Hob's Show HN: The Local AI App Gap Big Labs Left Open

Hob is a free, MIT-licensed desktop AI app that runs open-weight models on your own machine, and its Show HN post reached about 144 points and 33 comments because it solved a problem the frontier labs have not: keeping files where they are (Hacker News). The thread is a useful case study in launching an open-source AI tool. The packaging works; the models underneath are still small.

TL;DR

• Hob is a free MIT desktop app for Mac, Windows and Linux that runs Gemma, gpt-oss and similar open-weight models locally (Hacker News).
• It reached about 144 points and 33 comments on 27 October 2026, with privacy, offline use and a working installer as the pitch (Hacker News).
• Commenters pushed back on model quality and on yet another local chat app, not on the idea of local AI (Hacker News).
• The maker said the hardest part was getting smaller models to use tools reliably, which is where the product work actually sits (Hacker News).

Disclosure: ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 27 Oct 2026; ZeroRaised has no commercial relationship with the companies named.

Local-first AI is a product design in which the model and the user's data run on the user's own device, so the core job still works without a network connection and files do not have to be uploaded to a vendor.

What is Hob and why did it reach the front page?#

Hob is a desktop app from maker pancomplex, who goes by Christian, and it runs on Mac, Windows and Linux (Hacker News). It is free and MIT-licensed, with the code on GitHub. It runs open-weight models such as Gemma and gpt-oss locally, so files never leave the machine and the app works offline (Hacker News).

The feature list goes beyond a chat box: it picks the best model for your hardware automatically, runs sandboxed Python, offers web search through SearXNG, supports MCP, does image recognition with Qwen3-VL, and includes local Whisper dictation (Hacker News). It is built on Electron and llama.cpp, with sandboxing through sandbox-exec, Landlock or a low-integrity process (Hacker News).

That is why the post travelled: the thing installs, finds a model your laptop can run, and keeps working when the network drops.

Why is local-first a gap the big labs left open?#

The frontier labs sell API calls and a hosted chat product. They have little reason to optimise for a 16 GB laptop running someone else's weights, and none to make the offline case easy. A local-first app is a different job: hardware detection, model downloads, disk management, sandboxing and an interface that survives a slower model.

That leaves a real gap. Privacy and data residency are not abstract for lawyers, clinicians and accountants. A promise that files never leave the device is easier to verify than a vendor's retention policy.

The gap is integration and trust, not a new model. You cannot train a frontier model, but you can own the last mile.

What did the Hacker News thread criticise?#

• User wilg complained that models of about 18 GB are left on disk without cleanup, and wrote that "The models are really dumb, mainly." (Hacker News).
• User jrm4 wrote that "there are a zillion of these" (Hacker News).
• User dfltr pushed the other way: "having something Just Work(tm) makes getting started so much easier." (Hacker News).

The maker did not dodge the tool-use point. Christian said "The hardest part was definitely to get the smaller models to use tools reliably", and noted that machines with 16 GB of RAM or more do well (Hacker News).

The category is crowded, the raw models are weaker than an API, and the thing users reward is the boring work of making it just work.

What can founders learn from the launch?#

Lead with a constraint you remove, not a category you join. "Local AI assistant" puts you next to a zillion others. "Your client files never leave your laptop" gives a buyer a reason to care.

Ship something people can run in one step. The most positive comment in the thread was about how easy it was to get started (Hacker News). A free MIT licence lowers the cost of trying you, and a working installer lowers it further.

Answer the hard comment with specifics. The maker conceded the tool-use weakness and named the hardware that works (Hacker News). Our guide to writing a Show HN post covers the format, and marketing an open-source project covers the follow-up.

Where is the opportunity beyond another local chat app?#

The interesting space is vertical, offline and confidential: document review for a small law firm, notes for a clinic with poor connectivity, field work where the laptop is the only computer, or an agent working inside files a company will never upload.

Each needs a workflow, file formats and vocabulary a general chat app lacks. The model can be swapped; the integration and the trust cannot. That is the argument in our guide to building a defensible AI product and in open weight vs closed AI models.

Hob also supports MCP (Hacker News). A local agent that calls local tools is more useful than a local chatbot, and MCP is becoming the standard way to wire tools up. A tool server is also an attack surface, which is why agent-proofing a product matters even offline.

What this means for bootstrapped founders#

• Pick a buyer with a confidentiality problem. Privacy sells to people who already have a rule to follow.
• Compete on the install, the sandbox and the cleanup. The thread's complaints were about disk use and model quality, and both are product decisions you control (Hacker News).
• Be honest about model limits. Say which jobs work, which need a bigger machine and which should stay on an API.
• Price the workflow, not the tokens. Local inference has no per-token bill, so charge for the job, the team or the support. Our guide to pricing an AI product covers the shape.
• Plan for support. Hardware variance is your cost centre, and it is why a thin wrapper is easy to clone but hard to keep working.

Where ZeroRaised fits#

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have shipped a local-first tool or an AI workflow a real person can try, you can submit your product and get a page that states what it does, who made it and how to run it. There is no promise of traffic or results.

For the next analysis, join the ZeroRaised newsletter. The pricing page explains the current listing options.

Sources#

• Hacker News: Show HN, Hob
• GitHub: FrigadeHQ/hob