How to Use Hob: The Beginner's Guide (First Launch, Models, Files, and Offline)
CodexFuse Team · 2026-11-03T05:06:53.587Z · 13 min read
Learning how to use Hob starts with a ten-minute setup: you download one free installer from hob.chat, launch it, and Hob — Frigade's free, open-source, MIT-licensed AI desktop app for Mac, Windows, and Linux — picks an open-weight model that fits your machine and downloads it, all with no account and no subscription. Everything below walks you through the full beginner's arc: the install moment, the auto-pick model lesson, your first file job, the offline rule, and the connectors. One housekeeping note before we begin: Hob is not hob.dev, an unrelated agentic-workspace product from a different company (founderland.ai flagged the collision on October 27, and the collision still exists today). And if you dig through the release history and see the name "Garth" in the v0.2.0 notes — same product; Frigade renamed it to Hob. This guide covers the current release, Hob 0.2.6 (October 30, 2026).
Hob's pitch is short enough to quote from its own homepage: "A free, open-source alternative to Claude and ChatGPT that runs open-weight models like Gemma and gpt-oss locally on Mac, Windows and Linux. Private, and works offline." (github.com/FrigadeHQ/hob, fetched 2026-11-03). The rest of this guide is about making that pitch true on your machine.
What Is Hob, Exactly?
Hob is a desktop AI assistant that runs open-weight language models entirely on your own computer instead of in someone else's cloud. It is made by Frigade, a Y Combinator-backed company, its source code is on GitHub at github.com/FrigadeHQ/hob under the MIT license, and it runs on Electron bridged with llama.cpp for local inference (Metal on Mac, CUDA or Vulkan elsewhere, per the vendor's docs).
What makes it beginner-friendly rather than just open-source is the auto-pick: "Hob picks the best frontier open-weight model that fits your computer and sets everything up for you" (hob.chat, fetched 2026-11-03). You never touch a model file, a quantization table, or a config flag unless you want to. If you already run local models the hands-on way, the Ollama beginner guide covers that lane; Hob is the one-installer, desktop-app lane.
Step 1: Download Hob and Check Your Machine First
The direct answer: get the installer from hob.chat or the Releases page at github.com/FrigadeHQ/hob/releases, then make sure your machine meets the bar — Apple Silicon Mac, Windows PC, or Linux; at least 8 GB of RAM; 10–25 GB of free disk for a typical setup (hob.chat, fetched 2026-11-03).
Hob's hardware guidance is refreshingly specific, and these are vendor claims, so take them as the vendor's promises rather than independently tested fact: 16 GB is "good for everyday" use, 32 GB and up is "the best," and the largest model needs roughly 75 GB of disk (hob.chat, fetched 2026-11-03). On Windows, the app is unsigned, so SmartScreen will show a warning — click "More info" and then "Run anyway" to proceed (hob.chat, fetched 2026-11-03).
On macOS the installer goes into your Hob app-data folder and models live there too; on Windows and Linux you can pick a different drive on the first screen if your system drive is tight (github.com/FrigadeHQ/hob docs, fetched 2026-11-03). The GitHub repository is public, MIT-licensed, and on November 3, 2026 showed 456 stars, 38 forks, and 4 watchers — up sharply from the 236 stars founderland.ai reported on October 27.
Step 2: Learn the Auto-Pick Model Lesson (This Is Hob's Core Trick)
The direct answer: Hob chooses a Gemma 4 variant for most computers based on your RAM — E4B Compact on 8 GB Macs, E4B on 16 GB, and the 26B model from 24 GB — and reserves gpt-oss 120B for workstations, per the vendor's current model docs; you can switch models any time (github.com/FrigadeHQ/hob docs, fetched 2026-11-03).
Vendor claims, stated once for honesty: the models are "ranked on agent work," and the app splits memory between GPU and CPU when that is faster. The founder's own public example, quoted by founderland.ai (October 27, 2026), was an 8 GB Mac getting a small Gemma 4 model, a 128 GB Mac Studio getting gpt-oss 120B, and an RTX 4090 machine getting Gemma 4 26B — the exact variants in the shipping docs have since been refined (E4B Compact vs the earlier E2B wording), so trust the docs' tiers, not a week-old quote.
Expect the model download to take a while: models run 10–25 GB each, and the largest needs about 75 GB (hob.chat, fetched 2026-11-03). Download once; everything after that works offline. Here is the vendor's tier map, as published in the current docs:
| Your machine | Model Hob picks | Why it fits |
|---|---|---|
| 8 GB RAM Mac | Gemma 4 E4B Compact | Small enough to run well with memory left over |
| 16 GB RAM | Gemma 4 E4B | The everyday tier the vendor calls the sweet spot |
| 24 GB+ RAM | Gemma 4 26B | More capable model for roomier machines |
| Workstation (e.g., 128 GB Mac Studio) | gpt-oss 120B | The big model, for hardware that can feed it |
Two things make this genuinely low-friction for beginners: there is no account to create and no subscription — "Hob and its models are free" (hob.chat, fetched 2026-11-03) — and the release notes show the project is moving fast, with seven releases between October 26 and October 30, 2026, the latest being Hob 0.2.6.
