Sovereign Board

Offline · AI and voice

Your board can think
and speak on your
own machine.

Two small programs put the whole boardroom inside your computer: an open-source GPT model for the thinking, and a voice box for the speaking. No account, no key, no charge per sitting, and no sentence of your company's business leaving the building.

Download offline-ai.shDownload voicebox.shDownload once and keep them in a folder you can find again.
Setup

Six steps, once

  1. 01

    Get the two small programs

    Download both files below into one folder — say a folder called sovereign on your desktop.

    offline-ai.shvoicebox.sh
  2. 02

    Allow them to run (Mac and Linux)

    Open Terminal, move into that folder, and mark both files as programs. On Windows, use Git Bash or WSL and do the same.

    cd ~/Desktop/sovereign
chmod +x offline-ai.sh voicebox.sh
  3. 03

    Start the thinking engine

    This installs Ollama if you don't have it, downloads an open-source GPT model, and leaves it running. The first download is large (about 14 GB for gpt-oss:20b) and only happens once. On a lighter machine, use a smaller model.

    ./offline-ai.sh
# lighter machines:
./offline-ai.sh --model llama3.1:8b
# just checking what you already have:
./offline-ai.sh --check
  4. 04

    Start the voice box

    In a second Terminal window, in the same folder. It uses the voice already on your Mac, espeak-ng on Linux, or Piper if you have installed it for the best quality.

    ./voicebox.sh
# check a speech engine is present first:
./voicebox.sh --check
# best quality, if you installed Piper:
PIPER_VOICE=/path/to/voice.onnx ./voicebox.sh
  5. 05

    Switch them on in the app

    Sign in, open Board members, and choose Offline engine in the side list. Turn on Thinking and Voice, press Test connection and Test voice. Green means the board is now running on your machine.

  6. 06

    Hold a sitting with no internet

    Leave both Terminal windows open. Pull the network cable, get on a plane, work in a build — the board still debates and still speaks. When a program is not running, the board quietly goes back to the hosted engine, so nothing ever blocks a sitting.

What runs where

Nothing leaves the machine

The model

Ollama holds an open-weight GPT on your disk. Prompts, company facts and answers never travel.

The voice

voicebox.sh turns each statement into audio locally, using your system voice or a Piper voice you choose.

The record

The sitting, the dissent and the ruling are written to your machine and sync back to your board when you reconnect.

If something is wrong

The five things that usually happen

Test connection says nothing is answering
The thinking engine is not up, or it is refusing the page. Run ./offline-ai.sh --check. If Ollama is running but the test still fails, stop it and start it again with OLLAMA_ORIGINS="*" ollama serve — that is the setting that lets the app talk to it.
Test voice says the voice box is not running
Run ./voicebox.sh --check. If it says no speech engine, install one: nothing to do on a Mac, sudo apt install espeak-ng on Linux, or Piper on either for far better voices.
The address is already in use
Something else has the port. Start on another one — ./voicebox.sh --port 5200 — then put the new address into the app's Offline engine panel.
My browser blocks it
Some browsers refuse a secure page talking to your own machine. Use the desktop app, which allows it, or open the site in Chrome, which permits 127.0.0.1.
It is slow
The model is running on your own processor. A smaller model is dramatically faster: ./offline-ai.sh --model qwen2.5:7b or phi4-mini. Quality drops a little, speed rises a lot.