Technology

The layers personal intelligence needs.

Aura is not a collection of apps. It is a stack: presence, intelligence, memory, perception, a model, a runtime, and verification around all of it. Scroll, and follow one request down through the layers and back up to you.

  1. 01KIRAIn Atlas

    You ask.

    Out loud or typed. KIRA hears you and shows that Atlas is listening. Intelligence should not feel like another dashboard.

  2. 02ATLASPreview

    Atlas works out what you need.

    An answer, a live search, a visual, an edit. It plans the work before any model runs.

  3. 03TRACEIn development

    Trace remembers.

    What you saved, and why it mattered. Your own context, brought back with your permission.

  4. 04OMNISCOPEResearch

    OmniScope looks.

    If there is an image or footage, it sees it. It can tell a letterbox bar from a genuinely dark scene.

  5. 05NUCLEUSResearch

    The model reasons.

    Today Atlas runs open-weight Qwen 3.5 models. Nucleus is our research into smaller, sharper ones. Bigger is not the only path to better.

  6. 06BARX 2.0Shipping in Atlas

    Barx runs it on your chip.

    It reads your machine and picks the model size, memory, threads and GPU layers that fit right now. The decision takes under a millisecond.

  7. 07GLASSBOXLite in Atlas

    GlassBox checks, then you get the answer.

    Each claim is checked against its sources and marked supported, partial or unsupported. You see the verdicts, not just a confident tone.

Barx 2.0 · Shipping in Atlas Runtime

Same app. Different machines.

Every time Atlas answers, Barx reads the machine: memory, pressure, cores, GPU, disk, power. Then it picks the model, context, threads and GPU layers that fit. Try it. This is the actual Barx policy, running in your browser.

Pick a machine
Model qwen3.5:9b
Context
8192
Threads
10
Stays warm
30m
Video
2160p
Model layers GPUCPU

    The same policy has Python and PowerShell ports for Windows and Linux, parity-tested at 42 of 42 identical plans. The machine presets here are examples.

    GlassBox Lite · in Atlas Verification

    See why it said that.

    Every Atlas answer passes through GlassBox. It pulls out the claims, then checks each cited claim against the exact source text the model read. You get a verdict and the evidence, not just a confident tone. It runs in milliseconds with no model call.

    • Supported
    • Partial
    • Unsupported
    • Uncited

    Matching is lexical today, so a faithful paraphrase can read as partial. We would rather show you that than hide it.

    QuestionHow far away is the Moon, and is it moving?

    The Moon orbits at an average distance of about 384,400 km from Earth [1]. It is slowly drifting away, by about 3.8 cm each year [2]. The drift is caused by tides transferring Earth's rotational energy to the Moon [2]. In a few million years it will be too far to cause solar eclipses.

    In development and research

    What we are building next.

    We label everything by what it is today. A product is something you can use. Research is a direction we are investigating.

    Trace icon

    In development

    Trace

    An on-device memory layer. Thoughts, voice, screenshots, photos and links, kept with why they mattered and brought back when they help. Processed on your device where possible, synced through your own platform account where needed.

    Research

    OmniScope

    Visual intelligence. Our first problem: fill a 16:9 screen with a 2.39:1 film without cropping a genuinely dark scene. It samples small patches about every 0.7 seconds and stops the moment real image appears in the bars.

    Research

    Nucleus

    Our foundation model research: efficient reasoning in small models that run on CPUs and Apple silicon without a server. We will publish results when they are benchmarked and reproducible.

    Models & credits

    Built on open shoulders.

    Atlas runs open-weight models locally through Barx. We name them, because trust starts with telling you what is inside.

    ComponentUsed forLicense
    Qwen 3.5 (4B, 9B)Language model, run locallyApache 2.0
    OllamaLocal model serverMIT
    Z-Image Turbo via mfluxLocal image generationApache 2.0 / MIT
    Whisper.cppSpeech recognitionMIT
    Kokoro-82MNeural voiceApache 2.0
    FFmpegVideo and audio processingLGPL / GPL
    Three.jsWorlds and visualsMIT

    Qwen is a trademark of its owner. Aura is not affiliated with or endorsed by the Qwen team.