Hob: an open-source, on-device alternative to Claude Cowork and ChatGPT

Frigade's Hob reached Hacker News as a Show HN with a blunt pitch: an open source, on-device alternative to Claude Cowork and ChatGPT. The FrigadeHQ/hob repository on GitHub puts it more plainly still, calling itself "an easier way to use AI on your own computer", with downloads for Mac, Windows and Linux.

The word doing the work is "easier". Open-weight models such as Gemma and gpt-oss can run on personal hardware, but someone who wants that has to decide which model their machine can carry and then get it running. Hob's answer is to make the decision for you. In the README's words, it "picks the best frontier open-weight model that fits your computer and sets everything up for you."

The payoff it promises is concrete: fill out PDFs and work through documents with an assistant that is "free, private, and works offline." If that holds, the file you are working on never has to leave your laptop. The catch sits in the same sentence. The model you get is whatever fits your computer, so your hardware sets the ceiling on what the assistant can do.

What Hob says it does

The README is short on promises and specific about the ones it makes. Hob is pitched at document work: filling out PDFs and working through documents, "and much more". It is described as private and able to work offline, which follows from where the model runs. Nothing in that workflow requires a hosted model.

The models are the open-weight kind, with Gemma and gpt-oss named as examples. The phrasing is "like Gemma and gpt-oss", so those are illustrations rather than a closed list. The app ships for all three desktop platforms, and the repository is tagged desktop-app, local-llm and offline-ai, which matches the claim.

The model picker is the product

Plenty of tools will run a local model if you hand them one. Hob's claim is that you never have to choose. It "picks the best frontier open-weight model that fits your computer" and handles setup. The repository structure reflects that emphasis: alongside the app sits a separate package called picker, and a top-level folder called catalog.

That design cuts both ways. A newcomer gets a working assistant without learning about model sizes or runtimes. But "best that fits" is a judgement made on the user's behalf, and "frontier" is relative to the open-weight models available. On a lighter machine, the best model that fits may be a modest one.

What the repository shows

The code is a TypeScript monorepo managed with pnpm: there is a pnpm-workspace.yaml, a pnpm-lock.yaml, a shared tsconfig.base.json and an .nvmrc pinning the Node version. Top-level folders split the work into app, catalog, packages/picker, docs and site, with CI living in .github/workflows. At the time of writing it had 208 commits.

The repository's topic tags fill in some of the stack. It is tagged electron, which fits a single desktop app shipped to three operating systems, and llama-cpp, a common engine for running models locally. It is also tagged mcp, ai-agent and ai-assistant. Treat the tags as signposts rather than a feature list; the README does not spell out how each is used, so check the docs folder before relying on a specific integration.

Open source on MIT terms

Hob is free and released under the MIT license, one of the most permissive available: you can read the code, fork it and build on it, provided the copyright notice travels with it. For a tool whose selling point is privacy, being able to inspect what it does with your files is part of the value, not a side benefit.

The project has a CONTRIBUTING.md and had drawn 38 forks against 427 stars, with 4 people watching. Those are early numbers for a Show HN launch, and they describe interest rather than maturity. It is also a company project: the README credits Frigade, so its direction is set by one team rather than a foundation.

Limits and what to watch

The biggest variable is not the software but the machine under it. Hob's promise is the best model that fits your computer, which means its usefulness scales with RAM and processor rather than with a subscription tier. Anyone weighing it against Claude Cowork or ChatGPT should test their real documents on the hardware they actually own, not take the comparison in the headline as settled.

Watch, too, how the picker explains its choices, how the catalog of models grows as new open-weight releases arrive, and whether the agent and mcp tags turn into documented features. With 208 commits behind it, Hob is young; the README sets out a clear idea, and the next few months of commits will show how far it goes.

This article draws on Open Source Alternatives to Reflector? and answers to it (1, 2), on the Stack Exchange network, licensed under CC BY-SA 4.0.