Three local AI cores reason independently, then synthesize one grounded answer. Offline first. Private by default. Built to be checked.
Counts what can be counted, and weighs what holds. The Analyst pins claims to evidence and refuses to round up.
Reads what is felt and what matters. The Humanist keeps the human stakes in frame when the numbers try to flatten them.
Hunts what is silenced and what hides. The Skeptic argues against the strongest emerging consensus, not against you, so easy agreement gets tested.
Like a medical conference, three specialists examine the question from different angles, then deliver one careful synthesis. Analyst checks the facts. Humanist reads the room. Skeptic asks what's missing.
Three local GGUF models, Analyst, Humanist, Skeptic, each examine your query from a different angle. Their perspectives are scored by core weights derived from query type, cross-probed against each other for weak claims, then synthesized into one answer with a parsed self-reported confidence tag for every response.
Runs on your hardware via llama.cpp. No cloud dependency. Your conversations, vault, and persistent memory never leave your device.
Local knowledge base with tiered sourcing. Cross-encoder reranking. Citations enforced as [Vault:Source] with strict-mode checking.
The Skeptic argues against the strongest emerging consensus between the other two cores, not against the user. Catches the place where smart panelists agree without checking each other.
Every response carries three confidence signals, heuristic (evidence-derived), model-stated (parsed from the synthesizer), and blended. The system is honest about how confident it has the right to be.
Most AI asks for trust. ZOTHEOS keeps receipts: signed results, a local track record, a log of every mind-change, and an offline mode you can watch prove itself. All of it lives in the product, not in a press release.
Download a signed receipt for a result and verify it offline. No account, no server, no trust required.
Forecasts the system logged, scored against what actually happened. A local track record with a reliability diagram.
When the council changes its mind, the full chain stays: v1 to v2 to v3, with the evidence that moved it. No memory-holing.
Every flagged correction is kept with provenance in an immutable log, and approved or rejected where you can see it.
Not a toggle and a promise. A network ledger, plus a canary that attempts an outbound call in front of you so you can watch it get refused.
ZOTHEOS also measures its own catch rate. The method is published; the numbers print when they are ready, not before.
Everything lives in one overlay: a fuzzy command palette, keyboard driven, arranged in three zones. A search that matches nothing offers to run itself as a fact-check, a memory search, or an investigation. No dead ends.
A selection of the thirty. The Ethics Sentinel runs underneath them all: it refuses to help build systems of control, and it always answers fact-seeking accountability questions.
You do not shop for models. ZOTHEOS measures your machine and seats the council it can honestly run, from workstation to pocket. The three roles stay the same; only their size changes.
On T0 the app says so out loud: one mind, not a council. Honesty is part of the interface.
Most AI memory tools remember for you, they keep your context, you stay passive. ZOTHEOS does the opposite. Memory awareness activates only inside Projects. Inside one, ZOTHEOS stays focused on the facets you're exploring and quietly teaches you how to organize what you learn.
A real query through the ZOTHEOS fusion pipeline. Three cores reason independently, then one synthesized answer emerges with full attribution.
Hosted on personal hardware via Tailscale Funnel · may be offline
Algorithms reward what keeps you scrolling, not what helps you understand. ZOTHEOS is a calm, reliable alternative: a tool that helps people learn and solve problems without surveillance, manipulation, or paywalls shaping the outcome.
Most assistants are products of an ad model, a surveillance model, or both. ZOTHEOS is a tool, built to one purpose.
ZOTHEOS
[Vault:Source] tags. Strict mode rejects answers that cannot be grounded.I built ZOTHEOS because I was tired of tools that treat attention as the product and truth as optional. The internet learned to keep us scrolling; I wanted something that helps people think instead.
So I gave it three minds that disagree on purpose, a memory that teaches rather than hoards, and a hard rule: it runs on your machine, and your data stays there. No investors to please, no ad model to feed. Just the work, done right.
Drop your email to get the invite ahead of public launch. We'll only email you about ZOTHEOS milestones. No newsletter spam, no third-party sharing. Same privacy posture as the product.
Support funds development, testing, and releases, and keeps ZOTHEOS permanently independent. No investors. No ads. Just the work.