI knows you before lunch.
⬡ The exit ritual
The whole workspace, checksummed, yours.
Export, verify, leave — the drill every claim on this page answers to.
What you get
Owning an AI means owning the stack it stands on. Each layer below is a concrete guarantee, audited claim by claim in the sovereignty ledger that ships with the repo — where an unknown is marked unknown, never rounded up.
The complete map — twelve things that are yours here, each one shipped, none of them a promise:
Everything inside the wall is yours. Touch a district — or note the open wall at 12, where the boat is always waiting.
Counterintuitively, the ability to leave is the strongest reason to trust us.
— the whole product philosophy, in one line
sovereign verify backup.json proves an archive is intact, any time, on any machine.Open trial · fair-source licensed
The trial is the product — the real thing on your machine, not a demo with scripted answers. No account, no sign-up: your trial lives in a Docker volume you own, and deleting that volume deletes every trace of you. That's the exit working before you've even committed.
Have Docker? You're about two minutes away:
docker run -d --name sovereign -p 127.0.0.1:4321:4321 -v sovereign:/state --add-host=host.docker.internal:host-gateway -e OLLAMA_BASE_URL=http://host.docker.internal:11434 -e SOVEREIGN_TOKEN=pick-a-long-secret ghcr.io/mlmrx/sovereignai:latest
Then open http://localhost:4321/#token=pick-a-long-secret — with the
secret you chose — and the guided setup takes it from there. Drop in a ChatGPT or Claude export and
your AI knows you within minutes.
⬡ Step three, same day: take your first export. Settings → Data
hands you the whole workspace as one checksummed file (on an installed copy,
sovereign export --encrypt). It's the best advice we've been given about our own product:
do it before the AI knows anything worth keeping, so you're never trapped — not by us, and not if this
project someday stalls. A vendor should show you the way out before asking you to stay.
Bring a model, your dial: the command already points at an Ollama running on your
machine — that's the host.docker.internal part — or skip Ollama and use your own API key
for a frontier model in setup, disclosed on screen whenever it's used. Done trying?
docker rm -f sovereign && docker volume rm sovereign — gone means gone.
One file for Windows, macOS, or Linux from the Releases page — runtime, app, and UI inside. Make it executable and run it.
Windows: irm https://raw.githubusercontent.com/mlmrx/SovereignAI/main/scripts/install.ps1 | iex.
macOS and Linux: curl -fsSL https://raw.githubusercontent.com/mlmrx/SovereignAI/main/scripts/install.sh | sh.
Needs Node 22. Docker Compose if you prefer containers.
sovereign byoc deploy --host you@your-box puts a hardened instance on any Linux machine with SSH
and Docker — health-checked upgrades, export-to-owner, verifiable delete.
The source is on GitHub, the binaries are on its Releases page, and zero runtime dependencies means the audit is actually finishable. Someone else running the ops? Ask about the managed edition below.
Fair source · FSL-1.1-MIT · MIT after two years
Every line that runs is there to read — and with zero runtime dependencies, the audit is one you can actually finish. Found something? Open a pull request. Contributions ship under the Developer Certificate of Origin: sign your commit, no CLA to sign.
Binaries for Windows, macOS, and Linux are on the Releases page. Prefer someone else running the ops while you keep the boundary, the export, and the exit? Ask about the managed edition — a person answers.
Managed edition: hello@mysovereign.ai — we store nothing you don't send.