📦 miaai-lab/qwen3.8-flash-next-single-dgx-spark-tensorfold

Qwen 3.8 Flash Next on a Single DGX Spark

A single desktop-sized computer can now serve a massive AI model that processes five conversations at once, each handling over two hundred thousand words of context, while also analyzing images and videos. The project, called Qwen 3.8 Flash Next on a Single DGX Spark, solves the problem of needing a server cluster to run modern, high-capacity AI locally. By using clever memory management and a set of custom patches, it speeds up reading long prompts by nearly two times without changing a single word of the output.

🆔 @hackernewsgithubprojects

📦 iamunbounded/save-token-jev-clean

save-token-jev-clean

save-token-jev-clean stops your coding agents from summarizing their own work into useless mush. Most tools squash old context into a lossy summary, but this project asks a specialized model exactly which tool calls still matter, keeping your actual code and instructions exactly as they were. It works across different agents like Codex and Claude, so you stop losing crucial details in long sessions. If you want to see how much you are actually saving, there is even a local dashboard tracking every cut. This is a clever way to keep your AI sharp without burning through tokens or forgetting what you asked it to do.

🆔 @hackernewsgithubprojects

📦 using-system/otelyssey

How Otelyssey Automates Plugin Management

Drop your OpenTelemetry plugin into Otelyssey and the repository instantly handles the validation, listing, and release tracking for you. It acts as a self-running marketplace where you simply open an issue, and the automated pipeline checks your submission against the Agent Plugins format before listing it publicly. This removes the manual effort of maintaining compatibility lists, ensuring your tool stays discoverable and up to date without extra work. It is a practical way to manage AI coding agent extensions while keeping the community organized and current. If you build plugins, this system keeps your work visible and maintained. That is the core value of Otelyssey.

🆔 @hackernewsgithubprojects

📦 linxule/mcp-music-studio

MCP Music Studio

You can now turn your AI assistant into a fully functional music production suite that writes sheet music and performs live electronic sets on demand. The mcp music studio is an open source server that lets your favorite AI models compose scored melodies in classical notation and immediately render them as interactive sheet music. It features more than thirty instruments, style presets that instantly add basslines or drums, and a live coding playback environment powered by Strudel and Tidal Cycles. The notes highlight perfectly as they play, allowing you to edit the arrangement on the fly, export standard audio files, or even run synchronized visualizers.

🆔 @hackernewsgithubprojects

📦 lyuyiqi/open-jev-fast

OpenJev-Fast: 6x Faster AI Inference

Turns a slow local AI server into a lightning fast responder by rewriting its core calculations in raw GPU code. While standard setups take over a quarter of a second to answer a simple customer service question, this specialized backend cuts that wait time to just seventeen milliseconds. It achieves this by fusing hundreds of tiny computer operations into a single efficient command, eliminating the lag caused by waiting for data to move back and forth. The result is a drop in replacement that gives you the exact same model answers but with a speed boost so significant it feels instantaneous.

🆔 @hackernewsgithubprojects

📦 wasi-master/wmkeyboard

Wmkeyboard: The Private Android Keyboard

Stop letting apps snoop on your typing. Wmkeyboard is a private Android keyboard that runs entirely on your phone. It supports over three hundred and fifty languages, including complex input methods for Chinese and Japanese, all without sending your data to a cloud. The coolest part? It uses offline intelligence to guess what you are about to write, right on the device. No servers, no tracking, just fast, accurate typing you can actually trust. If you care about privacy, this is the one tool you need to try.

📰 https://news.ycombinator.com/item?id=49922608

🆔 @hackernewsgithubprojects

📦 amywork777/lipflow

Type by Mouth: The Silent Dictation Hack

Lipflow turns your silent mouth movements into typed text using just your Mac’s webcam. Hold the right Option key, silently mouth your words, and they appear at your cursor. This local tool solves the awkwardness of voice dictation by eliminating noise entirely. It runs entirely on your device, using an AI model that reads lips to capture speech without sound. An optional language model fixes any misread words. It is a clever way to type without touching the keyboard or making a sound. If you want private, silent typing on your Mac, this is the tool to try.

