Run a private, Claude-style AI coworker on your own machine: hands-on with Hob

Hob is a free, MIT-licensed desktop app that runs open-weight models entirely on your computer — no account, no subscription, and no telemetry. This guide installs Hob 0.2.1, lets it pick the model your hardware can actually run, and puts it to work on a real spreadsheet before taking it fully offline.

• Who: Hob, a free, open-source desktop AI app by Frigade, MIT licensed.
• What: A Claude-style coworker that runs open-weight models locally via llama.cpp — files, code, PDFs, images, charts, and plugins, with approval before any folder change.
• When: Hob 0.2.1, released 26 October 2026; Hob itself was created 14 October 2026.
• Models: Gemma 4 for most computers (E4B Compact on 8 GB Macs, E4B on 16 GB, 26B from 24 GB) and gpt-oss 120B for workstations; a Qwen3-VL 4B vision model follows the main download.
• Installers: About 193 MB (Mac), 416 MB (Windows), and 532 MB (Linux) for version 0.2.1.
• Cost and account: Free, no account, no subscription, no limits, and no telemetry — but Web access, PyPI installs, and connectors you add do go online.

Cloud chatbots are convenient, and they are also a leak by design: every document you paste, every spreadsheet you upload, and every question you ask travels to someone else's server. Local chat apps fixed the leak but usually gave you a bare chat box — no tools, no file work, and a model picker that assumed you already knew which 8-billion-parameter file your laptop could run.

Hob tries to close that gap. It is a free, open-source desktop app from Frigade that behaves like a Claude or ChatGPT coworker — it works on your files, runs code, reads PDFs and images, and asks before it changes anything — except its models run entirely on your own computer through llama.cpp. There is no account, no subscription, and no telemetry. You do not pick a model on day one: Hob checks your hardware and downloads the best open-weight model it can run well.

This guide walks through exactly that setup: install Hob 0.2.1, let it choose your model, give it a real spreadsheet job, take it offline, and save the workflow so next month's version is one sentence away. Everything about Hob below comes from its own docs and its 0.2.1 release, checked on 27 October 2026.

What you will need

• A Mac, Windows, or Linux computer.
• Room for the download: the Hob 0.2.1 installers are about 193 MB (Mac), 416 MB (Windows), and 532 MB (Linux), plus the model download that follows on first launch.
• No account, API key, or subscription — Hob asks for none of these.
• A file you already own to test with: a spreadsheet, a PDF, or a folder of documents.
• Optional: Python 3, only if you want to generate the sample spreadsheet in Step 3 with the script below. Otherwise, make the same table in any spreadsheet app.

Step 1: Download the right installer

Go to the Hob releases page and pick the installer for your machine: Hob-mac-arm64.dmg for Mac, Hob-windows-x64.exe for Windows, or Hob-linux-x86_64.AppImage for Linux. The README links point at the latest release, so if a newer version than 0.2.1 is showing, take that one and note the version — the steps below do not depend on the exact build. Open the installer and follow its prompts; there is no account screen to get past, because there is no account.

If you would rather start from the project page than the release list, hob.chat links to the same downloads.

Step 2: Let Hob pick your model

On first launch, Hob checks your hardware and downloads the best model it can run well. Do not skip this wait, and do not judge speed until it finishes: the main model download comes first, and a small vision model follows it. That vision model is Qwen3-VL 4B, and it is what lets Hob look at photos, screenshots, receipts, charts, and error messages you paste in.

What you get depends on your machine. Hob's docs say it picks from a curated list: Google's Gemma 4 for most computers — E4B Compact on 8 GB Macs, E4B on 16 GB, and 26B from 24 GB — and OpenAI's gpt-oss 120B for workstations. Models are ranked on agent work, meaning multi-step tasks and tool use, and Hob splits work between graphics and system memory when that is faster. You can switch models any time, so the first pick is not a commitment.

Once Hob is ready, ask it something small and check the answer before you trust it with files: for example, ask it to explain what it can do with a spreadsheet. If it answers sensibly, you are set up. If it stalls, the most likely cause is still the model download — check that first before reinstalling anything.

Step 3: Give it a real job: the spreadsheet test

A chat box proves nothing. A file job does. Make a small sales table — either in your spreadsheet app or with this Python script, which writes sales.csv in the folder where you run it:

import csv

rows = [
    ["region", "product", "units", "revenue"],
    ["North", "Widget", 40, 1200],
    ["North", "Gadget", 25, 1750],
    ["South", "Widget", 30, 900],
    ["South", "Gadget", 20, 1400],
    ["East", "Widget", 55, 1650],
    ["East", "Gadget", 35, 2450],
    ["West", "Widget", 20, 600],
    ["West", "Gadget", 15, 1050],
]

with open("sales.csv", "w", newline="") as f:
    csv.writer(f).writerows(rows)

Now put that file in a Hob project — Hob works on attached files, or on folders grouped into projects with their own instructions — and give it one concrete job:

Chart revenue by region from sales.csv, and summarise what stands out, with the numbers.

