What Is Free Local AI "Hob"? 5 Features You Can Use Without an Internet Connection
機械翻訳 / Machine-translated
A
AI Friends
1日前・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever thought, "I'd love to have an AI read my company documents, but I'm worried about sending them to an external server"? A free app called "Hob" has arrived to address exactly that concern. In this article, we'll walk you through Hob's features, how to use it, and how it differs from other tools.
Hob is a desktop app published as open source on GitHub. It is developed by Frigade and released under the MIT license (a permissive license that allows free use and modification).
It aims to serve as a free alternative to Claude and ChatGPT, with the AI running entirely on your own PC.
This news was introduced by GIGAZINE on November 6, 2026. Supported operating systems are Windows, Mac, and Linux. From GitHub Releases, you can get a DMG for Mac (arm64), an EXE for Windows (x64), and an AppImage for Linux (x86_64).
Incidentally, the strength of local AI (AI that runs on your own PC) is that your text never has to be sent over the internet. Hob aims to reduce the hassle of getting started with it.
One place many people get stuck with local AI is choosing which model to use. Models (the data that acts as the AI's brain) are smarter the larger they are, but they also consume more memory.
When Hob launches for the first time, it checks your PC's memory and GPU (the component that accelerates image processing and computation). It then displays a list of model candidates that your machine can run.
Choose the model you want, download it, and the chat screen opens. You can change the model later.
ScriptByAI's introduction article includes a rough guide for Mac users:
The app supports 32 languages, including Japanese. However, there are no fine-tuning controls for settings such as Temperature (how varied the responses are), context length (how much text the AI can remember at once), or GPU usage.
In other words, it is designed more for people who want to get started right away than for those who enjoy tweaking settings.
Hob doesn't stop at chat. It reads files, writes Python scripts, and runs them.
The GIGAZINE article includes an example of handing Hob a CSV and asking it to "total sales by month, create a bar chart, and explain the changes in Japanese." Hob wrote and executed the script, then displayed the chart and aggregated results in the chat.
Other features include:
Approval is required when modifying a folder, and undoing actions is also supported. You can rest easy knowing there is little risk of AI rewriting your files without permission. Note that filling in PDF forms is a Mac-only feature.
Consider an accounting staff member at a small company who compiles a sales CSV at the end of each month. Previously, they would aggregate figures in a spreadsheet and manually create charts. With Hob, they simply hand it the CSV and ask it to "aggregate by month and make a chart." The sales data never leaves their machine.
Next, consider a legal staff member who needs to read a contract PDF from a business partner. Many companies prohibit entering contracts into external AI services under internal rules. With Hob running locally, they can organize the key points entirely on their own device.
Finally, think of someone who needs to organize documents while traveling on a plane or bullet train. With a model already downloaded, they can continue chatting and processing saved files even without an internet connection.
Hob is not the first tool for running local AI. Two prominent examples are Ollama and LM Studio.
The Ollama official website states that it has over 9 million monthly installs and more than 180,000 GitHub stars. It is a tool that can run models both locally and in the cloud. The LM Studio official website highlights an agent feature that assists with document creation, coding, and PC operation.
Hob is positioned for users who want to leave model selection and setup to the app itself. ScriptByAI's page lists LM Studio and similar tools as related resources, but does not include a detailed comparison. Any judgment of performance differences will need to wait for further evaluation.
Since it supports 32 languages including Japanese, it can be used with Japanese-language documents. For Japanese companies that handle personal information or confidential materials, having an option that keeps data in-house is a significant development.
That said, there are some points to be aware of.
According to GIGAZINE, Hacker News saw positive comments along the lines of "you can get started right away." At the same time, there were also criticisms that "the model isn't very smart," and that an 18 GB model cannot be easily uninstalled.
Local models may fall short of the latest cloud models in terms of capability. Free disk space is also required. Start by trying it with non-critical data to see how it performs on your PC before using it in earnest.
In the privacy settings, turning off "Web Access" will prevent web searches and internet access from code. Also keep in mind that connecting to external services via MCP does involve network communication.
A. Yes. According to ScriptByAI, local use requires no account, subscription, or per-message charges. The license is MIT.
A. After downloading the model file, you can continue chatting and processing saved files offline. Web search and integration with external services do require a connection.
A. The README does not provide specific figures. Based on ScriptByAI's guidelines, 8 GB of RAM on a Mac is recommended for running Gemma 4 E4B Compact. Speed depends on your RAM and GPU memory.
A. Yes. For analyzing screenshots and similar content, it uses a small model called Qwen3-VL, which is downloaded separately.
Start by testing it with non-critical data to see how far it runs on your PC.
This article is a cross-post from AI Friends.
