📦 laravel/vet

Vet: The AI Security Guard for PHP

Vet is the dependency auditor that makes your codebase actually secure. Every time you update your packages, it shows you the exact code changes before they land on your machine. The real magic is that it hands those updates to the coding agent you already have, like Claude or Gemini, and lets it read through the files for you. If the agent finds a security risk, it flags it immediately. If the code looks safe, it passes. You just look at the failures and trust the rest, saving hours of manual checking.

📰 https://news.ycombinator.com/item?id=49701618

🆔 @hackernewsgithubprojects

📦 frankweb33/flybrain-robot-bridge

FlyBrain Robot Bridge: Turning Video Into Motion

A simple webcam can now act as a bridge to control a physical robot without any human input. This project, called FlyBrain Robot Bridge, uses optical flow to track movement in video and feeds that data into a mock neural backend. The system processes these visual signals and sends simple motor commands over a local network to drive servos. It is designed as a proof of concept to test how a digital brain can interface with real-world hardware. The code is open for anyone to experiment with. If you want to see how software can drive motion, this is a fascinating place to start.

🆔 @hackernewsgithubprojects

📦 tau-intelligence/mujoco-drones-gym

MuJoCo Drones Gym

Fly a swarm of quadcopters inside a physics simulator that actually models the wind pushing against them. This project, called MuJoCo Drones Gym, lets you train artificial brains to keep small drones stable and flying in tight formations, even when gusts of air try to knock them over. The really cool part is that it runs on your graphics card, so you can simulate hundreds of these flights at once to learn faster. Instead of guessing how to balance, the computer figures out exactly how to fight the wind just by trying thousands of scenarios in a row.

🆔 @hackernewsgithubprojects

📦 yzhao062/catchbench

CatchBench: The Audit for AI Agent Failures

CatchBench is the benchmark that finally audits agent failures across their entire lifecycle. It solves a specific mystery by asking what an auditor can actually catch at three distinct moments: the declared plan before a run starts, a growing stream of data while it executes, and the finished trace after it ends. By pinning down exactly when evidence is available, it reveals where our current safety nets break down. It is a clever way to test if we can spot a bad agent before, during, or after it messes up.

🆔 @hackernewsgithubprojects

📦 brig-sh/brig

Brig: The Sandboxed Coding Agent

Brig lets you run an AI coding agent inside a tiny virtual machine on your own computer, ensuring it cannot touch anything you did not explicitly give it. The tool creates a completely isolated environment where the agent only sees the specific project folder you point it at. It cannot reach your personal files, your password manager, or any other part of your system. If the AI makes a mess or runs a bad command, the damage stays locked inside that single folder. When you are finished, you simply throw the whole sandbox away and start fresh.

📰 https://news.ycombinator.com/item?id=49709131

🆔 @hackernewsgithubprojects

📦 anuj0456/openarch

OpenArch: LLMs Written for Humans

Every major AI lab hides the secrets of their models in code optimized for speed, making the architecture nearly impossible to read. OpenArch fixes this by rewriting the internal structure of giants like Llama and DeepSeek from scratch into a single, readable file per model. The goal is clarity, not competition. You can finally see exactly how attention mechanisms and expert routing actually work without wading through production optimizations. If you want to truly understand how these models think, this is the clearest map available.

📰 https://news.ycombinator.com/item?id=49693384

🆔 @hackernewsgithubprojects

📦 jub0t/concat

Cut videos locally with Concat

Cut, trim, and caption your videos entirely on your own machine with Concat. This open-source editor gives you the full power of professional video tools without the monthly subscription or the nagging watermarks. The coolest part is that the AI features, like automatic subtitles and text-to-speech, run locally right on your computer. Nothing ever leaves your device, so your private clips stay private while you still get those slick automated edits. It is a native app that feels fast and smooth, giving you complete control over your creative process.

🆔 @hackernewsgithubprojects

📦 bashebr/ai-native-sdlc

AI-Native SDLC: The Agent That Needs Your Permission

Your coding agent can now build your entire app in an afternoon, but it still has to ask you before it touches production. AI-Native SDLC is a ready-to-use skill that turns this idea into a working system. It walks your agent through the full journey from a vague idea to a live deployment, pausing at every critical step to get your explicit approval. It treats the process like a map where every decision is logged and checked. The agent does the heavy lifting, but you hold the keys to the car. It is not magic, it is just a very disciplined workflow.

🆔 @hackernewsgithubprojects

📦 mekion/the-bimo-project

Bimo: The 3D Printable Walking Robot

The Bimo Project is a two-legged robot you can print yourself that actually walks. It is not a toy. It is a real research platform that lets you train a brain in a computer simulation and then plug that exact same logic into the physical robot without any extra tweaking. Usually, moving from digital to real life is a nightmare of adjustments, but here, the robot just gets up and starts walking straight away. You can control it with simple Python scripts or let it run on its own tiny computer board.

🆔 @hackernewsgithubprojects

📦 yesianrohn/water

Water: The 90% Text Recognition Breakthrough

Water reads the wildest, most chaotic text styles found in real-world signs, stickers, and logos with a precision that used to be impossible. This research project, accepted for a major computer vision conference, stops relying on messy, inconsistent real photos to teach computers how to see. Instead, it builds its own massive, synthetic world of artistic text using over eleven thousand unique fonts and tricky layouts like curved paths and warped perspectives. By feeding this clean, controlled data into a specialized model, it finally breaks through that stubborn ninety percent accuracy barrier on the hardest text benchmarks.

