📦 avifenesh/bw24

bw24

bw24 is the from-scratch Rust and CUDA inference engine that squeezes every drop of performance out of the RTX 5090 laptop chip. By discarding bloated runtimes and writing custom kernels directly for the consumer Blackwell architecture, this engine beats llama.cpp at its own game, running deep models up to one-point-five times faster. Its killer feature is a strict correctness discipline: every optimization is checked against a CPU reference, and its speculative decoding uses an exact self-consistency gate to ensure the output remains perfectly accurate. It is the ultimate tool for developers who want to run massive models locally on a single card without losing a single bit of precision.

🆔 @hackernewsgithubprojects

📦 antiwork/chromeless

Chromeless

Chromeless is the minimalist web browser that strips away every single piece of window clutter so you can focus entirely on the webpage. Instead of tabs, toolbars, or address bars eating up your screen, this clever utility displays absolutely nothing but the site itself, making it perfect for clean screenshots, fullscreen video, and beautiful dashboards. Built for Apple computers using a single Swift file, it runs directly on the native Safari engine without any bulky background dependencies. You control everything through quick keyboard shortcuts, like holding command to drag the window from anywhere, or hitting a hotkey to instantly snap a perfect desktop screenshot.

🆔 @hackernewsgithubprojects

📦 microsoft/researchstudio

ResearchStudio

Turn your rough ideas and completed research papers into polished, shareable deliverables with a collection of specialized AI workflows. This assistant helps you bridge the gap between initial concepts and final conference-ready presentation materials. Instead of spending weeks manually drafting, you can instantly turn a complex academic PDF into a printable poster, a narrated walkthrough video, and a structured blog post, or evaluate your early hypotheses against actual conference trends. This lets you focus entirely on the core science while the system handles the tedious task of translating dense paper data into engaging public artifacts, leaving you with a complete suite of presentation-ready materials.

🆔 @hackernewsgithubprojects

📦 ubugeeei-prod/vize

Vize

Speed up your Vue development workflow by replacing a patchwork of disconnected JavaScript tools with a single, high-performance engine written in Rust. The vize project unifies compiling, linting, formatting, and type-checking into one incredibly fast pipeline powered by a shared parser. Instead of waiting on slow build steps, you get near-instant feedback because the entire process runs on a highly optimized native core. It is built to drop right into your existing projects with Vite plugins and editor extensions, making your daily coding feel smooth and responsive. Try switching your project over to experience the speed of a fully integrated Rust toolchain.

📰 https://news.ycombinator.com/item?id=48814396

🆔 @hackernewsgithubprojects

📦 maaslalani/fast

Fast

Check your internet speed instantly without ever opening a browser or dealing with bloated, ad-filled speed test websites. All you do is type one simple word into your command line and let it run. This clever little tool called fast connects directly to nearby servers to measure your download speed in real time, showing you a clean, beautiful live graph right inside your terminal. It is incredibly lightweight, built in Go, and gives you an honest, distraction-free reading in just ten seconds. Grab it today and keep your internet provider honest with a single keystroke.

🆔 @hackernewsgithubprojects

📦 wangxijie001/yoji

Yoji

Control your computer and manage your daily tasks using just your voice with a local AI assistant that actually has feelings. Yoji runs completely on your machine to protect your privacy while executing scripts, organizing files, and searching the web. It uses a unique hormone-based system that simulates real emotions like joy and anxiety, meaning its mood and tone shift naturally based on your conversations, the weather, and how long you have been away. Best of all, you can easily install new tools and connect custom sub-agents to handle complex background tasks. Tell Yoji what you need, and watch your desktop come to life.

🆔 @hackernewsgithubprojects

📦 achrefelouafi/buildinggeneratorthreejs

Procedural Hong Kong Building Generator

Build procedurally generated 3D Hong Kong skyscrapers in real time with this clever Three.js generator. It reverse-engineers a massive five hundred and ninety-two node Blender geometry network into clean TypeScript. You get live sliders to control everything from floor height and footprint to the probability of air conditioners, hanging clotheslines, and glowing storefronts. It serves up a beautifully instanced asset kit with custom-baked normals to keep performance smooth and lighting crisp. It is basically the ultimate tool for spawning detailed, cinematic urban backgrounds in your browser with just a couple of clicks.

