Hob is a free desktop app that runs an open-weight model locally: Reveneau

Why it mattersA team that cannot send files to a cloud model gets a local app. It reads folders, runs sandboxed code and uses MCP.

A privacy team that is told it may not paste a client file into a cloud chat has few options. The desktop apps from Claude and OpenAI need an account and send the file to a hosted model.

A solo developer at FrigadeHQ, Christian, posted Hob to Hacker News two hours ago. The Show HN post has 28 points and 12 comments at the time of writing. Hob is a free, MIT-licensed desktop app for Mac, Windows and Linux. The author says it works on local files the way the Claude Cowork and ChatGPT desktop apps do. The model runs on the user's own computer.

What the app does

The author lists what Hob does on first install. It picks the best open-weight model for the hardware. The current default models are Gemma 4 and gpt-oss. Hob writes sandboxed Python code for small tools like data processing, document conversion and chart drawing. It reads and changes files in a folder. It takes images into a Qwen3-VL model for recognition. It records voice through a local Whisper model. It calls MCP servers the user adds. Web search is on by default, powered by SearXNG, and can be turned off in settings.

How it is built

The author says Hob is Electron bridged with llama.cpp running in its own process. The code sandbox can only write to the chat's folder and has no network unless the user asks for it. The sandbox uses sandbox-exec on macOS, Landlock on Linux and a low-integrity process on Windows. On first load Hob tries to find a compatible graphics card. If none works it falls back to CPU only.

The author says the hardest part was getting smaller models to call tools reliably. Four-billion-parameter models sometimes said "here's your chart" without calling the chart tool, or got stuck in a call loop. The author built guards around the tool call rather than only tweaking prompts.

The caveat the author names

The author writes that local models are weaker than frontier cloud models. For most tasks, the post says, a user does not need the newest frontier model. From Hob's own tests, systems with 16 GB of RAM or more do well. A machine with a GPU or more RAM gets close to a hosted frontier model's output. These are Hob's own figures, not an independent test.

What it is up against

The thread on Hacker News lists other local desktop AI apps the author tried before building Hob. Jan, OpenWork, Chatbox and Lobe are named. One commenter writes that there are many of these tools. Another names Osaurus as a privacy-focused option. A third names PasClaw as a tool built for local embedding. Hob itself is MIT licensed on GitHub.

Source

Primary source: FrigadeHQ/hob on GitHub. Discussion: Show HN on Hacker News.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.