A 12 MB menu bar app that turns your Mac into a local LLM server

What it does

LlamaBarn is a native macOS menu bar application built on llama.cpp that downloads, configures, and serves local LLMs from a built-in catalog. It exposes a local OpenAI-compatible API at localhost:2276, so chat apps, editors, CLI tools, or scripts can connect without modification. Models load on demand and unload when idle, and all inference stays on your machine.

The interesting bit

The app is only 12 MB because it leans on llama.cpp for the heavy lifting, yet it handles the tedious parts—quantization fallbacks, Hugging Face cache sharing, and per-Mac tuning—automatically. It also filters its model catalog to show only what your specific hardware can actually run, which saves you from downloading weights that won’t fit. The “barn” metaphor is apt: it is genuinely just a small shelter for much larger animals.

Key highlights

• 12 MB native Swift app; models live in the standard Hugging Face cache directory.
• Auto-configures optimal settings and offers quantized fallbacks if a model exceeds your Mac’s resources.
• OpenAI-compatible local server works with Chatbox, VS Code, Zed, Cline, Continue, Claude Code, and custom scripts.
• Built-in WebUI available at localhost:2276 while the app is running.
• Experimental network exposure to local devices or specific IPs (e.g., Tailscale), though the developers flag security risks.

Caveats

• You are currently limited to the built-in model catalog; custom model weights are not yet supported.
• Only one model can be loaded at a time, and per-model configuration variations (e.g., different context lengths) are on the roadmap but not implemented.
• Exposing the server beyond localhost is experimental and flagged as a security risk by the developers.

Verdict

Mac users who want a fire-and-forget local LLM backend for their favorite chat UI or editor extension will find this a tidy fit. If you need multi-model concurrency, custom fine-tunes, or non-Mac hardware, this barn isn’t big enough yet.