Sarja
Early release · 0.1.0

Sarness

A private AI workstation that runs on hardware you control.

Sarness (Sarja Harness) is a native Linux desktop app for working with AI models hosted on your own GPU server. The desktop is the interface; a separately installed engine does the inference. Your prompts and media stay on infrastructure you own.

Downloads & installationRequirementsView on GitHub

How it fits together

Two components, installed separately.

The app on your desktop does not run models. It connects to the Sarness Engine on a GPU server — the same machine, a box on your network, or one reached over Tailscale or an SSH tunnel.

  1. 1 · Server

    Sarness Engine

    Python service with a job queue and GPU workers, running llama.cpp and Diffusers-based pipelines.

  2. 2 · Desktop

    Sarness Desktop

    Tauri app that connects with an API token, stored in your system keyring.

What you can do with it.

  • Streaming AI chat

    Chat with GGUF models through llama.cpp, or any OpenAI-compatible endpoint, with Markdown, highlighted code, stop, token statistics and a thinking toggle.

  • Model selection

    Installed models and a curated catalogue with VRAM figures, plus Hugging Face search and install, resumable downloads and llama.cpp runtime management.

  • Image attachments

    Attach images for vision-capable models, and text files as context, directly in a conversation.

  • Branching, editing, regeneration

    Edit a message or regenerate a reply and move between the branches that result, without losing the original.

  • Image generation and editing

    Prompt-to-image, reference-image (img2img) and mask-painted inpainting, with advanced controls exposed only where the model supports them.

  • Video generation

    Text-to-video, and image-to-video for the models that support it, with time estimates measured on your server.

  • Node-based workflows

    An editor that shows the actual graph simple mode runs, with typed sockets, validation, presets, and import/export of workflow files.

  • Library and projects

    Every generation with full metadata, searchable and filterable, with collections, tags, favourites and a trash with restore.

Capabilities depend on the models you install and the GPU in your server. Sarness 0.1.0 is an early release; no application screenshots are published here yet.

Requirements.

Sarness Desktop

  • Linux on x86-64 (.deb, .rpm or AppImage)
  • Network access to your engine server — Tailscale, LAN or an SSH tunnel
  • OpenSSH client if you use the guided install or SSH tunnels

Sarness Engine (server)

  • Ubuntu 22.04 or newer, x86-64
  • NVIDIA GPU with driver 570 or newer
  • Python 3.11+ with venv, and about 15 GB for the runtime plus model storage

Sarness 0.1.0 is built for Linux only. There are no Windows or macOS builds.

Version 0.1.0 · Linux x86-64 (Ubuntu 22.04 or newer tested)

Downloads.

Sarness 0.1.0 has been built, but public downloads are not published yet. The checksums below identify the exact build that will be released; each file becomes a download link here the moment it is hosted.

Sarness Desktop

On the computer you work at. Choose one package.

  • Desktop app

    Debian / Ubuntu package

    Desktop application for Debian, Ubuntu and derivatives.

    Sarness_0.1.0_amd64.deb · 5.4 MB

    SHA-25680488bc70b070b918120e212f78f95cbbd01e4f6264a7dbdf9ab23d8d708f280
    Not yet published

  • Desktop app

    Fedora / RHEL package

    Desktop application for RPM-based distributions.

    Sarness-0.1.0-1.x86_64.rpm · 5.4 MB

    SHA-256829ebc8efe053a531e87bb86bdfa4cd5753ebb19bf882ef818db957959a18f94
    Not yet published

  • Desktop app

    AppImage

    Portable desktop application; no installation required.

    Sarness_0.1.0_amd64.AppImage · 86.8 MB

    SHA-2561861af1d848ac70b226e0193e71203cb0ec63044f0340655d02b1e35f064582d
    Not yet published

Sarness Engine

On the GPU server. Install this first; it is a separate component.

  • Server engine

    Sarness Engine bundle

    The inference engine, installed separately on your GPU server.

    sarness-engine.tar.gz · 98 KB

    SHA-2569cb59ebdc95ac3aea6a047a5339c26f14bbe65b7225d1cbc020c992fc08d8a0f
    Not yet published

  • Server installer

    Engine install script

    Installs the engine bundle on the server as a systemd user service.

    install-engine.sh · 7.6 KB

    SHA-256fec6b3c7bf98c08052b4b7235e1a32f38302a5cfdf2d398d0d7c15b2c7392e92
    Not yet published

Verify a file with sha256sum <file> and compare it with the value above.

Install the engine.

On the GPU server, with the bundle and installer in one folder:

bash install-engine.sh --source sarness-engine.tar.gz --bind tailscale --with-llama

The installer prints an API token once. It runs in user space as a systemd user service and does not use sudo or change drivers or firewall rules. Alternatively, the desktop app can install the engine over SSH.

Install the desktop.

sudo apt install ./Sarness_0.1.0_amd64.deb
# or: chmod +x Sarness_0.1.0_amd64.AppImage && ./Sarness_0.1.0_amd64.AppImage

Launch Sarness, enter the server’s address and port 8742, paste the token, choose Test connection, then Save & connect. Then open Models to install what suits your GPU.

Want a private AI environment designed for you?

Sarness is how we work with private models ourselves. We also design and operate private AI infrastructure for organisations.

Talk to SarjaBrowse services