JustVugg/colibri is built on a stubborn idea: a model with hundreds of billions of parameters should run on the hardware you already own. It is a pure C inference engine with zero dependencies that streams the experts of a mixture-of-experts model from disk, treating VRAM, RAM and storage as one memory hierarchy. The README shows a 744B model answering in about 32 seconds on an ordinary desktop, and a built-in web dashboard visualizes which expert tiers are active during a conversation. Treat it as a research platform rather than a finished product, since the authors promise correct semantics, not speed.
webadderallorg/Recordly targets anyone who records product walkthroughs but does not want to learn video editing. The desktop app captures your screen, then suggests zooms, smooths the cursor, places everything inside a styled frame with wallpaper and shadow, and lets you pin a webcam bubble wherever it fits. A drag-and-drop timeline handles trims, speed changes and annotations, work saves as project files you can reopen, and an extension marketplace adds extras such as click sounds and device frames.
People juggling several codebases at once will recognize the problem yandanp/Connexio solves: terminal windows pile up until nobody remembers which server runs where. This Tauri and Rust app gives each project its own workspace with tabs that survive a restart. It spots runnable scripts in package.json, Makefile and similar files so a common task becomes a single click, and it wraps in SSH connection management, git status, a small code editor and a file tree. A practical first step is moving your two most active projects in and counting how many stray windows disappear.