Two things to check in the answer. First, the chart: Hob makes interactive charts in the chat, downloadable as PNG, so download it and keep it. Second, the arithmetic: in this sample file, East totals $4,100, North $2,950, South $2,300, and West $1,650, for $11,000 overall. If Hob's summary matches those totals, it read the file properly; if it does not, say so and ask it to show its working per region. That check matters more than whether the chart looks nice.

The same pattern works for the other file types Hob lists: attach a PDF and ask it to summarise it (it reads PDFs, including scans), or paste a screenshot of an error message and ask what it means. For code, Hob runs Python with pandas, numpy, openpyxl, python-docx, pypdf, and Pillow, plus any package from PyPI it needs — sandboxed to the chat's folder. And it asks for approval before anything in your folders changes, with Undo for every task, so approve a change only when you have read what it plans to do.

Step 4: Take it offline

This is the step that separates a local model from a local experience. Hob's model runs on your computer with llama.cpp (Metal on Mac, CUDA or Vulkan elsewhere), and your chats and files stay on your machine, with no telemetry. But Hob still goes online for some things: it downloads models and checks for new ones, and while Web access is on, it searches the web, lets its code fetch data such as stock prices, and installs Python packages from PyPI. Connectors you add also go online.

To prove the offline part to yourself, go to Settings and turn Web access off, then repeat a file-only job from Step 3 — the spreadsheet chart, or a PDF summary. That job should still work, because the model and your files never left the machine. A web search, on the other hand, should not: that is the expected behaviour with Web access off, not a bug. Turn Web access back on only when you actually want a web-connected job.

Step 5: Save the workflow as a skill

After Hob does something you will want again — the monthly sales chart, say — tell it: save this as a skill. Hob keeps the steps and the code that worked, and shows you both before saving, so read them before you confirm. Next time, you can run the same workflow without re-explaining it.

Under the hood, a plugin is a folder with a hob-plugin.json file. It can add skills in the open Agent Skills format, Python tools, and connectors to any MCP server, including .mcpb bundles. Hob's own web search, code, and PDF features are plugins too. You do not need to build a plugin to use Hob, but knowing the format explains why a saved skill is portable: it is a folder of steps and code, not a hidden setting.

Step 6: Switch models and languages

Finally, try the two freedoms that cost nothing here. First, switch models any time and rerun the spreadsheet job — a bigger model on a bigger machine should handle a messier file, and a smaller one should answer faster on a modest laptop. Second, write to Hob in your own language: it supports 32 languages, matching your computer's language (change it in Settings), and answers in whatever language you write in. If you work in French, Spanish, or any of the other supported languages, test the same prompt in that language before you decide Hob is an English-only tool.

What you built

You now have a desktop AI coworker that needs no account and sends no chats or files to a cloud provider: Hob installed, a model matched to your hardware, a checked spreadsheet job with a downloadable chart, an offline mode you have actually tested, and a saved skill for the next run. That is the whole point of Hob — not a demo of a local model, but a working setup you can hand real files to.

Honest limitations

• PDF form filling is Mac-only. Reading PDFs, including scans, works generally; filling in forms is listed for Mac.
• Your hardware picks your model. An 8 GB machine gets the compact model, not the 26B or 120B one. Expect a smaller local model to be slower and less capable on the hardest tasks than a cloud frontier model — Hob does not publish a benchmark in its docs to claim otherwise, and neither do we.
• The first run is a download, not an instant start. Judge Hob after the main model and the vision model have both downloaded, not while they are still arriving.
• Offline is a setting, not the whole story. With Web access on, Hob searches the web, fetches data for its code, and installs PyPI packages; connectors you add go online too. If privacy is the reason you installed Hob, leave Web access off except when you need it.
• This is a young project. Hob was created on 14 October 2026, and 0.2.1 — the release used here — landed on 26 October 2026. Expect rough edges and frequent updates; installed copies download new versions and install them when they quit.

Sources

• Hob docs: github.com/FrigadeHQ/hob/blob/main/docs/README.md — no-setup model picking, model tiers, privacy and Web access, file and plugin features.
• Hob releases: github.com/FrigadeHQ/hob/releases — Hob 0.2.1, 26 October 2026, and the Mac, Windows, and Linux installers.
• Hob project page: hob.chat — downloads and project overview.
• All Hob version, model-tier, installer-size, and feature details in this guide were checked against Hob's docs and release page on 27 October 2026. The sample spreadsheet is the author's own test file; its totals are arithmetic on that file, not a Hob output.