Frigade open-sources Hob, a local desktop AI assistant positioned against ChatGPT and Claude Cowork

A local alternative to hosted assistants

A new project called Hob has reached the Hacker News front page through a Show HN post, presenting itself as a free, open-source alternative to ChatGPT and Claude Cowork. According to the project's GitHub page, Hob is a desktop application built by Frigade and pitched as a simpler way to run AI on your own machine. It is distributed for Mac, Windows and Linux under the MIT license.

The core idea is where the AI runs. Rather than routing prompts through a cloud service, Hob runs open-weight models — the repository names Gemma and gpt-oss as examples — directly on the user's computer. The hardware bar is modest by current standards: the project says any machine with 8 GB of memory or more qualifies, with the Mac build targeting Apple silicon.

Automatic model selection

One of the more practical design decisions described in the repository is that Hob does not ask users to become experts in model selection. The app instead works out which leading open-weight model will fit the computer it has just been installed on, then handles download and setup on its own. That removes a layer of configuration work that has historically made local inference feel like a hobbyist project rather than a consumer product, where the user normally has to research and pick a model appropriate for their hardware.

Document work without the cloud

Feature-wise, the repository is brief. Hob is shown completing PDF forms and processing documents, with further capabilities promised beyond those. Because everything runs on the device, the app functions offline, and the privacy claim follows directly from the architecture: working on a document does not require sending that document anywhere.

The MIT license means the code can be inspected, modified and redistributed with essentially no restrictions, and the project links to a contributing page, suggesting the developers are open to outside contributions.

Why it matters

The gap Hob is aiming at is straightforward. Hosted assistants are capable, but they depend on an account and an internet connection, frequently on a subscription for serious use, and every prompt and attachment passes through the provider's servers. For anyone handling sensitive material — contracts, medical files, internal company documents — that data flow is the central objection, not a minor detail.

Running locally inverts the equation. A document processed on your own laptop stays there, the tool keeps working with no network at all, and there is no recurring fee to keep using it. The cost is capability: a model has to fit within the machine's memory, so users will necessarily be running something smaller than the largest hosted frontier models. Hob's own design, which selects a model based on what the hardware can accommodate, is an implicit acknowledgment of that ceiling. For privacy-sensitive work, a somewhat less capable model that never touches the network may be the better trade, and if open-weight releases continue to improve, the sacrifice involved keeps shrinking.