📦 pku-value-lab/omnireasoning

OmniReasoning: When Machines Actually Watch

Teach an AI to actually watch a video, not just stare at frames. OmniReasoning is a research toolkit that builds models capable of joint audio-visual reasoning, meaning they understand how sound, speech, and visuals connect to answer complex questions. The project includes a data engine called OmniQA that synthesizes millions of training samples with deep thinking traces, alongside a new benchmark to test if models can reason beyond what is strictly shown on screen. It solves the problem of AI hallucinating or missing the audio context that defines a scene. If you want to build or evaluate systems that truly comprehend multimodal content, this is your starting point.

🆔 @hackernewsgithubprojects

📦 ubugeeei-prod/corsa-bind

Corsa Bind: Native TypeScript Power for Any Language

Corsa Bind is the bridge layer that lets any programming language talk to the new native TypeScript compiler without breaking anything. It solves the problem of wanting fast, type-aware code checking in Rust, Go, or JavaScript without having to rewrite the entire compiler logic from scratch. The coolest part is that it keeps the heavy lifting in Rust for speed, then hands that performance to your favorite language through simple, stable interfaces. This means you can write custom linting tools or checkers for your projects without reimplementing how TypeScript understands code. It just works, stays fast, and keeps your code safe across different platforms.

🆔 @hackernewsgithubprojects

📦 hypersniper05/mcp-image-generator-uncensored

MCP Image Generator Uncensored: Private AI Image Creation

MCP Image Generator Uncensored is the self-hosted tool that puts full control of AI image creation back in your hands. This project runs a local server that generates and edits pictures without the usual content filters found in cloud services. You can create detailed scenes, remove backgrounds, or fix watermarks, all on your own computer. It connects easily to other tools through a standard protocol, making it a flexible addition to your creative workflow. The biggest draw is the freedom to create exactly what you want, without external rules getting in the way.

🆔 @hackernewsgithubprojects

📦 hoangp8/tokless

Tokless: The 30-Second Setup for AI Coding Agents

Tokless cuts the time to set up an AI coding agent from hours to under thirty seconds. It solves the messy problem of wiring the best coding tools into your AI assistant without any manual configuration. You just run one command, and it automatically connects tools for code indexing, context memory, and prompt compression directly into your agent. It works across all major platforms and coding assistants, handling the complex routing behind the scenes so you don't have to touch a single config file. It’s like having a professional install your entire AI toolkit in the blink of an eye.

🆔 @hackernewsgithubprojects

📦 silvertree2010/apolloshell

ApolloShell

ApolloShell lets you decorate your Mac desktop with a customizable sidebar that acts like a personal control center. It does not replace your file manager but adds a layer of widgets where you can pin a dock, search for apps, and monitor your system stats from a top panel. The standout feature is how easy it is to change the look; you simply edit one plain text file to pick colors and fonts, and the entire interface updates instantly without restarting anything. Because it is built from scratch in Swift for the latest macOS version, it feels fast and native.

📰 https://news.ycombinator.com/item?id=49977347

🆔 @hackernewsgithubprojects

📦 professorpalmer/mirai-s-ada

A 27 Billion Parameter Model on a 12 GB GPU

A 27 billion parameter AI model now runs entirely on a standard 12 gigabyte graphics card, delivering speeds that previously required much more expensive hardware. Mirai S Ada solves the problem of memory limits by using a clever tiered cache system that moves data between fast video memory and regular system RAM as needed. The real surprise is how fast it generates text when rewriting files you have already seen, because it copies existing context almost instantly instead of reprocessing it. This project is useful for developers who want powerful local AI without buying a server grade machine.

