📦 matraix-ai/matraix-persona-8b

MatrAIx Persona 8B: Testing Apps with 8 Billion Virtual Humans

Imagine testing your next app not with a generic user, but with millions of distinct digital personalities. The MatrAIx Persona 8B project allows developers to simulate realistic human behavior at a massive scale before launching anything to the real world. By creating detailed persona records that mimic diverse psychology and capabilities, this infrastructure lets you run your software through complex scenarios in chat, web, and native app environments. It turns vague user feedback into concrete, reproducible data. You can stress-test features against a population of simulated users to find edge cases you might have missed.

📰 https://news.ycombinator.com/item?id=49370103

🆔 @hackernewsgithubprojects

📦 indigo0445/yt-collate

The YouTube Music Client That Lives In Your Terminal

Stop opening your browser for music. Collate lets you play, search, and manage your entire YouTube Music library right inside your terminal, using only your keyboard. It is a fast, lightweight tool built for Linux that handles playback, playlist editing, and even downloading tracks without ever leaving your code editor. The coolest part is the Vim-style workflow, where you can copy and paste songs between playlists just like text, making library management surprisingly intuitive. If you love efficiency and hate tab fatigue, this is your new favorite daily driver.

📰 https://news.ycombinator.com/item?id=49356485

🆔 @hackernewsgithubprojects

📦 calnode/calnode

Calnode: The Self-Hosted Scheduling Engine for AI Agents

Calnode is a self-hosted scheduling app that runs entirely from a single static binary, meaning no external database servers are required. It combines built-in video meetings with an AI notetaker, all wrapped in a tiny footprint that fits on a five dollar virtual server. The standout feature is its native support for AI agents, allowing tools like Claude to find open slots and book meetings directly through a standard protocol. This removes the need for manual setup or complex API integrations. The codebase is deliberately small, making it easy to audit and trust.

🆔 @hackernewsgithubprojects

📦 osolmaz/pi-workflows

Watch Your AI Agent's Workflows Run Live

pi-workflows is the workflow engine that lets your coding agent run multi-step tasks you can actually watch in real time. You define these workflows as simple code graphs, and the system executes them directly inside your existing chat session, keeping all the context you already built up. The standout feature is the live terminal viewer, which draws a dynamic map of your workflow, showing exactly which step is running, where it paused, and how it will route next. This solves the problem of opaque automation, giving you a clear visual trace of the logic while it executes.

🆔 @hackernewsgithubprojects

📦 analogy-ai/fm-bench

FM-Bench: The Football Game That Tests AI Patience

FM-Bench is a football management game where the real opponent is your own attention span. You run a club for twenty in-game years, making hundreds of decisions with only noisy scout reports and no visible player stats. It is designed to expose where language models fail over long periods, specifically when they forget their own plans or leak money through sloppy tools. The engine is deterministic, so every run is a perfect audit trail. If your agent can keep a coherent strategy across decades of simulated time without the board firing them, you have found a weakness in current AI.

🆔 @hackernewsgithubprojects

📦 luckygirl-xu/awesome-agent-dynamic-graphs

Your AI Agent is a Living Network

That massive wall of text isn't a coding tutorial, it is a map of how AI agents actually change. Researchers just realized that building smarter AI is like growing a living network, not just writing code. They call this view self evolving agents as dynamic graph transformation. Instead of treating an agent as a static box, we see it as a constantly shifting web of memories, tools, and skills. When an agent learns something new, it doesn't just memorize a fact, it physically restructures its own internal wiring. This repository collects the papers proving that this biological style of growth is the future.

🆔 @hackernewsgithubprojects

📦 carterperez-dev/cybersecurity-projects

Cybersecurity Projects: 70 Hands-On Security Challenges

Cybersecurity Projects is the structured collection that turns abstract security concepts into seventy hands-on Python tools. It solves the confusion of knowing what to build by organizing projects into clear tiers from pre-beginner foundations to advanced engineering. The standout feature is the Foundations tier, which uses single-file scripts with heavy teaching comments to explain every line of code to total beginners. This approach ensures you understand the logic behind hash identifiers and header scanners before tackling complex systems. The repository includes full source code and certification roadmaps, making it a complete learning ecosystem.

