Hearth - your own AI, on your own hardware

A private, grounded AI agent that runs entirely on a machine you control. No cloud account. No API key. Nothing you type ever leaves the building.

Most "AI assistants" are a thin wrapper around someone else's server - your questions, your documents, your private notes all sent to a company that can change its terms, raise its prices, read your data, or shut you off. Hearth is the opposite: the weights live on your machine, inference happens on your hardware, and your knowledge base never touches a third party.

What you get

• A grounded agent. Answers tied to your material - your notes, docs, research - not a model guessing from memory. It cites what it found and refuses to invent facts it can't support.
• A one-command pipeline. ./hearth.sh takes you from raw notes searchable private memory a chat session. One entrypoint, no plumbing.
• Total privacy by construction. No telemetry, no phone-home, no API key. Pull the network cable and it still works.
• Bring your own model. Tuned for small fast open-weight models, but every model is a swappable default - point it at a bigger model on your hardware and re-run. Never locked in.
• Hands on the web. The agent can act: a local browser it drives itself - navigate, read, click, fill, submit - against a real Chromium on your machine (CDP bound to localhost only). It stops at any login or 2FA for you to step in, and never sends a keystroke off the box.
• A built-in self-check. A 17-point test suite proves the install is sound: deterministic memory, integrity-checked storage, grounded answers, working tools. Run ./hearth.sh gate any time.
• Readable, auditable code. Pure-Python core, plain bash runner, no opaque binaries. Read every line before you trust it.

What it's for

• A private research assistant over your own document pile.
• A knowledge base for a team that cannot send its data to the cloud - legal, medical, financial, anyone under a compliance regime.
• A sovereign starting point for builders who want a real local-AI agent they fully own, not a SaaS subscription.

Requirements

• A machine with Ollama installed (free, one-line install - Linux / macOS / Windows).
• Python 3.10+. ~5 minutes for the first run. No GPU required for the demo; a modest GPU makes larger models comfortable.

Pricing - buy it once, own it forever

No subscription. No per-seat metering. No upsell. You buy the kit, you own the kit.

One-time. Source-available - read and modify every line before you trust it.

How to buy (about five minutes) - pay in crypto, no account

1. Email themidgardcovenant [at] protonmail [dot] com with the tier (Solo / Team / Sovereign) and the coin you'll pay in. You get back a receiving address and the exact amount, locked for 30 minutes.
2. Send that amount from your own wallet.
3. Reply with your transaction id. (For Monero, also include the tx key - your wallet's "Show transaction key" - which privately proves payment and gives up nothing else.)
4. We confirm on-chain and email you the kit directly, attached to the reply. Usually well under an hour. No account, no link to expire, nothing to host.

We accept: USDC / USDT (on TRON-TRC20 or Ethereum-ERC20 - we quote the exact network), Bitcoin, Ethereum, and Monero (most in the spirit of the product). A bank/processor would force a legal identity between you and a tool whose whole point is sovereignty - crypto keeps it peer-to-peer.

Refunds: 14 days, no questions. A source-available product earns trust by standing behind itself.

How it works (the honest version)

1. Ingest - your notes (a simple JSON file) become a clean, deduplicated corpus.
2. Embed - each passage becomes a vector in an integrity-checked local file. Corrupt one byte and it refuses to serve rather than lie.
3. Retrieve - your question finds the most relevant passages by meaning, not keywords.
4. Ground - those passages are handed to a local model, which answers from them and cites them.

No step in that chain depends on a company you don't control.

1
2
3
4
5

ollama pull mxbai-embed-large
ollama pull qwen3:8b
./hearth.sh doctor
./hearth.sh ingest sample_bookmarks.json
./hearth.sh ask "what does this knowledge base say about local AI?"

Own your hardware, own your model, own your data. That's the whole idea.
Contact: themidgardcovenant [at] protonmail [dot] com