Hob chooses a local AI model, then puts it to work on your files

Hob, an open-source desktop assistant from Frigade, shipped version 0.2.0 on 26 October and reached 0.2.3 on 28 October. It offers Windows and Linux downloads, plus an Apple-silicon Mac installer, and works with files through locally running AI models.[1][2]

Creator Christian says Hob checks for compatible graphics hardware and falls back to CPU execution when needed, automatically choosing an open-weight model such as Gemma 4 or gpt-oss.[4]

The app uses Electron, with llama.cpp running separately. Generated Python handles jobs such as document conversion. Christian says its sandbox restricts writes to the chat’s folder and blocks network access unless requested.[4]

The current help documentation lists projects, spreadsheets, images and scanned PDFs among supported inputs; filling PDF forms is Mac-only. It says changes to user folders require confirmation and tasks can be undone.[3]

What still goes online

Local inference concerns where the model runs. Hob downloads models and checks for new ones. With Web access enabled, it can search pages, fetch online data and install Python packages. After setup, disabling Web access is the documented route for offline work; added connectors can still contact external services.[3]

The creator’s account also identifies the difficult part of the project: making smaller models actually use their tools. Some claimed to have made a chart without calling the chart tool; others repeated calls indefinitely. He describes addressing those behaviours in the surrounding software, after prompt changes proved insufficient.[4]

That makes completed work the meaningful measure of this assistant. A spreadsheet task should end with the requested file or chart, with the right inputs carried through the calculation. A confident chat reply is only one part of the job.[4]

Report an error or suggest a correction