🆔 @hackernewsgithubprojects

📦 midea-ai/semaplc

Sema PLC: Natural Language to Running Factory Code

Sema PLC takes a plain English sentence like 'latch the motor after a three second timer' and turns it into a real, running factory controller program. It is an open source coding workspace that uses an AI assistant to write, compile, and test industrial control code that normally requires specialized engineering knowledge. The system does not just guess at the answer; it actually deploys the logic to a virtual machine, forces the inputs, and checks if the machine behaves exactly as you described. This project solves the steep learning curve for factory automation by making complex machinery control accessible to anyone who can write a clear instruction.

🆔 @hackernewsgithubprojects

📦 coinve200k/coinve-200k

CoinVE-200K: Editing Videos With Multiple Instructions

CoinVE-200K is the massive video editing dataset that finally lets AI handle multiple changes at once. Most existing tools can only swap one thing in a video, like changing a background or removing an object, but real users want to do several things simultaneously. This dataset from Tencent provides over two hundred thousand examples where a single video gets edited by two to five different instructions at the same time. It includes precise masks showing exactly where each edit happens, ensuring the AI knows to change the subject without messing up the background. It also comes with a matching model and benchmark to test these complex edits.

🆔 @hackernewsgithubprojects

📦 sig-pku/adsl

AI Builds 3D Models From Scratch

Researchers at PKU built a system that doesn't just guess at 3D shapes, but actually writes code to construct them. Instead of generating blurry images, it creates precise digital assets by scripting their geometry. The process works like a team of agents: one plans the structure, another writes the build script, and a third inspects the final render to catch mistakes. If something looks off, they iterate until the model is right. You can type a simple prompt like 'a nightstand' and get a fully articulated, interactive 3D object that you can actually use in games or simulations. It turns plain text into tangible, editable digital hardware.

🆔 @hackernewsgithubprojects

📦 sharelab-sii/va-judger

VA-Judger: Teaching AI to Judge Audio and Video Together

VA-Judger is the first reward model designed to judge audio and video together, not separately. Most AI systems rate sound and pictures using different scores, which misses the real magic of synchronization. This project fixes that by learning from human preferences to spot when a video actually feels coherent. It uses a clever training method to help AI understand why one clip is better than another, rather than just giving a simple pass or fail. The result is much more realistic content where the sound perfectly matches the action. It stops the AI from cheating the system by prioritizing metrics over actual quality.

🆔 @hackernewsgithubprojects

📦 copilotkit/openbot

OpenBot: AI Coworkers With Real Computers

OpenBot gives each AI agent its own literal computer, complete with a private browser, personal files, and isolated login credentials, so it can actually do work without touching your main system. The platform intercepts every single action the agent takes before it happens, checks it against your strict security rules, and records it in an audit log, meaning you always know exactly what it did and why. It is not just a chatbot that talks; it is a governed digital employee that you can hand real tasks to, like opening a website or filling out a form, while keeping full control.

📰 https://news.ycombinator.com/item?id=49365575

🆔 @hackernewsgithubprojects

📦 zenbu-labs/terminal-code

Run VS Code in Your Terminal

Open your code editor right inside your terminal. terminal-code lets you launch the familiar VS Code experience directly from the command line without switching to a web browser. It works by rendering the editor interface using terminal graphics, so you get full editing power while staying in your workflow. The project handles complex setup tasks for you, like importing your existing VS Code settings, keybindings, and extensions with a single command. It also resolves keyboard conflicts between your terminal and the editor automatically. This is perfect for developers who love their terminal but need a robust coding environment. Stop fighting with split screens and just code where you are.

📰 https://news.ycombinator.com/item?id=49371439

🆔 @hackernewsgithubprojects

📦 kartik-3004/facemoe

FaceMoE: The AI Trick for Blurry Faces

FaceMoE is a new model from Johns Hopkins that solves a massive security headache: recognizing people in blurry, low-resolution security footage. Standard face recognition tools fail when images are pixelated, but this project uses a clever 'mixture of experts' trick to fix it. Instead of one giant brain, it uses specialized mini-brains that only wake up when needed to analyze specific parts of a face. This allows it to pick out identity clues even when the picture is terrible. It beats older methods on major benchmarks and is fully open source for you to try.

