Morax One
E
AppsMorax AI

Native macOS command post for local models

Native app

The models run local. So does the control.

A local model is fast and free, but raw: no interface, no memory of what it did, no guarantee about what it triggers. Morax AI adds the three layers a raw Ollama daemon doesn't have — a live view of the inference chain, a catalogue of reusable actions, and a human approval gate the moment an action would touch anything outside the machine.

Open Morax AI — indisponible. A release script exists in its repository, but no notarised build is published, and this hub has no access to the local one.

The only Ollama interface on the machine — a second one is deliberately not written.

The Morax AI overview on a dark background: a banner reporting the state of the inference chain, a compute-mode card naming the target model and the memory it needs, cards for unified memory and resident video memory, the chain from the NAS through Tailscale to Ollama with each port, a speed panel, and the models ready to load.
One Apple Silicon machine, the models it can hold, and the chain that reaches them. Every figure is read from the machine the app runs on.

What this product does

  • Observability

    Unified memory, resident VRAM, throughput in tokens per second, and the state of every link in the chain from NAS to Tailscale to relay to daemon, in one cockpit tab.

  • A catalogue of reusable actions

    An action pairs a source (free text, a local file, a shell command, a git diff, a remote NAS file over SSH), a model, a prompt template, and one or more destinations. Built once, run again without re-explaining it.

  • A mandatory approval gate

    Every action carries an effect level. The moment one would reach a third party — an email draft, an outbound message — it stops for a human decision. Nothing crosses that line unattended.

In detail

  • Nine tabs, one state

    Cockpit, Fleet, Actions, Playground, Comparator, Operations, Pipeline, Benchmarks, Settings — one SwiftUI app with a single observable AppState behind all nine, not nine disconnected screens.

  • State stays on the machine

    Model actions, proposals, the operation journal, NAS jobs and benchmarks are plain JSON under Application Support. No database, and this hub reads none of it.

  • Fleet and comparator

    See every installed model, its size and quantisation, load or unload it, and run the same prompt across several models side by side.

  • Built, not published

    scripts/build_app.sh produces the .app, and scripts/package_release.sh is written to sign, notarise and describe a disk image. The only build so far is signed ad hoc, so nothing is published yet — that is the next gap the distribution plan closes.

  • Not Morax AI Video

    That other page describes a face-cam generation service with no code behind it. This one is the opposite: real, running software with no public interface yet. Don't read one page as a status update on the other.

Who it is for

Triaging a long text or diff
A saved action reads it, applies the right model and prompt, and writes the result where it's needed — journal entry, file, or clipboard.
Choosing between two local models
The same prompt runs on both in the Comparator, side by side, before either is wired into an action.
An action that could reach someone else
The gate holds it until Etienne approves it by hand — nothing external is ever sent unattended.

Morax AI

One command post for the models on Etienne's Mac, gated by a human at the moment it matters. Built and running today, still without a public build.

Open Morax AI — indisponible. A release script exists in its repository, but no notarised build is published, and this hub has no access to the local one.