UpTrajectory Review

A developer named Christian has released Hob, an open-source desktop application for Mac, Windows, and Linux that mimics the file-handling capabilities of the Claude and ChatGPT desktop apps. The critical difference: Hob runs entirely on your own hardware using open-weight models like Gemma 4 and gpt-oss, never transmitting your data to external servers. The app automatically detects your system capabilities on first launch, defaults to CPU-only mode if your graphics card proves unstable, and includes built-in web search via SearXNG that users can disable. It also supports the Model Context Protocol (MCP), recognizes images, and executes sandboxed Python code for tasks like data processing and chart generation.

For small-business operators handling sensitive financial records, client contracts, or proprietary designs, Hob addresses a genuine vulnerability in mainstream AI workflows. When you drag a folder of invoices into ChatGPT or Claude, that data leaves your machine and enters someone else's cloud. Hob keeps everything local—your chats, your files, your analysis. This matters particularly for businesses in regulated industries or those handling client data under NDA, where uploading documents to a third-party AI service might violate agreements or compliance requirements. The ability to process large PDFs and images locally, rather than hitting context-window limits or privacy barriers, could streamline document-heavy workflows without the legal overhead.

What distinguishes Hob from the crowded field of local AI interfaces is its focus on agency rather than conversation. Most open-source alternatives stop at chat; Hob reads directories, modifies files, and generates executable code. Christian's critique of existing tools—that they either paste plain text into tiny context windows or default to hosted models despite marketing themselves as privacy solutions—identifies a real gap. Skepticism is warranted regarding hardware demands, though. Running capable open-weight models locally requires substantial RAM and processing power. The auto-detection feature suggests the developer has wrestled with compatibility issues, but operators with older hardware may find the experience sluggish compared to cloud alternatives.

The second-order effects extend beyond individual productivity. If tools like Hob mature, they could shift the balance of power in the AI market by decoupling capability from subscription fees. Businesses currently paying monthly for multiple AI seats might redirect those funds toward hardware upgrades, creating a different kind of vendor lock-in—this time to GPU manufacturers rather than AI platforms. Additionally, the sandboxed code execution, while powerful, introduces attack vectors. Malicious prompts could theoretically manipulate the local environment, though the sandboxing mitigates this. The MCP support also expands the attack surface, as connecting external data sources requires careful configuration.

Watch how Hob handles the tension between local-only operation and the practical need for cloud-based features. The web search integration via SearXNG is a pragmatic compromise, but operators should verify whether their queries leak sensitive context when the tool searches online. The project is MIT-licensed and free, which eliminates procurement headaches but also means no service-level agreements or support guarantees. Test it first on non-sensitive data to gauge whether your hardware handles the model comfortably. If you operate in legal, medical, or financial services where client confidentiality is paramount, Hob represents a viable path to AI-assisted document processing without compliance violations—provided you accept the performance trade-offs of local inference.

“All your chats and files never leave your system.” — Hacker News (front page)

Takeaway: Test Hob on non-sensitive documents first to verify your hardware handles local inference before trusting it with client data.