Step 3: Say Hello — Your First Conversation, No Account Needed
The direct answer: open the app, finish the short playful onboarding (the release notes' old "Garth" name is just pre-rename history — same product), and type your first question. There is no sign-up screen and no "log in to continue."
The no-account design is worth appreciating, because every mainstream AI chat app starts by asking who you are. Hob skips it: "No account, no subscription" is one of the homepage's headline claims (hob.chat, fetched 2026-11-03). Optional external services you choose to connect later may have their own fees, but Hob itself and its models cost nothing.
Hob speaks 32 UI languages (hob.chat, fetched 2026-11-03), and it ships with on-device speech-to-text dictation — the docs describe it as "Speech to text, on your computer" (github.com/FrigadeHQ/hob docs, fetched 2026-11-03). You can also paste photos or screenshots into the chat; a small Qwen3-VL vision model is downloaded after the main model to handle image input (github.com/FrigadeHQ/hob docs, fetched 2026-11-03).
Step 4: Put Hob to Work on Your Files (The Arc That Justifies the Install)
The direct answer: give Hob a real file job — a thick PDF, a messy spreadsheet, or a code file — and it reads or edits it locally, asking before it changes anything and offering an Undo per task.
This is the moment a local AI stops being a demo and becomes a tool. Hob reads and edits documents, including PDFs — even scanned ones — and PDF form filling works on Mac (github.com/FrigadeHQ/hob docs, fetched 2026-11-03; chatgate.ai, October 28, 2026). Spreadsheets are a first-class use case: drop one in and ask for an analysis, and Hob can produce interactive charts you can download as PNGs, powered by sandboxed Python with pandas, numpy, openpyxl, and Pillow available (github.com/FrigadeHQ/hob docs, fetched 2026-11-03).
Folder access is explicit and permission-first: Hob asks "Let Hob use your 'Sales' folder?" style questions and asks again before changing files (hob.chat, fetched 2026-11-03). For code, paste the file or point at the project and ask for a review — the sandboxed Python environment also includes PyPI access for real data work. If you would rather run models on a machine you rent instead of your desk machine, the self-host an AI chatbot on a VPS guide covers that lane — and what VPS hosting is explains the underlying idea. Hob's lane is your own computer; the VPS lane is a server you control.
Step 5: Master the Offline Rule (One Toggle Is All That Stands Between You and the Cloud)
The direct answer: Hob works fully offline after the model downloads — web search is the only built-in online feature, and you control it under Settings → Privacy → Web access.
"Works offline" is a headline claim on hob.chat (fetched 2026-11-03), and the mechanics are simple: with web access off, the only internet Hob needs is for downloading models, app updates, and occasionally software packages for tasks you request. Search itself is interesting — Hob searches "through a metasearch engine that runs on your computer" (github.com/FrigadeHQ/hob docs, fetched 2026-11-03), which founderland.ai (October 27, 2026) attributes to SearXNG based on the founder's own Hacker News comment. Even with search on, queries run through that local engine rather than a vendor-hosted pipeline.
One privacy note, stated once because it matters: as of November 3, 2026, Frigade publishes no Hob-specific privacy policy — hob.chat/privacy returns a 404 page. The vendor's privacy claims ("Hob has no telemetry"; chats and files stay on your system) live in the hob.chat FAQ and the open-source code, and the FAQ is candid that web tools and connectors "send the information needed for their actions to external services," which have their own privacy policies (hob.chat, fetched 2026-11-03). Judge the privacy claim on that evidence.
Step 6: Connect Hob to Your Other Apps (MCP and Connectors)
The direct answer: Hob connects to your tools two ways — built-in connectors (Notion, Linear, Figma, GitHub, plus Mail & Calendar on Mac) and any MCP server via plugins, including .mcpb bundles.
A plugin is just a folder with a hob-plugin.json file, and it can add skills, Python tools, and "connectors to any MCP server, including .mcpb bundles" (github.com/FrigadeHQ/hob docs, fetched 2026-11-03). There is a nice human touch in the docs: after you ask Hob to do a repeated task, you can say "save this as a skill" and it remembers the pattern for next time.
The connectors arc is where local AI quietly becomes more useful than a cloud chatbot, because Hob can reach the apps you already use. The guardrails are explicit: folder access is granted per folder, Hob asks before changing files, and every task gets an Undo option (hob.chat, fetched 2026-11-03). Treat connectors as the one area where "runs locally" has an asterisk — connecting Notion or GitHub necessarily sends the information needed for the action to that service.
