Open-source Hob brings Claude Cowork-style workflows to your laptop, no cloud required

Developer Christian has presented Hob on Hacker News: a free desktop app under the MIT license for Mac, Windows and Linux. It does with local files what the Claude and ChatGPT apps do: it reads and edits documents in a folder, analyzes images, runs code in a sandbox and draws charts. The difference is that the model runs on your machine. The post collected 162 points and 39 comments, and the source code is in the FrigadeHQ/hob repository.

What existing local clients were missing

Before starting, the author tried every local AI app he could find. None of them could do the one thing that mattered: work with files the way cloud assistants do. The most advanced approach he saw was pasting a text file into a small context window. That broke down on files of just a few kilobytes. And nobody handled images or large PDFs in any sensible way.

More sophisticated solutions, he observed, defaulted to hosted models, open or closed. That defeats the privacy point of running locally. Some projects also pushed toward accounts and an open-core business model. Hob is built on the opposite principle: chats and files never leave the system, and the internet is needed only for web search (which can be turned off in settings) and for MCP connectors that the user adds personally. Without a network connection, the app works fully offline.

How

Hob is built: Electron, llama.cpp and three layers of isolation

The app is written in Electron and talks to llama.cpp running in a separate process. On first launch, Hob checks whether a compatible GPU is present and picks the best open model for the hardware. If there is no GPU, it falls back to the CPU. The author admits this took some work: some GPUs looked compatible on paper but crashed on the very first request.

Code written by the model runs in a sandbox. The sandbox can write only to the current chat's folder and has no network access unless the user allows it. Each platform uses its own mechanism:

• macOS: sandbox-exec
• Linux: Landlock
• Windows: a low-integrity process

Through these scripts, the model builds one-off tools for data processing, document conversion and chart generation.

Small models and tools

According to the author, the hardest part was getting small models to call tools reliably. Models with four billion parameters happily wrote "here's your chart" without ever calling the charting tool, or got stuck in an endless loop of calls. Tweaking prompts barely helped on small models. Most of the time went into programmatic guardrails around the calls.

The full list of what the project includes:

• General-purpose models: Gemma 4 and gpt-oss, open weights
• Image recognition: a small Qwen3-VL model
• Voice input: a local Whisper model
• Web search: based on SearXNG, can be turned off in settings
• Support for MCP connectors that the user adds
• Create, read, edit and delete operations on files in the local file system
• Python code in an isolated sandbox: data processing, document conversion, charts
• Automatic selection of the best model for the hardware on first launch
• Fully offline mode
• MIT license; Mac, Windows and Linux

Who it suits and what to expect from quality

The author does not dispute that local models are weaker than cloud flagships. But for most everyday user tasks, he argues, the latest frontier model is rarely needed. By his own measurements, systems with 16 GB of RAM or more perform well, and with a discrete GPU or more memory, the results come close to what a hosted model delivers out of the box. This is the developer's own assessment; there are no independent benchmarks yet, and he explicitly asks the community to test Hob on other configurations.

The practical takeaway for readers is this. If you work with contracts, financial spreadsheets or internal documents that cannot be sent to the cloud, Hob is worth installing and running on files typical for your work. Check three things: how stable tool calling is on your model, how the app behaves on a machine without a discrete GPU, and how well it handles large PDFs. Teams building their own agent products should look at the code: the setup of an isolated llama.cpp process, OS-level sandboxes and tool-call guardrails solves problems that anyone trying agents on small models will sooner or later run into.