Hob

• Summarizing a folder of sales spreadsheets and notes into a Word report with charts, without uploading the files anywhere
• Filling in a PDF form on a Mac
• Cleaning, merging or analyzing CSV and Excel files with Python that runs in a local sandbox
• Reading receipts, screenshots or scanned documents and turning them into a table
• Working through documents on a plane or a locked-down network with Web access turned off
• Saving a repeated workflow, such as a monthly spending report from a bank CSV, as a skill you can rerun with a slash command

Hob's pitch is narrow and clear: the file-centric coworker experience of Claude or ChatGPT's desktop apps, with the model running on your own machine and nothing to pay for. What sets it apart from local chat apps is the tooling: a sandboxed Python runtime, PDF and image handling, approvals and per-task Undo, and Agent Skills and MCP support, plus automatic model selection so non-technical users never pick a quantization. The ceiling is the hardware. Its own FAQ ties near-frontier quality to 32 GB of RAM, and there is no fallback to a cloud model for harder tasks. It is also weeks old. For privacy-sensitive document work on a well-equipped Apple silicon Mac or PC, it is worth trying; for heavy or high-stakes work, a cloud coworker is still the stronger tool.

About Hob

Hob comes from Frigade, a Y Combinator (Winter 2023) startup, and its repository was created in October 2026. It is an Electron app built around llama.cpp (Metal on Mac, CUDA or Vulkan elsewhere). On first launch it checks your hardware and downloads the open-weight model it ranks best for agent work on that machine: Gemma 4 E4B Compact on 8 GB Macs, E4B on 16 GB, Gemma 4 26B from 24 GB, and OpenAI's gpt-oss 120B on workstations, with a small Qwen3-VL model added for images. You can switch models at any time, but every model runs locally; no option to plug in a cloud model or API key is documented. The workflow is the one desktop AI coworkers made familiar. You attach files or grant a folder (folders can be grouped into projects with their own instructions), describe the job, and Hob plans the steps: reading spreadsheets, Word files, PDFs (including scans) and photos, running Python with pandas, openpyxl, python-docx and pypdf in a sandbox that can only write to the chat's folder, drawing interactive charts, filling in PDF forms on Mac, and writing results you can open and share. It asks before changing anything in a granted folder and records each task's file changes so you can undo them. It extends through plugins: skills in the open Agent Skills format, Python tools, and MCP connectors, including remote servers with OAuth and Claude Desktop's .mcpb bundles. After a task you want to repeat, you can ask Hob to save it as a skill with the code that worked. On Mac a built-in plugin works with the Calendar and Mail apps: it reads and adds events, searches and reads email, and writes drafts that you send yourself. Privacy is the point: chats and files stay on your machine, there is no telemetry and no account, and web search runs through a SearXNG metasearch instance on your own computer. That instance still sends your queries to public search engines, so Hob offers a single Web access switch that keeps it fully offline apart from model downloads and updates. The tradeoffs follow from running locally: quality depends on your RAM (Hob says it gets close to Claude or ChatGPT with 32 GB or more), Intel Macs are not supported, models take 10 to 75 GB of disk, the Windows build is not yet code-signed, and the app is at version 0.2.x.