Download

Wyrdsekai 0.1.0 — Apache-2.0

The installers are large because they are self-contained: a JVM, an inference runtime, and the embedding model travel with the package. You should not have to assemble a Java toolchain to talk to your companion.

Install

• Linux

  curl -fLO https://github.com/Wyrdsekai/wyrdsekai/releases/download/v0.1.0/wyrdsekai_0.1.0_amd64.deb
  sudo dpkg -i wyrdsekai_0.1.0_amd64.deb
  wyrd start

• macOS

  curl -fLO https://github.com/Wyrdsekai/wyrdsekai/releases/download/v0.1.0/Wyrdsekai-0.1.0.pkg
  open Wyrdsekai-0.1.0.pkg
  wyrd start

  On Apple Silicon the companion thinks locally through MLX, which wyrd setup installs for you. Intel Macs need one extra step: MLX is Apple-Silicon only and the package does not bundle an Intel inference engine, so run brew install llama.cpp first — or point the node at an API key, or at another machine in the household. Setup will tell you which case you are in rather than failing quietly.

• Windows (PowerShell)

  curl.exe -fLO https://github.com/Wyrdsekai/wyrdsekai/releases/download/v0.1.0/Wyrdsekai-0.1.0.msi
  msiexec /i Wyrdsekai-0.1.0.msi
  wyrd start

• Relay

  A relay is the door your phone comes through. It is deliberately tiny — no JVM, no models — and it never sees inside your world; it shuffles encrypted bytes between your phone and your household. If you have nowhere to put one yet, the public relay will do — but running your own is the point, and it is three commands.

  curl -fLO https://github.com/Wyrdsekai/wyrdsekai/releases/download/v0.1.0/wyrdsekai-relay-0.1.0.tar.gz
  tar xzf wyrdsekai-relay-0.1.0.tar.gz
  cd wyrdsekai-relay-0.1.0
  sudo sh relay.sh relay.example.com

Verify before you run

You are about to run 1.8 GB of someone else's software as root. Check it arrived intact:

curl -fLO https://github.com/Wyrdsekai/wyrdsekai/releases/download/v0.1.0/SHA256SUMS
sha256sum -c SHA256SUMS --ignore-missing      # Linux
shasum -a 256 -c SHA256SUMS --ignore-missing  # macOS

A checksum served from the same host as the file proves the download was not corrupted in transit. It does not prove the host was not compromised — for that, build from source. Both paths are documented, and neither is treated as the paranoid option.

The first start

Expect roughly ten minutes. wyrd start runs first-time setup, fetches about 5.3 GB of model weights, and brings up the inference containers before the server answers. Then:

sudo wyrd invite bootstrap --name yourname   # mint a steward invite
ssh yourname@localhost -p 7022               # the code is your password

Full instructions, including building from source and multi-machine households, are in INSTALLATION.md.

Mobile

The Android and iOS clients are built from the clients directory. They are not in an app store for this release: you build and sideload them, and they reach your household through your own relay.

Models

The companion models are open weights (Apache 2.0), hosted on Hugging Face: huggingface.co/wyrdsekai. Installers fetch them automatically during wyrd setup — you only need these links for mirrors, re-quantization, or fine-tuning.