Hob: A Free, Open-Source Local AI App for Your Files (What Businesses Should Know)

On 26 October 2026 a developer named Christian posted Hob to Hacker News as a "Show HN". Within about 12 hours it had passed 130 points and 30 comments. The pitch is simple: a desktop AI assistant that works on your files like the Claude and ChatGPT desktop apps, except the model runs on your own computer and nothing about your chats or documents goes to a cloud service.

This post covers what Hob is, what is confirmed so far, its limits, and how a business should think about local AI. Everything below comes from the Hacker News post, its comments and the project's GitHub page. We have not run Hob ourselves, so we do not claim benchmark results.

Hob at a glance

What Hob does

The author tried other local AI apps first and found that none could do what the ChatGPT and Claude apps do with your files. Hob is built to close that gap. The listed features are:

• Automatic model choice. On first install it works out the best open-weight model your hardware can run.
• Sandboxed Python. It writes small tools on the fly for data processing, document conversion and charts.
• Local file access. It can read, create, edit and delete files in a folder you give it.
• Image recognition through a small Qwen3-VL model, and local dictation through a Whisper model.
• Web search powered by SearXNG, which you can switch off.
• MCP support, so you can connect your own tools. (If MCP is new to you, see our explainer on WebMCP and AI agents.)
• Offline use, with no internet needed once models are downloaded.

How it works under the hood

According to the author, Hob is an Electron app that talks to llama.cpp running in its own process. If it cannot find a compatible graphics card it falls back to CPU only. The author notes that cards that looked compatible but crashed on the first query were a recurring problem.

Generated code runs in a sandbox that can only write to the chat's own folder and has no network access unless asked. The sandbox uses sandbox-exec on macOS, Landlock on Linux and a low-integrity process on Windows.

The hardest engineering problem, the author says, was getting small models to use tools reliably. Four-billion-parameter models would announce "here's your chart" without ever calling the chart tool, or loop forever on a tool call. The fix was building guards around the model rather than only tuning prompts.

Limits and early feedback

Hob is days old, and the Hacker News thread already shows where it is rough:

• It is new. The project page shows 26 stars and one contributor in the 0.2.0 release notes. Expect bugs and fast changes.
• Local models are weaker. The author says so directly, and argues that for most tasks most people do not need the newest frontier model. Their own evals suggest 16 GB or more of RAM does well.
• Disk use. One commenter reported about 18 GB of models installed with no built-in way to remove them after deciding the tool was not useful for real work. The author apologised and said cleanup would come in a later update.
• Install location. Another commenter wanted models on a different disk. They found that the HOB_USER_DATA setting covered their case.
• No LAN model support yet. A commenter asked to point Hob at a model hosted elsewhere on the local network. The author asked a clarifying question and has not announced it.
• Crowded space. Commenters named alternatives including Open WebUI, LM Studio, Jan, Osaurus and the llama.cpp app. Hob's angle is a simpler setup that picks the model for you.

When local AI makes sense for a business

Most businesses we speak with use hosted AI, and that is usually the right call for quality. Local AI becomes worth testing in a few cases:

1. Sensitive files. Contracts, HR records, customer data or financial documents that your policy says cannot go to a third-party service.
2. Offline or low-connectivity work. Field teams, travel, or sites with unreliable internet.
3. Cost control. Repetitive document work where per-seat or per-token fees add up, and a good-enough local model is acceptable.
4. Learning and prototyping. A free, MIT-licensed tool is a cheap way to find out what on-device models can and cannot do.

It is a poor fit when you need the best possible reasoning, have older or low-memory machines, or must give staff a supported, audited tool. In those cases a hosted assistant, or a custom setup, is safer.

A sensible way to try it

• Use a spare or non-critical machine first, with at least 16 GB of memory.
• Copy a small set of non-sensitive documents into a test folder and give Hob only that folder.
• Check the output against the source yourself. Small models can sound confident and still be wrong.
• Leave web search off if you want a fully offline test.
• Check free disk space before the first model download.
• Write down what worked, what did not, and whether a hosted tool did the same job better.

What it means for your AI plans

Hob is one more sign that "AI on your own files" is no longer limited to cloud apps. Open-weight models are good enough that a free desktop app can attempt document work privately. Whether that is good enough for your business depends on your data, your hardware and your tolerance for weaker answers.

If you are deciding between hosted tools, on-device AI and a custom assistant built on your own data, our guide to AI agents for business and our list of practical AI uses for small business are good places to start. If you want help choosing or building something, see our AI services or contact our team.

Sources

• Hacker News: Show HN: Hob - Open-source, on device alternative to Claude Cowork and ChatGPT, posted 26 October 2026.
• GitHub: FrigadeHQ/hob and its releases (v0.2.0 and v0.2.1, 26 October 2026).

Frequently asked questions

What is Hob?

Hob is a free, open-source desktop app for Mac, Windows and Linux, made by Frigade and released under the MIT license. It lets you chat with an AI that can read and change files in a folder, look at images, run sandboxed code and draw charts, using open-weight models that run on your own computer.

Is Hob really private and offline?

According to its author, your chats and files never leave your computer. The only things that go online are web search, which can be turned off in settings, and any MCP connectors you add yourself. It is described as working fully offline once the model is installed.

Which models does Hob use?

Currently open-weight models from the Gemma 4 and gpt-oss families. On first install Hob checks your hardware and picks the best model that fits. Image recognition uses a small Qwen3-VL model and dictation uses a local Whisper model.

What computer do I need to run Hob?

The project page lists Apple silicon Macs, Windows and Linux with 8 GB of memory or more. The author says machines with 16 GB or more of RAM do well, and that a GPU or extra RAM gets results closer to hosted frontier models.

Is Hob as good as ChatGPT or Claude?

No, and the author does not claim it is. Local models are weaker than the newest frontier cloud models. The argument is that most everyday tasks do not need the newest frontier model. Test it on your own documents before relying on it.

Is it safe to let an AI change files on my computer?

Hob runs generated Python in a sandbox that can only write to the chat's own folder and has no network unless asked, using sandbox-exec on macOS, Landlock on Linux and a low-integrity process on Windows. It is still a very new project, so try it on copies of your files first.

Can a business use Hob?

The MIT license allows commercial use. Whether it fits depends on your data rules, your hardware and how much quality you can trade for privacy. For regulated data, ask your security or legal lead before adopting any new tool.

Last reviewed 27 October 2026. This article is general information, not a guarantee of results. Hob is changing quickly and details may differ by the time you read this.