Questions about running your own AI

What is SovereignAI?

SovereignAI is a private AI command center that runs entirely on hardware you control. It is one file with zero runtime dependencies, needs no account and no cloud, and works offline. It combines chat, a memory system where every fact carries provenance, document knowledge, and imports of your history from other AI assistants.

How do I try SovereignAI right now?

One Docker command runs the real product on your machine — no account, no sign-up. Run:

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 using the secret you chose, and the guided setup takes it from there. The command already points the container at an Ollama running on your machine; no Ollama means you can use your own API key in setup instead. Deleting the container and its volume deletes every trace of the trial.

Does SovereignAI work offline?

Yes. With a local model running on your machine, the entire product works with no internet connection: chat, memory, document search, imports, and exports. Keyword search always works offline; semantic search additionally requires a local embedding model.

Can I see what leaves my machine before it does?

Yes, and exactly: before any message goes to a remote provider, SovereignAI shows the system prompt, the memories placed in it, the retrieved document excerpts, the prior messages, and your new message, each expandable, with sizes and the destination host. You can send, cancel with nothing sent, or stop asking for that provider and revoke that later. The preview is assembled by the same code as the real request, so it cannot differ from what is sent. Local endpoints never ask because nothing leaves, and every answer from a remote provider carries a receipt of how much left. There is no classifier deciding what counts as sensitive; you see the actual content.

Is SovereignAI open source?

No, and we do not claim to be. The core is fair source under FSL-1.1-MIT. You can read the code, self-host it, fork it, and use it internally without restriction; the only barred use is launching a competing commercial SovereignAI offering. Each release converts to the plain MIT license two years after it ships, so the code outlives the company.

Which AI models can I use with it?

Local open-weight models are the default, through Ollama or any OpenAI-compatible server such as llama.cpp or vLLM. Frontier-class sparse mixture-of-experts models — gpt-oss-20b and gpt-oss-120b, Qwen3.6-35B-A3B, Gemma 4 26B-A4B — run locally too, on one NVIDIA gaming GPU plus host RAM through the FreeToken engine, which SovereignAI recognizes, health-checks, and sizes for your machine. You can also use a frontier model with your own API key. Models are chosen per persona and swappable in minutes, and when a remote model is used the disclosure appears on screen at the point of use.

Can I import my ChatGPT, Claude, or Gemini history?

Yes. SovereignAI imports the data exports those services are legally required to give you, plus a documented generic format for anything else. Imports are idempotent, so re-running them does not duplicate data, and every imported record keeps provenance pointing back to its source conversation.

How do I know my data is really deleted?

Deleting a memory zeroes the bytes on disk rather than hiding a row, using SQLite secure_delete. Because there is no cloud copy and no hosted control plane, there is no second location holding what you deleted. Write-ahead log frames are recycled by the database engine over time, which is noted in the code rather than glossed over.

How does the export work, and can I verify it?

One command writes everything to a documented format with a SHA-256 manifest per table plus a digest over the whole archive. A separate verify command checks an archive without importing it, so you can test a year-old backup without touching live data. Exports can be encrypted with AES-256-GCM under a scrypt-derived passphrase that never leaves your machine, and import refuses an archive whose checksums do not match.

What hardware do I need to run a private AI?

The application itself is lightweight and runs comfortably on an ordinary laptop; the model decides your requirements. A small open-weight model in the 3-to-8 billion parameter range runs on most modern machines with 16GB of RAM. Sparse mixture-of-experts models reach further: through the FreeToken engine, an NVIDIA RTX 30-series or newer GPU plus 32GB of RAM runs the 20-to-26-billion-parameter class such as gpt-oss-20b, and 128GB of RAM runs gpt-oss-120b. SovereignAI checks your available memory — and your GPU — and recommends models that will actually fit before you download them.

How is this different from just running Ollama myself?

Ollama runs models; it does not give you memory, provenance, knowledge, or an exit. SovereignAI is the layer that turns local inference into something that knows you over time: durable memory where every fact records which model wrote it and which conversation it came from, local document parsing and search, imports of your history from other assistants, and a verifiable export. SovereignAI uses Ollama as one of several inference backends.

How is this different from Perplexity's Portable Computer?

Portable Computer runs Perplexity's agent on an NVIDIA DGX Spark today (RTX PCs to follow, Linux first) and asks before sending a step to a cloud model — an instinct we share. It still needs a Perplexity subscription and account, its Gmail, Drive, Slack and GitHub connectors sign in on your behalf without saying where those credentials live, and it does not say where your memory lives, how you export it, or whether you can read the code. SovereignAI needs no account, holds no credentials, runs on the laptop you already have (small models anywhere, frontier-class sparse models on a 4 to 8 GB GPU plus RAM through FreeToken), records where every memory came from, shows exactly what leaves before any remote call, and exports everything into a documented, checksummed file. They are also different jobs: theirs is an agent harness for tasks; ours is the memory, knowledge, and exit you own underneath whatever you run.

What are SovereignAI's limitations?

The database is not encrypted at rest, so the honest answer today is operating-system full-disk encryption. The model layer is borrowed: even open weights are a third party's artifact trained on data nobody can audit. Rented GPU compute is tenancy rather than ownership, and those trade-offs are printed before you provision. A full layer-by-layer self-audit is published as the Sovereignty Ledger.

Does SovereignAI train on my data or send telemetry?

No. There is no telemetry, no analytics, and no hosted control plane that your instance reports to. Your data never trains anyone's model. When you choose to use a remote provider for a message, that provider's terms apply to that call, and the product tells you on screen at the moment it happens.

What does SovereignAI cost, and how do I get access?

Self-hosting the core is free: an individual never has to pay to own their own AI. The open trial is one Docker command — the real product, no account. The source is on GitHub under a fair-source license, single binaries are on its Releases page, and the managed edition is by request. Commercial offerings come by addition, such as a managed edition and organization features.