What is Hob?
Hob is a desktop app for Apple Silicon Macs, Windows and Linux that does file work in the style of the Claude and ChatGPT desktop apps with a local model. On first launch it measures the machine and downloads the strongest open-weight model that stays above about 15 tokens a second: Gemma 4 E2B on an 8 GB Mac, E4B at 16 to 18 GB, Gemma 4 26B A4B from 24 to 64 GB, and gpt-oss 120B on a 128 GB Mac Studio. It runs llama.cpp in its own process inside an Electron shell, with a small Qwen3-VL model for images and Whisper for dictation.
The agent reads and edits files in folders you share, asks before changing one and can undo changes, writes and runs Python in a sandbox to process data and draw charts, reads PDFs including scans and fills in forms (on Mac), and works with spreadsheets and Word files. Web search through SearXNG is the only built-in online feature and can be switched off; Mail and Calendar connectors are built in on Mac, and MCP connectors can be added.
It is made by Frigade, a Y Combinator-backed company, under the MIT licence. The repository was created on 14 October 2026, launched on Hacker News on 26 October, and reached v0.2.6 on 30 October.
Is Hob worth it?
Hob's contribution is a decision most local-AI apps leave to the user: which model, at which size, for this machine. It picks the strongest one that keeps above about 15 tokens a second with room for a reasonable context, from Gemma 4 E2B on an 8 GB Mac to gpt-oss 120B on a 128 GB Mac Studio, downloads it and gets out of the way. Jan and LM Studio give you a catalogue; Hob gives you an answer.
The second difference is that it acts on files rather than chatting about them. The author's Show HN post says he built it after finding no local app that could read and change files in a folder, look at images, run code in a sandbox and draw a chart. The sandbox can write only to the chat's folder and has no network unless asked, using sandbox-exec on macOS, Landlock on Linux and a low-integrity process on Windows; file edits need approval and can be undone.
Privacy is the strong suit. The site says Hob has no telemetry, models and chats stay on the machine, and with web access turned off in settings it runs fully offline apart from model and app updates. SearXNG search and any MCP connectors you add send what their actions need to those services.
The limits are the models and the age. Gemma 4 E4B on a 16 GB laptop is far from a frontier model, and one Hacker News commenter's verdict was that the models are "really dumb, mainly"; the author agrees they trail cloud models. The project is four weeks old, with about 530 stars and 3 contributors, the Windows build is unsigned, and early users reported 18 GB models left on disk with no way to clean them up, which the author said the next update would fix.
Against PawWork, which sends your files to whichever cloud model you pay for, Hob trades answer quality for keeping everything on your own disk.
Where it earns its place
A GP practice manager on a 32 GB MacBook
Fill in scanned PDF forms and summarise policy documents without uploading patient-adjacent files
Share the folder, turn off web access, and let Hob run Gemma 4 26B A4B locally. Free beyond the laptop; answers on long, subtle documents need checking.
A sales analyst on a Windows PC with an RTX 4090
Chart revenue by region from a spreadsheet
Hob runs Gemma 4 26B fully on the GPU, writes Python in the sandbox to read the file and draw the chart, and saves the result in the chat's folder.
An IT admin at a 25-person charity
Offer staff a free assistant that keeps data on their own laptops
Install the MIT-licensed app on each machine with at least 8 GB of memory and 10 to 25 GB of free disk; there is no central management, and 8 GB laptops get the smallest Gemma model.
Before you commit
- Setup
- Download the Mac arm64 .dmg, Windows x64 .exe or Linux AppImage from GitHub; first launch picks and downloads a model, needing 10 to 25 GB of disk, or about 75 GB for the largest.
- Integrations
- Built-in SearXNG web search, Mail and Calendar on Mac, user-added MCP connectors such as Notion or Linear; llama.cpp, Qwen3-VL for images and Whisper for dictation run locally.
- Data and privacy
- No telemetry, according to the site; chats, files and models stay on the machine. Web search and connectors send data to those services, and web access can be switched off in Settings. MIT licence.
- Cost at scale
- Free at any scale; the cost is hardware, with at least 8 GB of memory and 16 GB or more recommended, and 24 GB or more on a Mac for Gemma 4 26B A4B.
- Limits
- Local models trail frontier cloud models; PDF form filling is Mac only; no Intel Mac build; Windows build unsigned; model clean-up on disk was reported missing in early releases.
What’s good
- Chooses and installs the right model size for your hardware
- No telemetry; runs fully offline with web access off
- Sandboxed code with no network, approval before file edits, and undo
- Reads scanned PDFs and fills forms on Mac
- MIT licence, free with models included
What’s not
- Local model quality on 8 to 16 GB machines trails cloud models
- Four weeks old with 3 contributors
- Models take 10 to 25 GB, about 75 GB for the largest
- Unsigned Windows build; Intel Macs not supported
- Little control over which model it runs
Best for
Non-technical people who want a private, offline assistant that works on their documents and do not want to choose or configure a model.
Hob alternatives
Hob: common questions
- Is Hob worth it?
- Local AI with the setup taken away: it chooses the model for you and lets it work on your files, which is exactly right for a non-technical user and a limit for anyone who wants to choose. We rate it 3.7 out of 5.
- How much does Hob cost?
- Free and open source (MIT), models included. No account, subscription or paid tier; connected services may charge their own fees.
- What are the downsides of Hob?
- Local model quality on 8 to 16 GB machines trails cloud models; four weeks old with 3 contributors; models take 10 to 25 GB, about 75 GB for the largest; unsigned Windows build; Intel Macs not supported; little control over which model it runs.
- Who is Hob best for?
- Non-technical people who want a private, offline assistant that works on their documents and do not want to choose or configure a model.
- How much does Hob cost for a team or in production?
- Free at any scale; the cost is hardware, with at least 8 GB of memory and 16 GB or more recommended, and 24 GB or more on a Mac for Gemma 4 26B A4B.
- Is Hob safe for company data?
- No telemetry, according to the site; chats, files and models stay on the machine. Web search and connectors send data to those services, and web access can be switched off in Settings. MIT licence.
- What are the best Hob alternatives?
- The closest alternatives we have reviewed are Jan, AnythingLLM and PawWork.
Answers checked Nov 11, 2026.