🆔 @hackernewsgithubprojects

📦 realskyrin/ns-game-apple-arm-port

ns-game-apple-arm-port: Native Switch Game Porting

ns-game-apple-arm-port is the agent skill that translates Nintendo Switch games into native Apple Silicon code. This project acts as a detailed instruction set for AI assistants, teaching them how to rewrite Switch instructions into Rust and Metal graphics. It solves the complex problem of making console games run directly on Mac and iPad hardware without emulation. The standout feature is its strict evidence-based validation, which ensures the game actually plays rather than just compiling. It provides a clear, structured path for developers to verify every visual and audio detail. This makes a huge leap in performance possible.

🆔 @hackernewsgithubprojects

📦 respanai/respan-gateway

Respan Gateway: One Endpoint for All LLMs

Point your existing chat client at a single local server and swap out the model underneath without touching your code. Respan Gateway is a tiny Rust binary that sits between your app and every major language model provider, translating requests so an OpenAI script can call Claude or Gemini as if nothing changed. The coolest part is the fallback routing: if the primary model is down, it automatically retries on a backup provider so your users never see an error. It handles images, voice, and tools too, all behind one simple URL. One server, every vendor. Your code stays exactly the same.

🆔 @hackernewsgithubprojects

📦 liqiiiii/jeva

Jeva: The AI That Doesn't Blink

Jeva is a four-billion parameter model that processes two thousand distinct options in a single pass without ever losing track of their order. While standard AI models treat text as a linear stream where position dictates meaning, this system builds a decision engine from first principles that treats every candidate as an independent, reusable memory unit. This means the model can rank thousands of policy documents or support tickets by scoring them individually, then caching those representations to assemble new, dynamic decision sets instantly.

🆔 @hackernewsgithubprojects

📦 nx-nass/nasscad

NassCAD: CAD Without the Cloud

Model complex machinery right in your browser without sending data to a cloud server. NassCAD runs entirely on your computer, handling 3D shapes and engineering files locally for total privacy. The most interesting feature is its script console, which lets an AI assistant write code to build detailed mechanical parts like a four-cylinder engine. You paste that code into the app to generate and animate the model instantly. It is a personal project focused on local processing, meaning your designs never leave your device. If you want to create 3D models with full data ownership, this tool gives you the power without the subscription.

🆔 @hackernewsgithubprojects

📦 spovalova/openterminal

Open Terminal: The AI Trader That Actually Works

Open Terminal lets you ask a plain English question and instantly see a live workspace of charts and data tables that answer it. Unlike generic chatbots that just guess numbers, this tool proves every figure it gives you by pointing directly to the source chart on your screen. It runs entirely on your own computer, meaning your private trading data never leaves your machine to some outside server. This solves the annoying problem of juggling eight different tabs to find one specific piece of market information. Instead, you get a single, connected view where you can actually edit and rerun the data yourself.

🆔 @hackernewsgithubprojects

📦 fulcrumgenomics/riker

Riker: The Fast Sequencing QC Tool

Riker is a command line tool that analyzes DNA sequencing data up to thirty eight times faster than the standard software used for it. It takes a single file containing your genetic data and checks it for quality issues, like coverage gaps or sequencing errors, in one quick pass. Instead of requiring a heavy Java environment or large installations, it runs as a tiny, standalone program that needs no extra setup. This makes it incredibly easy to install and run on any computer. For anyone working with sequencing data, this tool turns a slow, cumbersome check into a fast, simple step that gets out of your way.

🆔 @hackernewsgithubprojects

📦 ikow/php2shell-poc

The PHP Flaw: php2shell-poc

A single file upload request can now execute code on any unpatched PHP server without needing a password or login. This repository, php2shell-poc, documents two critical flaws in PHP's built-in upload tracking feature that have existed since twenty twelve. The researcher found that the system stores file names exactly as sent, allowing attackers to hide dangerous code inside a file name that the server later reads as instructions. The project includes a ready to use lab environment and scripts to verify these flaws safely. If your application runs on PHP, you need to update to the latest patch immediately.