Hob vs the Alternatives: When Hob Wins, and When It Doesn't
The direct answer: pick Hob when you want a polished one-installer desktop app with automatic model setup; pick Ollama when you want a bare-bones model runner to build on; consider Jan or Chatbox when you want a different UI on the same local-model idea; keep ChatGPT or Claude when you need frontier-cloud capability.
The local-AI desktop space has several names beginners bump into — Jan, Chatbox, Lobe Chat — and they are comparison members here, not rivals with their own beginner guides on this site. The meaningful comparison is the lane: the Ollama beginner guide is the hands-on CLI path for people who enjoy terminals; Hob is the product path for everyone else. Here is the decision table:
| You want | Best pick | Why |
|---|---|---|
| A desktop app that just works, local-first | Hob | One installer, auto model pick, no account, open source |
| Full control via the terminal | Ollama | The CLI runner everything else builds on; more manual |
| A different UI over local models | Jan or Chatbox | Same local-model idea, different interface choices |
| Maximum model capability, cloud okay | ChatGPT / Claude | Frontier cloud models still beat anything local |
| AI on a rented server, not your desk | VPS self-host | Your own machine but remotely — the VPS AI-chatbot guide |
Be honest about the tradeoff: a Gemma 4 E4B on your laptop is not a frontier cloud model. Hob wins on privacy, cost (zero), and offline work — not on raw capability.
7 Beginner Mistakes to Avoid With Hob
The direct answer: the seven mistakes below are the difference between a smooth first day and an afternoon of confusion. Skim them before you download anything.
| Mistake | What happens | Do this instead |
|---|---|---|
| Confusing it with hob.dev | You sign up for an unrelated product | Only download from hob.chat or github.com/FrigadeHQ/hob |
| Downloading the wrong installer | A build for the wrong OS | Each release ships 12 assets (github.com/FrigadeHQ/hob releases, fetched 2026-11-03) — pick your platform |
| Skipping the disk-space check | The model download stalls mid-way | Keep 10–25 GB free, ~75 GB for the largest model (hob.chat, fetched 2026-11-03) |
| Expecting cloud-model quality | Disappointment on hard questions | Use Hob for private, everyday work; use the cloud for the hardest tasks |
| Leaving web access on by default | You aren't actually offline | Check Settings → Privacy → Web access before sensitive work |
| Granting a whole drive at once | Hob sees more than the task needs | Grant folders one at a time when asked |
| Assuming a privacy policy exists | Trusting a document that doesn't | Know that no Hob-specific privacy policy is published yet — the claims rest on the vendor FAQ and the open-source code |
Sources
Every fact in this guide was verified on November 3, 2026 from a live page: the Hob GitHub repository and docs (github.com/FrigadeHQ/hob, releases page), the vendor homepage (hob.chat), founderland.ai's launch coverage (October 27, 2026), chatgate.ai's product explainer (October 28, 2026), the grafy.ai and steadyvane.com launch pieces, and a hands-on spreadsheet walkthrough on Medium (October 29, 2026). Vendor claims are labeled as vendor claims throughout.
Frequently Asked Questions
Is Hob really free, or is there a paid tier?
Free. "No account, no subscription. Hob and its models are free" (hob.chat, fetched 2026-11-03). Optional external services you choose to connect may have their own fees, but Hob itself costs nothing.
Which model will Hob install on my computer?
Hob picks automatically based on your hardware, per the vendor's current docs: Gemma 4 E4B Compact on 8 GB Macs, E4B on 16 GB, 26B from 24 GB, and gpt-oss 120B for workstations — and you can switch models any time (github.com/FrigadeHQ/hob docs, fetched 2026-11-03).
Does Hob work without an internet connection?
Yes, after the one-time model download. Web search is the only built-in online feature, controlled under Settings → Privacy → Web access; otherwise Hob only needs the internet for downloads and updates (hob.chat and vendor docs, fetched 2026-11-03).
What was Garth, and why do the release notes mention it?
Garth was Hob's earlier name. The v0.2.0 release notes (October 26, 2026) mention "Polish Garth onboarding" and replacing "legacy fire-themed startup copy" — the product was renamed to Hob, so treat any "Garth" reference as pre-rename history.
Is my data safe with Hob?
Hob's privacy story is this: AI runs on your computer and your chats stay with you, per the vendor, and "Hob has no telemetry" per the hob.chat FAQ — but Frigade publishes no Hob-specific privacy policy as of November 3, 2026, so verify that yourself before doing sensitive work, and know that connectors send the information needed for their actions to external services (hob.chat, fetched 2026-11-03).