🆔 @hackernewsgithubprojects

📦 git-pkgs/proxy

Proxy: The Caching Proxy That Kills Supply Chain Attacks

Proxy is the caching proxy that finally stops malicious package updates before they hit your builds. It works by intercepting downloads from registries like npm and Python's PyPI to store them locally, which speeds up your installs. But the real magic is the version cooldown. Instead of instantly serving the newest release, it hides brand new versions for a set number of days. This simple delay acts as a safety net. If a hacker releases a compromised update, your systems never see it during the quarantine window. It buys you time to verify the package is safe before it ever touches your code.

🆔 @hackernewsgithubprojects

📦 overmind-core/overmind

Overmind: The AI Agent Improvement Engine

Overmind turns the messy real-world data from your AI agents into a continuously improving, high-performing system. You point the platform at your agent's codebase, and it automatically ingests production traces to find where the AI is failing. Then, it uses that data to fine-tune a new model specifically for your needs, without you needing to build any complex machine learning infrastructure. The best part? You get to keep the new model weights. You can download them, retrain them, or roll them back at any time. It is a practical way to make your AI agents smarter using the very data they generate every day.

🆔 @hackernewsgithubprojects

📦 galaxygeneralrobotics/latent

LATENT: The Robot Learning to Play Tennis

LATENT is the robotics pipeline that teaches a human-sized robot to play tennis by watching humans. This project solves the difficult problem of how a machine can learn complex, athletic movements from motion capture data that is often messy and imperfect. It works by simulating the robot’s body inside a computer and training it to copy the subtle balance and timing of a real person’s swing. The truly surprising part is that it does not need flawless data to succeed; instead, it learns to smooth out human error and translate those rough movements into precise robot actions.

🆔 @hackernewsgithubprojects

📦 dmoshehun-prog/learn-from-materials

Learn From Materials: Turn Books Into Interactive Pages

learn-from-materials is the study tool that actually verifies what it teaches. It takes any PDF, book, or slide deck and transforms it into a single, self-contained webpage where every concept links directly back to the exact source page. You get interactive diagrams, built-in quizzes, and a clear distinction between facts from the book and anything the AI adds on its own. Since it runs completely offline, your data never leaves your machine, giving you total privacy. It is the perfect way to turn dense, boring reading into an active, checkable learning session. If you want to study smarter without trusting blind summaries, this is your new favorite workflow.

🆔 @hackernewsgithubprojects

📦 moorew/be-more-hailo

Be More Agent: Your Private AI Friend

Turn your Raspberry Pi into a private AI companion that listens to you, thinks about what you say, and talks back, all without sending data to the cloud. Be More Agent runs entirely on your local hardware, using a specialized chip to process your words so your conversations stay strictly inside your home. The project adds a web interface, letting you chat from your phone or laptop, while the system handles everything from recognizing your voice to generating spoken replies on its own. It even looks at objects through a camera and shares its thoughts, making it a surprisingly capable desk buddy.

🆔 @hackernewsgithubprojects

📦 samwhosung/benilla

Benilla: The Entire Game in Rust

Run the entire classic World of Warcraft client without any original engine code. Benilla is a from-scratch implementation built on the Bevy game engine using Rust. It doesn't just render graphics; it speaks the exact network protocol of the old server, meaning you can log in, quest, fight, and trade just like the year two thousand and seven. The most impressive part is the interface. It runs the actual Lua scripts from your own game installation, so your favorite old addons work perfectly. It’s a massive technical feat that proves you can rebuild a complex system entirely from the ground up.

🆔 @hackernewsgithubprojects

📦 ikaleio/lm-detector

LM Detector: Find Out What Model You Are Actually Getting

LM Detector helps you find out if the AI model you are paying for is actually the one you think it is, by having it write down a bunch of random numbers. It turns out every large language model has a unique numerical fingerprint when doing this, so you can compare that list against a database of known models to see who really answered you. It is a clever way to check if a third-party service is secretly swapping in a cheaper, lesser model without you noticing.

