Back home

Hob

x+hn+web 0.78
Visit source

Intro

Hob is a free, MIT-licensed desktop AI app from Frigade for Mac (Apple silicon), Windows, and Linux. Models run on your machine; chats and files stay local by default. The pitch is Claude/ChatGPT-desktop-style file work without sending your folder to a hosted model — currently advertising Gemma 4 and gpt-oss via llama.cpp.

Site: hob.chat. Code: FrigadeHQ/hob. Version 0.2.x shipped around 2026-10-26, then hit Hacker News Show HN (~149 points / 34 comments). This radar window (2026-10-27 16:32–20:32 Asia/Shanghai) still saw multi-language posts. An earlier same-day 08:32 pass dropped Hob for low traction; Show HN + official site + GitHub now corroborate the named launch.

Explain

What it claims (site / Show HN / README)

  1. Local-first: no account/subscription; claimed no telemetry. Built-in web search (SearXNG) can be disabled; MCP connectors you add may go online.
  2. Hardware-aware setup: on first launch it tries to pick an open-weight model that fits RAM/bandwidth (author examples range from small Gemma 4 on 8 GB Macs up to larger MoE / gpt-oss on high-RAM machines).
  3. File tools: read/edit authorized folders, fill PDFs, vision via a small VL model, sandboxed Python (sandbox-exec / Landlock / low-integrity), charts.
  4. Stack: Electron shell + llama.cpp in a separate process; optional local Whisper for dictation.

Positioning

The author says prior local apps either pasted tiny text into a short context window or defaulted to hosted models/accounts. HN commenters note the category is crowded; Hob's bet is privacy + less setup friction.

Why this radar term

Short name, but branded domain + Frigade shipping + verifiable Show HN thread in/near this window — recorded as product name Hob.

Usage

Try it

  1. Downloads: https://hob.chat/
  2. Source/releases: https://github.com/FrigadeHQ/hob
  3. Show HN: https://news.ycombinator.com/item?id=50246238
  4. After install, let it fetch a model; turn off Web access in Privacy settings for offline-ish use; grant folder access only as needed

Good fit

  • People handling sensitive local docs who refuse cloud chat uploads
  • Users who want a packaged local assistant instead of hand-wiring Ollama/LM Studio

Caveats

  • Small local models are weaker at reliable tool use than frontier cloud models; quality scales with RAM/GPU
  • Windows SmartScreen may warn on an unsigned/new publisher build
  • Models can consume tens of GB; uninstall carefully
  • Points/stars move; figures are as observed at scan time