Ooor is a 1.5MB MIT-licensed desktop manager that turns llama.cpp into a local AI workstation

What Ooor is

According to a post on dev.to by the tool's developer, Ooor (pronounced "O-or") is a free, MIT-licensed Windows application that wraps the llama.cpp command-line ecosystem in a graphical workstation. The executable weighs roughly 1.5MB and is built with C# and WinForms on .NET Framework 4.8, a runtime that ships with Windows 10, so there is no Electron layer, no Node dependency and no large framework download. The developer says it launches instantly and idles in single-digit megabytes of RAM.

The app folds four roles into that one binary: an engine manager that installs and switches llama-server builds for CPU, CUDA, Vulkan and SYCL; a model library that scans the disk for .gguf files; a download manager paired with a Hugging Face model browser; and a streaming chat console that doubles as a local agent host.

Local-only by default

Privacy is the design boundary. The model service binds to 127.0.0.1, there is no telemetry, no account system, and the only outbound traffic happens when a user explicitly asks an agent to fetch a URL. Once a service is running, Ooor exposes an OpenAI-compatible endpoint at 127.0.0.1:6080, so clients that accept a custom base URL — the post names Cursor, the Continue extension for VS Code and Cherry Studio — can point at local models with a one-line configuration change.

For context, the developer's comparison table puts Ollama at around a 150MB installer with roughly 50MB of idle memory, and LM Studio at more than 500MB installed and about 800MB of RAM at idle due to its Electron base. Those figures are the author's own rather than independent measurements, but the install-size gap is the pitch: Ooor leans on an OS-bundled runtime instead of shipping one.

Engines, models and downloads

llama.cpp releases move quickly, so Ooor decouples the engine from the GUI. It downloads official release archives, extracts them into a config directory, lets multiple versions coexist and switches between them with one click — an engine upgrade never forces an app upgrade, and vice versa.

The model library presents a table of filename, multimodal projection file (the mmproj companion used by vision models), disk location, size and free-form notes, marking each entry as internal or externally referenced. A soft-delete option removes a model from the list while keeping the file on disk, so users comparing quantizations can switch back without re-downloading multi-gigabyte weights.

The Hugging Face browser sorts repositories by downloads, likes, task category, gated status and last update, lists every GGUF variant in a repo with an estimated hardware requirement per quantization, and can route through the hf-mirror mirror or GitHub proxies. The downloader handles parallel multi-chunk transfers with per-chunk status, automatic retries and resumption after the app is closed and reopened.

Chat, agents and profiles

The chat console streams responses and renders a tokens-per-second sparkline plus chunk timing for each reply, which makes it possible to compare inference speed across engine builds or quantization levels without running a separate benchmark.

The agent layer is the feature the developer highlights most. Built-in tools cover fetching a URL, reading and writing files, listing directories, executing shell commands and a key-value memory that persists across turns. There are guardrails: reads and directory listings are confined to allowed root directories, while file writes and shell commands require manual confirmation before they run. Ooor also supports the Model Context Protocol, so external MCP servers sit alongside the built-in tools, and the repository includes a sample SQLite MCP server. Profiles tie things together by snapshotting engine, model, runtime parameters and agent bindings — for instance a small CPU model for quick questions versus a large CUDA model with tools attached for code work. The interface supports English and Simplified Chinese, switchable at runtime.

Why it matters

Local LLM tooling has largely settled around Ollama and LM Studio, and Ooor argues for a third path: a featherweight front end that matters on older or constrained hardware and for anyone wary of Electron's memory cost. Its decoupled engine management is a practical answer to llama.cpp's release cadence, and its confirmation-gated agent layer with MCP support sketches a fully offline assistant that can read local files, fetch a page and write results without a cloud service in sight.

The caveats are worth stating plainly: everything here comes from a single self-published post by the developer, so the size and performance numbers are the author's own; the app is Windows-only, unlike cross-platform rivals; and the hardware-requirement estimates for quantizations are heuristics. Windows users who want a minimal llama.cpp workstation can judge for themselves — the source lives on GitHub under rhettli/Ooor-desktop, with a project site at ooor.cc.