Hob: A Free Local AI Assistant for Files, Code and Offline Work

Frigade introduced Hob on October 26, a free desktop AI assistant that runs open-weight models on your computer and works with local files. It combines model selection, document handling and code execution in one app, making local AI useful for more than a chat window.

The practical question is whether a smaller model can finish a task reliably. Running a model locally is already possible in many tools. Hob is trying to make the steps around that model feel like a finished desktop product.

What Hob does

The official repository offers downloads for Apple Silicon Macs, Windows and Linux, with at least 8 GB of memory. Hob is MIT-licensed and free, with no account or subscription required. The October 26 version 0.2.1 release provides Apple silicon ARM64, Windows x64 and Linux x86_64 installers; an Intel Mac installer is not listed.

You can attach files or group folders into projects, then ask the assistant to work through them. The useful example is a spreadsheet: a model can reason about the request, run Python to process the rows and return a chart. The answer comes from a tool working on the data rather than the model trying to reproduce calculations in prose.

That distinction matters for local AI. A small model does not need to memorize spreadsheet functions or calculate every total itself. It needs to choose the right operation, write working code and interpret the result accurately.

Which models and hardware does Hob support?

According to Hob's documentation, the app examines your hardware and downloads a suitable model. Its listed choices include Gemma 4 E4B Compact for 8 GB Macs, E4B for 16 GB machines, Gemma 4 26B from 24 GB, and gpt-oss 120B for workstations. You can change the model yourself.

Inference uses llama.cpp, with Metal on Mac and CUDA or Vulkan elsewhere. This separates the desktop interface from the software doing the model computation. It also means the experience depends on your machine, not just the model name.

The minimum memory requirement should be read as an entry point, not a promise of equivalent speed or quality across supported computers. A file task that feels comfortable on a workstation may be slow or unreliable on a small laptop.

Local does not mean every feature is offline

Hob says chats and files remain on-device and that it has no telemetry. The documented network exceptions include model downloads and checks, web search, Python package installation and network-enabled code when Web access is turned on. Connectors added by the user can also communicate online.

Local inference and an entirely offline workflow are different settings. Someone using Hob to work on confidential material needs to understand both. Turning Web access off narrows the workflow, but installing an online connector introduces a separate destination for whatever that connector receives.

The project describes approval confirmations before changes to folders, task undo and sandboxed code execution. These are useful design choices, not independent evidence that every possible file or plugin operation is safe. Keep a backup of important work and review requested changes rather than treating local execution as a guarantee.

Plugins make the assistant more than a chat app

The plugin documentation describes three extension types: written skills, Python tools and Model Context Protocol connectors. A plugin uses a hob-plugin.json file, and the app also accepts .mcpb bundles.

Skills explain how to perform a task. Python tools provide executable operations. MCP connectors connect the assistant to other software, either locally or online. Keeping those roles separate makes it easier to understand what a plugin adds and what access it might require. The connector documentation explicitly says MCP servers are not sandboxed. Some tools need approval, while tools classified as reads can run without asking. The sandbox described for Python execution should not be assumed to cover every connector.

Hob's own code, web and PDF features use this plugin system. PDF form filling is listed for Mac specifically, which is an important qualification if that is the workflow you want to use on Windows or Linux.

Where Hob still needs to prove itself

In the launch discussion, creator Christian described difficulty getting small models to use tools consistently. Examples included claiming to have made a chart without calling the tool and getting stuck in repeated calls. He says the work included building guards around these failures.

That is a more useful limitation than a generic claim that local models are smaller. An assistant can write a convincing explanation while failing to produce the requested artifact. The test is the saved file, the correct calculation and the completed task, not how confidently the answer is phrased.

The same discussion includes readers pointing to existing local-assistant alternatives. Hob should therefore be evaluated on its setup and file workflows, rather than treated as the first app to make local agents possible. No independent comparison in the launch evidence establishes that it matches frontier hosted assistants.

For developers and technically comfortable users, this is a concrete project to evaluate on a small, repeatable task. Use a copy of a spreadsheet, ask for a transformation with a known answer, and check both the result and the permissions requested.

The interesting direction is whether local assistants can make modest models dependable enough for ordinary work. Better setup helps people start. Reliable tool use is what will make them come back.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product launches and engineering ideas into clear stories that developers understand and share.

Book a 30-minute call.