Open-source Hob lets users run Claude‑like AI locally

An open‑source desktop application called Hob lets anyone run a Claude‑style AI assistant directly on their computer, bypassing cloud services. The tool, released by Frigade, supports locally‑hosted models such as Gemma and gpt‑oss and works on macOS, Windows and Linux without an internet connection.

What is Hob and why it matters?

Hob is positioned as a free, open‑source alternative to commercial chat agents like Claude and ChatGPT. By installing the software on a personal machine, users obtain a private AI that never sends prompts or responses to external servers. This design addresses growing concerns about data leakage, subscription costs, and reliance on internet connectivity for everyday AI tasks such as drafting text, answering questions, or summarising documents.

Which models does Hob run and how are they selected?

The application ships with a model‑selection routine that scans the host hardware and chooses the most suitable “frontier” open‑weight model. Currently supported models include Gemma, an efficient transformer released by Google, and gpt‑oss, a community‑maintained replica of OpenAI’s GPT architecture. Hob automatically downloads the appropriate model binaries and configures them to run with the llama‑cpp inference engine, which is lightweight enough to operate on machines with as little as 8 GB of RAM. Because the models are open weight, users can inspect the code and weights, ensuring transparency.

How does Hob compare to proprietary services like Claude and ChatGPT?

Feature	Hob (open‑source)	Claude / ChatGPT (cloud)
Privacy	All processing stays on‑device; no network traffic needed.	Prompts are transmitted to provider servers; data may be logged.
Cost	Free to download and use; only hardware cost applies.	Subscription fees or pay‑per‑token pricing.
Scalability	Limited by local CPU/GPU; inference speed depends on hardware.	Massive data‑center resources give faster responses and larger context windows.
Updates	Community‑driven releases; new models must be added manually.	Providers roll out continuous model improvements behind the scenes.
Offline capability	Works fully offline after initial download.	Requires internet access for every request.

In practice, Hob is well suited for developers, researchers, or privacy‑focused creators who need occasional assistance without exposing sensitive material. Power users who demand the very latest model size or multilingual breadth may still prefer cloud APIs.

Getting started: installing and using Hob

1. Download the installer from the project’s website — hob.chat — or from the GitHub releases page. Packages are available for Apple‑silicon Macs, Intel Macs, Windows and Linux.
2. Verify system requirements: at least 8 GB of RAM and a recent CPU; a GPU is optional but can speed up inference.
3. Run the installer; the app is built with Electron, so it opens like any other desktop program.
4. Select a model: Hob will suggest the best‑fit model based on your hardware, or you can manually pick Gemma, gpt‑oss, or any compatible llama‑cpp model you have downloaded.
5. Start a chat session: the UI resembles a typical messaging app. You can paste text, upload PDFs, or drag‑and‑drop files for the assistant to analyse.
6. Use offline features: because the model runs locally, you can work in environments without internet—air‑gapped labs, remote field sites, or simply on a personal laptop while traveling.

The repository’s README provides detailed commands for advanced users who prefer a terminal‑only setup, including how to point Hob at a custom model directory or adjust inference parameters such as temperature and token limit. The source code and release notes are hosted on GitHub: GitHub repository.

Where Hob fits into a Grafy workflow

Grafy already offers a Chat interface that can query dozens of cloud models side by side. For users who have installed Hob, the same conversational experience can be replicated locally by launching the Hob desktop client and then connecting it to Grafy Chat via the Operator browsing agent or by simply copying the output into a Grafy project.

If you are a developer looking to prototype code with an AI assistant, Grafy Code provides an integrated coding studio that can call external tools. You could configure a custom operator step that forwards a code‑generation prompt to Hob’s local endpoint, captures the response, and feeds it back into the editor—giving you the privacy of an offline model while still benefiting from Grafy’s version‑control and prompt‑graph visualisation. Both scenarios let you keep the bulk of your workflow inside Grafy while leveraging Hob for the parts that require strict data confidentiality or zero‑cost operation.

Frequently asked questions

How much hardware does Hob need to run comfortably?

Hob runs on any modern laptop or desktop with at least 8 GB of RAM. On CPUs, response times are a few seconds per request; a modest GPU (e.g., NVIDIA‑RTX 2060 or Apple‑silicon GPU) reduces latency dramatically, making the experience comparable to cloud services for short prompts.

Can I use Hob with proprietary models like Claude?

No. Hob is limited to open‑weight models that can be run with the llama‑cpp engine. Proprietary models are hosted behind provider APIs and therefore cannot be executed offline. Users who need Claude‑specific features must still rely on the official service.

Is my data truly private when using Hob?

Yes. All inference happens on the local machine; the application does not open outbound network connections after the initial model download. As long as the host system is secure, the prompts and generated text never leave the device.

---

Source: Hacker News — front page — Show HN: Hob - Open-source, on device alternative to Claude Cowork and ChatGPT