🆔 @hackernewsgithubprojects

📦 greentfrapp/testreel

Testreel: Programmatic Demo Videos

testreel is the video tool that turns your LLM agent into a professional screen recorder. It solves the headache of manually re-recording demos by letting you define interactions in a simple config file, so you can rerun the script whenever your app changes. The coolest part is how it builds the visual polish for you. Instead of just a flat screen capture, it overlays an animated mouse cursor with satisfying click ripples and frames your app in a clean, modern window. You get a finished, studio-quality product video without touching a single video editor, which is huge for showing off your web apps.

📰 https://news.ycombinator.com/item?id=47676489

🆔 @hackernewsgithubprojects

📦 opensource03/harnss

Harnss: See What Your AI Is Actually Doing

Run multiple A.I. coding agents side by side in a single desktop window, watching exactly what each one does in real time. Harnss turns raw code edits into readable side by side comparisons and shows live terminal output, so you see the changes happen before they save. Instead of guessing what a tool is doing, you get clear visual cards for every action, making it easy to trust the results. This solves the problem of blind faith in automated coding by giving you total visibility. If you want to understand what your agents are truly building, this is the window that shows you the work.

🆔 @hackernewsgithubprojects

📦 pvrlabs/aibadger

AI Badger: Stop Burning Tokens on Your Codebase

Stop burning your token budget by pasting your entire codebase into a chat bot. AI Badger is a local command line tool that acts like a smart gatekeeper for your code. It scans your project to find only the files that actually matter for your specific question. Instead of feeding the AI a novel of irrelevant code, you give it a tight, focused summary. You stay in control of what gets copied, and the AI gets a precise answer without the noise. It is the quickest way to get high quality code reviews without paying for context you do not need.

🆔 @hackernewsgithubprojects

📦 anabologyco-maker/claudia-bot

Claudia Bot: The AI Popstar Factory

Claudia Bot automates the production of high-quality music videos for an open character named Claudia. It guides a coding agent through the entire workflow, from generating song lyrics and analyzing beat maps to generating video clips and locking lip sync to the audio. The system ensures every sung word appears on screen at the exact right moment, turning a simple track into a broadcast-ready visual experience with measured precision. You control the creative direction while the software handles the technical heavy lifting of rendering and editing. This toolkit proves that complex video production can be managed by software agents with strict quality checks.

🆔 @hackernewsgithubprojects

📦 artcraft-android/lightcraft-android

LightCraft: The Open Source Lightroom Alternative

LightCraft is the open source photo editor that finally brings professional-grade image processing to your Android phone without the subscription. It is a complete reimplementation of the Adobe Lightroom engine, written in Rust to run directly on your device for speed and privacy. The most surprising part is that it handles complex tasks like noise reduction and color grading entirely locally, meaning your photos never leave your phone. You get the full suite of editing tools, from basic exposure adjustments to advanced masking, all wrapped in a clean, modern interface.

🆔 @hackernewsgithubprojects

📦 oaleex/shaolinmonksrecomp

Shaolin Monks Recomp: No Emulator, Just Native Code

A twenty-year-old Xbox fighting game now runs directly on your modern computer without a single line of emulation code. Shaolin Monks Recomp takes the original hardware instructions, translates them once into native machine code, and lets your graphics card render them at high resolutions using the latest driver standards. Instead of simulating the old console, this project bridges the gap between legacy hardware and modern performance. You just point it at your legally owned disc image, and the full campaign plays smoothly on Windows.

🆔 @hackernewsgithubprojects

📦 lululynne/maimai-xianpin

麦麦现拼: Strangers combine McDonalds orders

McDonalds is letting strangers combine their individual orders into one big family meal for a discount. This project, called Mai Mai Xian Pin, was built to test that idea. It connects to the real McDonalds menu system to check actual prices and available items. Then it uses its own logic to pair up people at the same store who want to share a larger bundle. This lets them save money on the combined cost while still getting their own specific pickup codes. It solves the problem of expensive single orders by creating a simple, local matching system.

🆔 @hackernewsgithubprojects