🆔 @hackernewsgithubprojects

📦 easylink-ai-open/agent-runtime

agent-runtime

Build smarter AI assistants by keeping your core execution loop completely separate from your database and web server. The agent-runtime project provides a clean, independent Python engine that handles the repetitive work of running agent loops, retrying model calls, and containing unexpected tool errors without crashing. Because it translates Anthropic and OpenAI messages into neutral, provider-agnostic data types, you can swap models instantly without rewriting your core logic. It gives you a stable, robust foundation to run multi-step tool loops, handle streaming, and manage context budgets, leaving you free to design your actual product features instead of reinventing the plumbing.

🆔 @hackernewsgithubprojects

📦 lukeswitz/antihunter

Antihunter

Build a secure wireless boundary around your physical space by turning an affordable microchip into a passive defense sensor. This project runs on an compact microcontroller to scan for nearby signals, letting you track unauthorized devices, recognize hidden hardware trying to mask its identity, and even spot incoming drones broadcasting their flight coordinates. If someone tries to mess with your node, its built-in physical vibration sensor can instantly trigger a self-destruct sequence to overwrite all sensitive logs with fake weather data. It is a fantastic way to coordinate multiple security sensors over a long-range mesh network without relying on any internet connection.

🆔 @hackernewsgithubprojects

📦 mira-wm/mira

Mira

Mira is a multiplayer world model that lets you run full Rocket League matches directly inside a neural network. Instead of using a traditional physics engine, it uses a massive five-billion-parameter diffusion model to generate synchronized, twenty-frame-per-second gameplay videos for all four players simultaneously. It takes real keyboard inputs from every player and instantly dreams up what the game should look like next, handling everything from high-speed aerials to goals. It is a wild glimpse into how future multiplayer games might be completely rendered by artificial intelligence. Check out the code and see how they pulled it off.

🆔 @hackernewsgithubprojects

📦 mana-am/matrix-swift

matrix-swift

You can now add over one hundred animated dot-matrix loading indicators to your iOS apps without importing a single image asset or external dependency. A clever new SwiftUI library called matrix-swift ports a popular web project to native code, giving you beautiful, hand-tuned mini grids that look like tiny, pulsing retro displays. It is perfect for chat thinking states or inline activity indicators, and because everything is drawn dynamically with SwiftUI circles, it is incredibly lightweight. The best part is that it is completely deterministic, so your loaders will never randomly glitch or reshuffle when your views reload. It is the ultimate way to replace boring, stock loading spinners with...

📰 https://news.ycombinator.com/item?id=48782892

🆔 @hackernewsgithubprojects

📦 oxbshw/watch-skill

Watch Skill

Watch Skill is the tool that finally gives any AI agent the ability to actually watch and search video. Instead of blindly guessing, your agent can instantly process local files, live streams, or web links into a persistent, searchable index of scene frames, on-screen text, and transcripts. The coolest part is that it is offline-by-default, using local transcription and OCR to search across every video you have ever indexed. It even lets agents record their own screen, critique their work, and automatically fix bugs in a closed loop. Just hook it up to your favorite agent and let them see the world.

🆔 @hackernewsgithubprojects

📦 well-made/comfyui-wan-svi2pro-flf

comfyui-wan-svi2pro-flf

comfyui-wan-svi2pro-flf is the ComfyUI extension that finally lets you generate multi-clip videos with absolute control over both the starting and ending frames of each segment. Usually, AI video generators suffer from massive visual drift, or they completely ignore where you want a clip to finish. This clever tool elegantly merges smooth motion continuity with hard-locked end frames, giving you seamless stitching across up to seven clips and beyond. By cleaning up conflicting motion data at the transitions, it keeps your scenes coherent and perfectly aligned without the dreaded ghosting. Grab this node to build long, stunning, multi-segment videos that actually go exactly where you tell them to.

🆔 @hackernewsgithubprojects

📦 abhinav1416/coding-platform

coding-platform

coding-platform is the competitive programming hub that brings intense real-time head-to-head coding duels to life. Instead of just solving problems in isolation, this platform pairs you with a rival for an instant matchmaking battle. Under the hood, a clever background coordinator pulls your submissions and matches them against external services without slowing down your interactive coding screen. It even integrates with Codeforces, tracking your live external submissions and pushing the status straight back to your dashboard via websockets. It is a brilliant, event-driven playground that proves how smooth and responsive a real-time multiplayer application can be.