🆔 @hackernewsgithubprojects

📦 google/mantis

Mantis: The AI Security Audit Tool

Point an AI agent at your code and let it hunt for bugs. Mantis is a new toolkit from Google that turns standard coding assistants into autonomous security auditors. It does not just guess at problems; it builds a map of your software, finds potential weaknesses, and then actively tries to break its own code to prove the bug is real. Once it confirms the issue, it writes the fix and tests that fix against new attacks. This closed loop means you get verified patches, not just warnings. It is a powerful way to catch security gaps before they reach production.

🆔 @hackernewsgithubprojects

📦 noizefield/audio-plugin-coder

Audio Plugin Coder: Describe, Don't Code

Audio Plugin Coder lets you build professional audio plugins just by describing them in plain English. The surprising part is that it handles the entire creation process for you, from the initial idea all the way to a finished file you can actually use in your music software. It works by guiding AI coding tools through every step, like designing the interface and writing the sound processing logic, so you do not have to learn complex programming languages first. You simply tell the system what you want, and it builds it on your computer.

📰 https://news.ycombinator.com/item?id=49710997

🆔 @hackernewsgithubprojects

📦 filliptm/comfyui-fl-yue2

Compose songs by clicking notes in ComfyUI

A piano roll editor now sits directly inside ComfyUI, letting you click to place notes and drag them to change pitch before rendering full songs. This project plugs the YuE2 music model into the ComfyUI interface, so you can visually draw a melody, add lyrics, and generate forty-eight kilohertz stereo audio from your sketch. It solves the problem of blind text-to-music prompts by giving you precise control over the musical structure before the AI fills in the rest. The editor handles chords, tempo, and sections, turning a blank canvas into a playable score that the model converts into sound.

🆔 @hackernewsgithubprojects

📦 arnegiacomo/fugleramme

The Vintage Bird Frame

Fugleramme turns your kitchen window into a living museum by displaying real, hand-cut bird illustrations from the eighteen-hundreds. This clever Raspberry Pi project listens to your garden using a microphone and local artificial intelligence to identify exactly which species are singing outside. Instead of showing boring data graphs, it prints a beautiful, textured collage of those specific birds onto an e-ink display, sizing them accurately to reflect their actual body weight. The result is a charming, low-power art piece that changes dynamically as the local wildlife visits your yard. It is a delightful fusion of vintage natural history and modern hardware.

📰 https://news.ycombinator.com/item?id=49711544

🆔 @hackernewsgithubprojects

📦 newspicel/sdrminusminus

sdrminusminus: The Modular Software Defined Radio

sdrminusminus treats your radio setup like a visual puzzle where you literally plug signal blocks together on a digital canvas. It is a client server system that lets you connect hardware devices, decoders, and displays without writing code, which solves the problem of complex radio software being too rigid for quick experiments. You can leave the heavy server near your antenna and control it from your browser, making it easy to test different setups. The most interesting part is the built in virtual signal generator, which lets you try the whole system without owning any hardware.

📰 https://news.ycombinator.com/item?id=49695805

🆔 @hackernewsgithubprojects

📦 shengshu-ai/turboserve

TurboServe: The AI Video Server That Saves Money

TurboServe is the serving system that finally stops wasting money on idle graphics cards while generating real-time video. When many users stream interactive AI videos at once, the system faces a tricky puzzle: it must keep every frame arriving on time while juggling different session lengths and sudden traffic spikes. TurboServe solves this by constantly moving user sessions between graphics processors and automatically scaling down the hardware when demand drops, rather than keeping every machine running at full power. This smart scheduling cuts both the worst-case delay for a video chunk and the total electricity bill by nearly forty percent compared to standard setups.

🆔 @hackernewsgithubprojects

📦 jakeatx/llamampere

LlamAmpere: 90 Tokens Per Second on One GPU

One graphics card pushes out ninety tokens per second even while juggling a hundred thousand tokens of context. LlamAmpere takes the popular Llama dot C plus plus inference engine and tweaks it specifically for Nvidia Ampere chips, like the RTX three thousand ninety. The problem is simple: as a conversation gets longer, the computer has to remember every single word it has ever said, and that memory eats up space and slows everything down. LlamAmpere compresses that memory aggressively without losing quality, and uses the model’s own draft head to predict words faster.

🆔 @hackernewsgithubprojects

📦 mystikolab/rustcast

RustCast: The Open Source Spotlight Alternative

RustCast is the open-source launcher that puts your entire computer at your fingertips, built entirely in Rust for speed. If you have used a command bar to quickly open apps or run calculations, this is the free, customizable version that lives in your menu bar. The single most impressive thing it does is window tiling, meaning you can snap any window to specific screen positions using a simple hotkey, which is usually a paid feature in other tools. It handles clipboard history, math, and file search with a snappy feel that proves Rust is a serious contender for desktop apps.

📰 https://news.ycombinator.com/item?id=49713307

🆔 @hackernewsgithubprojects

📦 joaquinruiz/specjudge

SpecJudge: Pick the right AI model

SpecJudge is the command-line tool that stops you from wasting money on the wrong AI model. You know how picking the right model usually feels like a guess? This fixes that. It reads the plan and specs you already wrote for your project, then uses a small AI running entirely on your own computer to figure out exactly how hard the work is. It compares that difficulty against every model you have access to, checking real prices and capabilities. The coolest part is that it demands proof.

📰 https://news.ycombinator.com/item?id=49206759

🆔 @hackernewsgithubprojects

📦 ethanplusai/jarvis

JARVIS: Your Mac's New Voice Assistant

JARVIS turns your Mac into a talking partner that actually builds software for you using the Claude Code you already pay for. Instead of typing commands, you simply talk to your computer, and it answers out loud, walking through your ideas with you until you settle on a plan. Once you agree, it writes the design down as a file, then kicks off a real coding session to handle the heavy lifting.

📰 https://news.ycombinator.com/item?id=49714089

🆔 @hackernewsgithubprojects