🆔 @hackernewsgithubprojects

📦 jimmy9704/tasktok

TaskTok: Smarter Image Fixing

TaskTok is the image restoration tool that stops guessing and starts solving specific problems. It fixes blurry or damaged photos by using smart tokens to boost details exactly where your classification, segmentation, or detection models need them. Think of it as a repair crew that only patches the cracks ruining the view, not the whole wall. This approach from the European Computer Vision conference makes your computer vision tasks sharper without wasting effort on unnecessary tweaks. If you want your AI to see clearly, TaskTok is the key to clearer, more accurate results.

🆔 @hackernewsgithubprojects

📦 xdarkzx/reaper-mcp

ReaperMCP: The AI Producer for REAPER

ReaperMCP lets your AI assistant directly drive REAPER to compose and mix entire tracks without touching a single plugin. It’s a local bridge that gives the AI full control over notes, effects, and mastering, meaning it can handle everything from writing a drum beat to applying genre-specific EQ and compression. The coolest part is that it runs entirely on your machine, so your music data never leaves your computer. You just tell the AI what you want, like an eight-bar orchestral arrangement, and it handles the technical setup, mixing, and final polish for you.

🆔 @hackernewsgithubprojects

📦 wang2122/sprix-sage-router

Sprix Sage Router: The Mid-Task Decision Maker

Sprix Sage Router makes a critical decision for autonomous AI agents while they are already working. It acts as a live manager that decides if an agent should continue alone, recruit help, or hand off the task entirely. This solves the problem of agents getting stuck or wasting resources on jobs they cannot finish. It uses live progress data to pick the best path, ensuring tasks are completed efficiently and within budget. This open-source tool is a fascinating look at how we manage teamwork for independent AI systems in the real world.

🆔 @hackernewsgithubprojects

📦 forrestchang/andrej-karpathy-skills

Make Your AI Coder Stop Guessing

Stop letting your AI coder guess your intentions and clutter your codebase with bloated, unnecessary logic. Andrej Karpathy Skills is a set of behavioral rules that forces AI tools like Claude to think before they type, keeping changes simple and surgical. Instead of silently assuming what you want, the AI now asks clarifying questions, pushes back on complex ideas, and only touches the specific code you requested. It turns vague prompts into verifiable goals with clear success criteria. This eliminates the frustrating need to clean up drive-by refactoring and overcomplicated solutions. If your AI code feels like a chaotic intern who won’t listen, this is the fix you need.

🆔 @hackernewsgithubprojects

📦 browser-use/macos-harness

Give an AI Complete Control of Your Mac

Imagine an AI that doesn't just browse the web, but actually operates your computer. macOS Harness is a lightweight tool that gives language models six simple commands to see your screen, click, type, and interact with apps like Spotify or Finder. Instead of needing a specific tool for every application, the AI writes its own code to handle tasks it encounters. It works in the background, so it never steals your mouse or disrupts your work. This project is interesting because it treats your Mac as a raw, open system rather than a locked-down environment.

🆔 @hackernewsgithubprojects

📦 k-dense-ai/mimeographs

Give Your AI Agent a Brain

Mimeographs lets your AI agent think like legendary minds, from Aristotle to Elon Musk. It’s not just a chatbot trick; it’s a way to borrow their actual decision-making habits. The repo packs ready-made mental models, principles, and even anti-patterns for each thinker, so your agent doesn’t just quote them—it reasons like them. Pick a persona, drop it into your setup, and watch your AI tackle problems with a new kind of sharpness. It’s a clever way to level up your tools without writing code from scratch. Next time you’re stuck, ask yourself: who would solve this? Mimeographs helps you find out.

🆔 @hackernewsgithubprojects

📦 diegoscarabelli/garmin-health-data

Garmin Health Data: Your Watch's Secret Database

One command pulls every metric from your Garmin watch into a single, local database file you can actually query. Garmin Health Data solves the problem of scattered health logs by downloading sleep, stress, and activity data from Garmin Connect and storing it in a local SQLite database for deep analysis. This tool is useful because it turns your wearable’s hidden data into an accessible resource without needing heavy cloud infrastructure. It also lets you manage multiple family accounts in one place. The result is a personal health archive that is fully under your control and ready for custom insights.

📰 https://news.ycombinator.com/item?id=47970512

🆔 @hackernewsgithubprojects