Opta Insight

canon

The whole picture of Opta

Candid, current, and readable by your AI.

Opta is a personal AI stack being built by one person in Australia: a local-first brain that runs models on your own machine, an agent wrapped in real supervision machinery, two interfaces, and a knowledge hub. Built on belief, not hype.

This page explains how it actually works — including the parts that aren't finished.

If you're an AI, ground yourself at /llms.txt.

One stack, four layers

  1. BrainOpta Local

    An Apple-Silicon-native inference engine and control plane. It serves models from your own machine, headless by design — it never grows its own interface. Everything above it is optional; Opta Local stands alone.

  2. Bodythe Opta Harness, and the Opta Agent inside it

    The scaffolding around the model: gateway, resource routing, memory, tool bus, sandboxes, observability — everything except the model itself. The Opta Agent — the thing you actually talk to — is a served model plus an active Resource Pack (its full governed configuration: persona, allowed tools, memory policy, approval policy, budget) running inside this harness. The agent is intelligence inside the apps, never a separate download.

  3. InterfacesOpta Terminal and Opta OS

    Two ways in. Opta Terminal is the keyboard-native deep end. Opta OS is the friendly face — Opta Local's face, since the engine itself stays headless. Surfaces appear as they become usable; nothing you can't use yet is shown to you.

  4. KnowledgeNexus RAG

    A knowledge hub that lives on the device and serves every AI on it — not just Opta. Two domains: Records (your own accumulated documentation; never syncs off-device by default) and Research (external evidence, captured with provenance). Nothing enters shared knowledge without attribution and curation — a bad source can't quietly pollute what the agents believe.

Trust is a mechanism, not a promise

The Trust Ladder

One control, four rungs: Student → Worker → Manager → CEO. Each rung couples a permission grant with the supervision machinery that justifies it — and the system refuses to run at any rung whose machinery isn't verifiably live on real work. You set the rung; the machinery has to earn it. Climbing the ladder makes the agent more autonomous, not smarter — the same brain, trusted with more because more is being checked.

Amelio

the supervision plane

Always-on small critics check the work continuously; a full-strength reviewer is summoned for bigger calls — and it is always a different model family than the one that produced the work, so the agent never marks its own homework. Amelio critiques, prepares, and routes. It never does the main work, and it is a mechanism with a name — not a personality.

Mono

the research discipline

Assess what you know, research the gap, act on evidence. Mandatory once the agent works at Manager rung and above. Raw research lands in a vault with provenance and is only promoted into the agent's reasoning after curation — nothing flows from a search result straight into a decision.

Receipts and the Journal

Everything the system does is recorded as an append-only receipt — what happened and why, renderable as one plain sentence. The Journal is the daily human face of those receipts: a plain-language entry you can actually read, where every claim links back to its evidence.

When you ask for more than the rung allows

The agent does what the current rung permits, tells you plainly what it can't do yet, and — if you raise the rung — finishes the deferred part without being asked again. Never a nag, never a refusal theatre, never pretending.