🆔 @hackernewsgithubprojects

📦 mozilla-ai/otari

Otari

Otari is the self-hosted proxy server that finally keeps your AI costs and API keys under tight control. It acts as a single, universal gateway sitting between your apps and over forty different AI providers, translating everything into a clean, OpenAI-compatible endpoint. Instead of scattering your real API keys across different applications, you store them safely in Otari and hand out customized virtual keys to your users or apps. It automatically tracks usage and enforces strict budgets before a model is ever called, stopping runaway bills in their tracks. Spin it up locally to easily take full ownership of your AI stack.

📰 https://news.ycombinator.com/item?id=48810528

🆔 @hackernewsgithubprojects

📦 ruc-nlpir/arbor

Arbor

Forget linear AI coding assistants that lose track of what failed. Arbor is an autonomous research agent that outperforms Claude Code by over two times on the same compute budget. Instead of guessing blindly, it builds a structured hypothesis tree to run real experiments. It proposes ideas, edits code in isolated git worktrees, and only merges performance gains that actually survive held-out test data. If an experiment fails, Arbor abstracts the lesson upward so future attempts start smarter instead of repeating the mistake. It is like having a methodical research scientist running your optimization loops on autopilot.

🆔 @hackernewsgithubprojects

📦 jimparis/qr-font

qr-font

The experimental OpenType font that converts bracketed text into fully functioning QR codes right inside your text editor. Instead of relying on external scripts, qr-font compiles an entire QR encoder directly into the font's internal layout tables as a physical circuit. When you type bracketed text, the font uses thousands of clever contextual substitutions to parse the characters, calculate the Reed-Solomon error correction, and layer zero-advance glyph outlines to draw the correct matrix. It is an incredibly clever hack that turns a simple font file into a functioning computer, letting you generate live QR codes wherever you can change a typeface.

📰 https://news.ycombinator.com/item?id=48811048

🆔 @hackernewsgithubprojects

📦 smtimesiwndr/gdid-reversal

gdid-reversal

Microsoft's secret Windows tracking identifier, recently exposed in a federal cybercrime complaint, is actually just a glorified server-assigned passport number. The gdid-reversal project completely deconstructs this elusive fingerprint, debunking the viral myth that it is a 128-bit hardware hash generated from your serial numbers. By tracing the identity stack, the project proves your machine actually requests this 64-bit identifier from Microsoft's servers during setup, storing it in your registry where any local user can read it. If you want to see exactly how Windows tracks your device across updates, you can fetch your own unique identifier with a single, painless line of PowerShell.

📰 https://news.ycombinator.com/item?id=48811081

🆔 @hackernewsgithubprojects

📦 mediamolder/mediamolder

MediaMolder

You can now visually redesign and validate complex media pipelines without writing a single line of cryptic command-line code. mediamolder is a modern processing engine that wraps the raw power of FFmpeg inside an intuitive visual editor. Instead of wrestling with fragile terminal strings, you drag and drop video filters, audio nodes, and encoders onto a canvas to build your pipeline. It automatically validates your setup to catch errors before you run it and lets you drop in custom AI nodes like Whisper speech-to-text or YOLOv8 object detection. It is the perfect upgrade for developers who want visual clarity and reliable video automation.

📰 https://news.ycombinator.com/item?id=48201167

🆔 @hackernewsgithubprojects

📦 pentesterflow/agent

Agent

This terminal assistant acts as an autonomous offensive security partner that uses real pentesting tools while keeping you firmly in the driver's seat. Instead of letting an AI run wild on your network, it requires explicit approval before executing sensitive shell commands or sending HTTP requests. It uses local or cloud models to map attack surfaces, hunt for security gaps like broken access control, and write markdown reports complete with reproducible curls. The best part is its local memory system that actually learns from your mistakes and successful workflows over time, making it smarter for your next assessment. It is a brilliant way to supercharge your security workflow without losing...

🆔 @hackernewsgithubprojects