📰 https://news.ycombinator.com/item?id=49824813

🆔 @hackernewsgithubprojects

📦 jsuarezruiz/nova.avalonia.ui

Nova Avalonia UI: The Missing UI Kit

Nova Avalonia UI solves a specific problem for cross-platform developers: finding ready-made, visually polished components that work perfectly on every operating system. While standard frameworks provide basic buttons and text boxes, this library offers a complete toolkit of advanced widgets, including interactive rating stars, scratch card overlays, and dynamic loading animations that automatically match your application's theme. It handles the complex layout math for you, arranging items in circles, arcs, or organic bubble clusters without you writing custom geometry code. This is useful because it turns a blank canvas into a functional interface in minutes, letting you focus on your app's logic rather than fighting with pixel-perfect alignment.

🆔 @hackernewsgithubprojects

📦 yestermester/thesimpsonsgamerecomp

Simpsons Game Recomp: 60 FPS on PC

The Simpsons Game Recomp is the native port that finally runs the 2007 Xbox 360 title at sixty frames per second without an emulator. It solves the problem of running old console games on modern hardware by translating the original code directly into C++, which means your computer runs the game logic natively instead of simulating it. This project is interesting because it uses a tool called RexXGlue to bridge the gap between old console kernels and new graphics drivers, giving you a crisp, high-performance experience on Linux, Windows, or Steam Deck.

🆔 @hackernewsgithubprojects

📦 jakeschincariol/arena-skill

Arena-Skill: 100 AIs Fight for One Answer

Arena-Skill is the Claude Code plugin that turns bad answers into a hundred-way tournament. When you are stuck with a weak response, it spawns a hundred sub-agents, each given the exact same task but a unique strategy card. They pair off, attack each other’s work, and a judge scores the matches until only one solution survives. It is a single-elimination bracket that forces the model to defend, revise, and prove its best thinking. You get the final answer, plus the specific attacks it survived to win. It never touches your project files, so you stay in control. If an answer feels wrong, let the arena decide who is actually right.

🆔 @hackernewsgithubprojects

📦 umamoorg/umamo

Umamo: The Free 2D Rigging Tool

Umamo lets you rig your own 2D characters right on a tablet using a pen, which is the killer feature here. It is an open source editor that opens the exact same files as Live2D Cubism, meaning you can switch tools without converting your hard work. This saves studios and hobbyists from being locked into expensive, closed source software while still getting professional results. The interface is built for touch and pressure sensitivity, making the modeling process feel natural rather than like a computer program. It runs on Windows, Mac, Linux, and Android, so you can work wherever inspiration strikes.

🆔 @hackernewsgithubprojects

📦 ntfargo/relapse-exploit

Relapse Exploit: The PS5 Browser Hack

Relapse Exploit gives you a way to break into a PlayStation Five’s security using just its built-in web browser. Instead of needing a custom hardware mod or a physical console connection, you simply point your console’s network settings to a specific address and load a webpage. That single action triggers a chain of technical tricks that slips past the system’s defenses, eventually giving you full control over the console’s core memory. It is a wild demonstration of how a standard browser can be turned into a backdoor for homebrew software.

🆔 @hackernewsgithubprojects

📦 xiaomingtongxueql/riot-account-checker-manager

Riot Account Checker and Manager

Check hundreds of gaming accounts in minutes without burning through your proxy budget. This open source tool automates the entire verification process for Riot Games accounts, handling the tricky captcha hurdles and sorting your data instantly. The coolest part is how it visually estimates the total value of every skin inventory, turning a messy list of logins into a clear, organized report with price tags attached. It saves you from manual clicks and expensive bandwidth waste, making bulk account management surprisingly fast and efficient. If you handle multiple game profiles, this automated valuation and cleanup system is a must have for keeping your data clean and accurate.

🆔 @hackernewsgithubprojects

📦 rgodha24/walgithub

walgit: The Git Server That Fits Anywhere

A single executable binary can host entire git repositories on a machine too small to store them. Walgit is a git server that points directly at an object store bucket, turning standard cloud storage into a scalable git backend without needing a database. It acts as a disposable cache, meaning every server instance is ephemeral and the bucket holds the source of truth. This design allows you to run the same code on tiny, cheap servers while serving massive monorepos that span gigabytes. By treating the object store as the primary record, the system eliminates the need for complex state management or replication logic.

📰 https://news.ycombinator.com/item?id=49852832

🆔 @hackernewsgithubprojects