Frigade, the Y Combinator W23 startup, released Hob on Oct. 26, a free, MIT-licensed desktop assistant that runs open-weight models locally on macOS, Windows and Linux. It is pitched as an on-device alternative to Claude Cowork and the ChatGPT desktop app.
The company announced the project in a "Show HN" post on Hacker News, where the thread had 162 points and 39 comments as of Oct. 27. The post's author, Christian, whose Hacker News profile lists him as CTO of Frigade and whom the company's About page names as a cofounder, described it as "a free, MIT-licensed desktop app … except the model runs on your own computer." According to him, Hob works on local files the way the Claude and ChatGPT desktop apps do.
What Hob can do
Hob currently runs Gemma 4 and gpt-oss. Image recognition is handled by a small Qwen3-VL model, and dictation by a local Whisper model. Because it supports the Model Context Protocol (MCP), an open standard for linking AI tools to outside data and services, users can add connectors, including remote servers.
The app can also write sandboxed Python for ad-hoc jobs like data processing, document conversion and charting. Christian wrote that the sandbox relies on sandbox-exec on macOS, Landlock on Linux and a low-integrity process on Windows, with "no network unless asked."
How Hob picks a model, and what it says about privacy
On first install, Hob chooses an open-weight model suited to the user's hardware. "It will automatically pick the best [model] for your HW on install," Christian wrote in a reply on Oct. 27.
His examples ran the range. An 8 GB Mac (M1/M2) gets Gemma 4 E2B, a 128 GB Mac Studio gets gpt-oss 120B, and a machine with an RTX 4090 graphics card runs Gemma 4 26B entirely on the GPU. Replies in the thread indicate the selection aims for a minimum of roughly 15 tokens per second, and that Windows and Linux support setups that split work between GPU and system memory.
On privacy, Christian wrote: "All your chats and files never leave your system." The only online features are an optional built-in web search, powered by SearXNG, and any MCP connectors a user chooses to add. Search can be turned off. Still, Frigade has not published a formal privacy document specific to Hob, and no explicit telemetry or account policy turned up in the research for this article, so the privacy claim rests for now on the developer's word and the open-source code.
