📦 vsitzmann/deckwerk

DeckWerk: The Local Slide Editor for Creators

Build a presentation where the video is just as easy to handle as a word document. DeckWerk is a desktop app that treats your slides as real files on your computer, meaning no cloud uploads and no login required. It is built specifically for researchers and developers who need to show off video and image-heavy demos without the friction of traditional software. The coolest part is that you can edit the actual code while watching the changes appear in real time. It feels like a bridge between writing a website and dragging a slide.

🆔 @hackernewsgithubprojects

📦 852wa/jizura

JIZURA: Turn Lyrics Into Motion Videos

Turn raw song lyrics into fully animated motion videos directly in your browser. JIZURA automatically designs the layout, text movement, backgrounds, and effects for every line of a song. The surprising part is how it works: you paste in your lyrics, load an audio file, and simply press a button to generate a completely unique visual style for the whole track. It detects the beat to sync the text with the music, then lets you randomize the design until you find a look you like. Everything runs locally in your browser, meaning your music and lyrics never leave your computer.

🆔 @hackernewsgithubprojects

📦 ybouane/videoflow

Code Your Videos: The VideoFlow Script

VideoFlow turns writing code into making actual video files. Instead of dragging clips around in a heavy editor, you write simple TypeScript instructions to place text, drop in images, and set up animations. The system compiles your script into a portable format that can render an MP4 directly in your browser or on a server. The coolest part is the resolution independence. You design your layout using relative units, so a composition built for a massive forty k monitor scales down perfectly to a standard phone screen without breaking the layout.

🆔 @hackernewsgithubprojects

📦 zhijiewong/openharness

OpenHarness

Your terminal can now run a full AI coding agent that actually lives inside your shell, and it does not force you to pick a single cloud provider. OpenHarness is an open source command line tool that connects to any large language model, whether that is a free local model on your own computer or a paid API. The coolest part is that it treats your codebase like a safe space. Every change the AI makes gets wrapped in a git commit automatically, so if the bot writes something terrible, you can just undo it with one command.

📰 https://news.ycombinator.com/item?id=47600371

🆔 @hackernewsgithubprojects

📦 ektogamat/threejs-conference

Threejs-Conference: Rain That Actually Collides

A browser scene renders a rainy cyberpunk alley where thousands of raindrops physically bounce off rooftops and car roofs without crashing the computer. This project, called Threejs-Conference, solves a common problem: making digital weather look real while keeping it fast enough to run on your phone. Instead of checking every drop against every wall, which is slow and messy, it uses a clever shortcut. The graphics card creates a simple height map of the street, then lets the rain calculate its own bounces directly on the hardware. This means the water behaves correctly without the main processor getting stuck.

🆔 @hackernewsgithubprojects

📦 wladimiravila/esp32s3-distributed-ai

ESP32-S3 Distributed AI

Three tiny circuit boards, each no larger than a coin, can work together to generate human-like text without touching the internet. This project, called ESP32-S3 Distributed AI, splits a massive language model across these three microchips, allowing them to chat over a direct wireless link to complete complex tasks. By dividing the heavy lifting among them, the system solves the problem of running powerful artificial intelligence on hardware that usually lacks the brain power for such jobs. You simply connect to the local network, type a prompt, and watch the boards collaborate to stream text back to your screen.

📰 https://news.ycombinator.com/item?id=49111137

🆔 @hackernewsgithubprojects

📦 xrip/cmp50hx-unlock

CMP 50HX Unlock

The CMP 50HX mining card has been artificially limited to a fraction of its true speed, but a new project called CMP 50HX Unlock reverses that restriction. It works by installing a special firmware patch during the early computer startup phase, which tells the graphics processor to run at its full factory speed instead of the throttled mining rate. This single change boosts the card's raw calculation power by over thirty times, making it perform like a high-end RTX 2080 Ti graphics card. The solution is simple to apply with a single installer for both Linux and Windows, and it also fixes slow data transfer speeds.

🆔 @hackernewsgithubprojects

📦 agentsea/nautilo

Nautilo: Your Own AI Workspace

Nautilo is the self-hosted workspace that puts you in total control of your AI partners. Think of it as a digital room where you bring in your own people and their digital assistants. You give each assistant a unique personality, memory, and voice, then you all write and plan together. Because you run the server and choose the models, your data stays yours. It is not just a chatbot; it is a shared creative space where humans and machines work side by side on real projects. You get the power of AI without giving up your privacy.

🆔 @hackernewsgithubprojects

📦 iapp-technology/openthai-systemone

OpenThai-SystemOne: The Tiny AI Brain

Classify customer complaints, score news topics, and control video games in real-time without generating a single word of text. That is the core promise of OpenThai-SystemOne, a tiny open-source model that answers specific questions about a situation in a single lightning-fast pass. It is built on a Qwen base but replaces the standard text-generation part with a specialized decision head that reads a situation and immediately outputs calibrated probabilities for your specific options. It handles Thai and English natively, making it incredibly useful for routing tickets, checking content safety, or even playing Doom by reading the game state.

🆔 @hackernewsgithubprojects

📦 frontiercs/frontiersmith

FrontierSmith: AI-Generated Coding Problems

An AI system is now inventing brand new algorithmic puzzles that human programmers have never seen before. The project called Frontier Smith solves a major bottleneck in software training by synthesizing open-ended coding problems at scale. Instead of relying on a fixed collection of old textbook exercises, this pipeline uses large language models to create fresh, complex challenges and even writes the test scripts needed to verify if your code works. This is useful because it allows developers to practice on novel scenarios that mirror real-world engineering, rather than just memorizing standard solutions.

