Haruspex User guide

Getting started

Haruspex is a desktop AI app for research and coding that runs on your own computer. It works on Linux, Windows and macOS. There is no account to create and no telemetry.

What Haruspex does

First run: the setup wizard

The first time you open Haruspex, a wizard asks how you want to run the model:

  1. Download a model (recommended). Haruspex checks your hardware (GPU, VRAM and RAM), shows the model it recommends, and lets you pick another from the list. You can also choose Use existing GGUF file to use a model you already have.

    With a graphics card of 8 to 24 GB and at least 32 GB of RAM, the wizard also offers Larger model using system RAM: Qwen 3.6 35B-A3B, which is smarter but slower. Use this instead picks it and turns on Settings → Inference → Let models use system RAM.

  2. Connect to an existing server (advanced). Enter the address of an OpenAI-compatible server. Nothing is downloaded.

After a download, the wizard sends a short test question to the model. If the test fails, you can Retry or Skip; the model may still work. Then press Start chatting.

Model files are several gigabytes. On a slow connection, Choose a different model lets you switch to a smaller one while the download runs.

To run the wizard again later, use Settings → Inference → Run Setup Wizard. Your existing models and settings are kept unless you change them there.

The main tabs

Tab What it is for
Chat Ask questions. The assistant can search the web, read and write files in a working folder, run Python, look at images, and use your email, calendar or other connected services if you turn them on.
Jobs Save a task and run it later, by hand or on a schedule: research, audit, guided planning, autonomous coding and asset generation. A badge shows when jobs are running or queued.
Shell A real terminal with an assistant beside it. Read-only by default: it suggests commands for you to run. You can open several shell tabs.
Code A coding agent that works in one project folder: it edits files and runs commands, showing diffs and command output. Sessions are saved. On Windows, inside WSL.

Along the top of the window you also find the server status (click it to open the logs), a light/dark toggle, the log viewer, the help list of keyboard shortcuts (? or F1), and Settings.

Check what the model tells you

AI models make mistakes. They can state wrong facts with confidence, misread files, and suggest commands that are wrong or harmful. The small models Haruspex uses on modest hardware make these mistakes more often than large cloud models.

The Shell assistant runs nothing by default. If you give it Full access, it runs commands in your real terminal: commands it flags as risky ask you first, but commands it thinks are safe run on their own. The Code tab works the same way in its project folder. Only use them on machines and projects you are willing to let the model change. Read every command before you run it, and keep backups.

Ask Haruspex about itself

In Chat or the Shell assistant, ask things like "how do I add a calendar?" or "is memory on?". The assistant reads this guide, and how this copy of Haruspex is set up (the model, which features are on), before it answers, and says so when the guide doesn't cover something. Remote guests get the guide but not your setup.

Where to go next