A developer identified as Christian has shared Hob on Hacker News, an open-source desktop application released under the MIT license for Mac, Windows and Linux. The tool replicates the core workflows of the Claude and ChatGPT desktop apps — handling files, reading images, running code — but executes the language model entirely on the user's own machine rather than sending data to a remote service.
Hob relies on open-weight models, currently Gemma 4 and gpt-oss, and automatically detects the most suitable configuration for the user's hardware at install time, falling back to CPU-only mode when no compatible graphics card is found. The app can write and run sandboxed Python code for tasks such as data processing or chart generation, perform web searches through SearXNG (which can be disabled), connect to MCP servers, recognize images, and read or modify files on the local filesystem. It is designed to work fully offline.
The author says he tested most existing local AI tools before starting the project and found that none could match what the proprietary desktop apps do with larger files or images — most simply pasted plain text into a limited context window, which broke down on anything beyond a few kilobytes. He also notes that the more capable open-source alternatives tend to default to hosted models, which undermines the privacy benefit that running everything locally is supposed to provide, often in service of an account-based or open-core business model.
Under the hood, Hob is built with Electron bridged to llama.cpp running in its own process. It remains an early-stage, solo-developer project, still far from the polish of the commercial products it aims to rival, but it reflects a growing interest in AI assistants capable of handling complex documents without relying on cloud infrastructure. Its long-term viability will depend on community adoption and on how well open-weight models keep pace with proprietary ones on these kinds of tasks.