🆔 @hackernewsgithubprojects

📦 nokia-applied-research/anyjev

AnyJev Makes LLMs Tell the Truth

A standard language model will change its answer just because you rearranged the list of options. AnyJev from Nokia Applied Research fixes that by turning any open model into a decision maker that gives you a real probability you can actually trust. It works without training, using a clever trick to average out the model’s bias so the confidence score means what it says. This lets you automatically handle a huge chunk of traffic safely instead of sending everything to a human. It is a neat way to make AI decisions reliable without touching the model weights, which is pretty clever for a tool that needs zero labels to start.

🆔 @hackernewsgithubprojects

📦 beniedev/moonbite

Moonbite: The Memory Layer for AI Agents

Moonbite is the continuity layer that gives long-running AI agents a real sense of time. Most AI systems forget everything the moment a chat session ends, which makes them useless for work that spans days or weeks. Moonbite fixes this by adding a structured memory system that remembers past events, tracks current context, and decides when an action actually matters. It does not replace your AI assistant but sits quietly behind it, ensuring the model knows what happened yesterday before it speaks today. If you are building agents that need to stick around, this is the missing piece that keeps them coherent.

🆔 @hackernewsgithubprojects

📦 pinto0309/ocec

OCEC: The 112 KB Eye Detector

The tiniest eye detector in the wild is a mere one hundred and twelve kilobytes, yet it judges if eyes are open or closed in less than a millisecond. Named OCEC, this ultra-lightweight model solves a specific problem: most computer vision tools are too heavy for real-time tracking on low-power devices. By analyzing tiny twenty-four by forty pixel crops, the project provides a training pipeline and several model sizes that balance speed and accuracy. This is useful for anyone building responsive interfaces or accessibility tools where processing power is limited.

🆔 @hackernewsgithubprojects

📦 pulse-energy-eu/slopshape

SlopShape: How to Spot AI Writing

SlopShape lets you spot AI-written business articles without reading a single word. Instead of looking for specific keywords or grammar errors, this research project measures the hidden structural habits of text. The team compared thousands of real human blogs against AI versions written by five different models. They found that AI posts follow predictable patterns in how they organize ideas, which remain visible even if you rewrite the sentences. This tool creates a reliable way to verify content authenticity on the web. If you rely on online articles for business decisions, knowing how to filter out machine-generated filler is becoming an essential skill for staying informed.

🆔 @hackernewsgithubprojects

📦 angerybob/s2-moe

How S2 MoE Makes AI Faster on Small Chips

S2 MoE is the inference engine that finally makes large AI models run fast on small, power-hungry chips. It solves a major bottleneck where running these complex models on devices like the Jetson Orin is usually too slow because of limited memory. The key trick is a smart guessing game where the model predicts its own next words, but only keeps the guesses that actually save time. By carefully choosing which parts of the model to use, it avoids wasting energy on useless calculations. This means you get speeds up to five times faster than the standard approach.

🆔 @hackernewsgithubprojects

📦 androoagi/starnet

Starnet: A Pixel Art Space Station for Real AI Agents

Your AI agents are no longer confined to a chat window. In StarNet, they live in a pixel art space station that mirrors their actual activity in real time. Every room represents a specific task, and every object you place grants a concrete permission. This visual layout is not decoration; it is the exact workflow your agents execute. You bring your own model keys, so your data stays local, while the agents perform genuine work with real tools and costs. You can run multiple specialized agents concurrently, each with its own workspace. The station acts as a live dashboard for your digital crew.

🆔 @hackernewsgithubprojects

📦 alban1911/rose

Rose: How to Get Every Skin in League of Legends for Free

You can play League of Legends with every single skin in the game without spending a cent, and it is all thanks to an open-source tool called Rose. It works by quietly running in the background, watching your champion select screen, and then swapping in the skin you picked the exact moment the game loads. The best part is not just seeing the skins, but the fact that it lets you share them with your friends in real-time through a cloud relay. Your entire lobby can see the same custom models simultaneously, turning your match into a coordinated fashion show.

🆔 @hackernewsgithubprojects

📦 codezeno/claude-code-usage-monitor

Claude Code Usage Monitor: Never Hit Your AI Limit Again

Claude Code Usage Monitor is the Windows taskbar widget that finally tells you exactly when your AI limits reset. If you live in coding agents, you know the panic of hitting a wall mid-task, but this little tool sits quietly in your system tray, reading your local logins to track usage for Claude, Codex, Cursor, and Grok without sending your data to a single third-party server. The coolest part is that it counts down your remaining allowance in real time, so you can pace your prompts like a pro gamer managing stamina. It is completely free, open source, and works on Windows ten or eleven.

🆔 @hackernewsgithubprojects

📦 mvschwarz/openrig

Openrig: Two AIs, One Brain

Openrig is the developer tool that makes two different AIs work as one team. It runs Claude Code and Codex side-by-side, coordinating them to handle complex coding tasks together. This solves the problem of AI agents working in silos, creating a unified system that can plan, build, and review code in a single workflow. If you want to see how multiple AIs can actually collaborate instead of just chatting, this is the project to watch.

🆔 @hackernewsgithubprojects

📦 danielravina/stemkit

StemKit Splits Songs Into Stems Locally

StemKit lets you pull every instrument out of a YouTube song and play them back individually, all without sending your files to the cloud. This desktop app runs entirely on your computer, meaning your private music library never leaves your machine, and you do not need any special accounts or internet keys to start using it. You simply search for a track or paste a link, and the software uses local processing to separate the vocals, drums, and bass into distinct tracks.

🆔 @hackernewsgithubprojects