📦 zyphra/zuna

Zuna: The Brain Wave Model That Fills In Missing Data

Zuna is the open-source EEG model that turns incomplete brain recordings into complete, clean signals using only electrode positions. Zuna learns the physical layout of sensors to reconstruct missing data, denoise messy recordings, and even predict what your brain was doing at spots where no sensor existed. Trained on millions of hours of data, it fits on a single consumer graphics card, making complex neuroscience research possible without expensive lab equipment. It is not a medical device, but a powerful tool for researchers to clean up brain data. The model handles everything from sparse consumer headbands to dense research caps by understanding where each sensor sits on the scalp.

🆔 @hackernewsgithubprojects

📦 wadezhao23/loom-novel

Loom Novel: AI Writing That Learns Your Style

Loom Novel turns your local computer into a personal novel-writing factory that actually learns your unique voice. Instead of generic AI text, this tool uses a team of specialized agents to draft chapters based on your world building and character cards, then watches you edit the output and memorizes exactly how you changed it. You simply paste a few paragraphs of your own writing to set the tone, let the system generate a chapter, tweak it by hand, and hit a learning button. The magic happens next time: the AI writes the following chapter with your specific phrasing and rhythm baked in.

🆔 @hackernewsgithubprojects

📦 microsoft/resource2skill

Resource2Skill: Turn Tutorials into Agent Skills

Microsoft Resource2Skill turns human tutorials into executable skills that AI agents can actually run. Instead of just reading text, the system digests video guides and articles to build reusable instructions for real software tools. An agent can then browse these distilled skills to generate web pages, design PowerPoint decks, create Excel workbooks, model 3D scenes in Blender, or even compose music tracks in audio software. It effectively transforms passive learning resources into active, automated workflows. This approach lets artificial intelligence agents perform complex tasks across diverse domains by combining and executing specific, proven techniques rather than guessing from scratch.

🆔 @hackernewsgithubprojects

📦 ohad6k/emulo

Emulo: Mine Your Coding Logs Into a Personal Agent Profile

Emulo reads your past coding sessions to build a private profile that teaches AI agents exactly how you work. Most agents start every task from zero, forgetting your specific debugging style or design preferences. Emulo fixes this by mining your history from tools like Claude Code and Codex, extracting the unwritten rules you actually follow. It creates a you.md file that loads before work begins, so your agent codes like you rather than a generic template. The process stays local, keeping your raw logs on your machine and only sending redacted text for processing.

🆔 @hackernewsgithubprojects

📦 mitkox/thinking-with-visual-primitives

Thinking with Visual Primitives

DeepSeek’s Thinking with Visual Primitives solves a sneaky problem where AI models get lost in complex images because natural language is too vague. Instead of just seeing clearer, this tool teaches the model to point while it reasons by using points and boxes as actual units of thought. It mimics how humans trace mazes with a finger to anchor abstract concepts to concrete coordinates. This approach drastically cuts down on the massive amount of image data the AI needs to process, letting it solve tricky spatial puzzles with surprising efficiency and accuracy that rivals top-tier commercial models without drowning in token costs.

🆔 @hackernewsgithubprojects

📦 sthamann/nyx-local-ai

Nyx Local AI: Offline Coding Agent

Nyx Local AI runs a full coding agent entirely offline right inside your editor with zero token costs. This extension automatically finds your local Ollama or LM Studio models and lets you chat, write code, and run commands without ever sending data to the cloud. You can even connect to other machines on your network to use their GPUs while keeping everything private. It gives you a Cursor-like experience on your own hardware, handling everything from code edits to file management securely. You get total control over your data and avoid monthly fees while still using powerful local models for real work.

🆔 @hackernewsgithubprojects

📦 kunchenguid/whathappened

whathappened: The X Briefing Agent

whathappened is a specialized tool for Grok Build that turns Twitter chaos into a structured, neutral briefing of what just happened, who is saying what, and where the real debates are happening. Instead of drowning you in screenshots, it actively searches public conversations to map out distinct opinion camps, tracks live arguments with direct post links, and filters out noise using an adaptive time window that adjusts based on how fast a story is moving. It is genuinely useful because it gives you the full picture of public sentiment and factual developments in one clean read, saving you from scrolling through hundreds of conflicting takes.

🆔 @hackernewsgithubprojects

📦 kiwi-init/bonsai

Bonsai AI: Turn Messy Thoughts into Ready-to-Code Prompts

Bonsai turns messy notes into ready-to-code prompts. This native macOS app gives you a single canvas to sketch ideas with shapes, attach local files or web tabs, and let an AI agent reshape the board live. The standout feature is a built-in semantic linter that checks if your prose is clear enough for automation. It quietly flags vague sentences while you type, ensuring you have a solid plan before you ever paste anything into a coding tool. It keeps the path from thought to action short and frictionless.

🆔 @hackernewsgithubprojects

📦 therealyt/git-knife

Git-Knife: Edit Commit Dates Easily

Git-Knife is the desktop tool that finally lets you edit commit dates and author details with a simple table view. Most GUIs lock those timestamps down, but this one opens them all up so you can fix your history without guessing. You get a clean list of commits where you can tweak messages, names, emails, or exact dates in one go. The coolest part is the bulk find and replace feature that scans your entire log for things like old email addresses and swaps them everywhere at once. It previews every change first and saves a backup before it does anything, so you never break anything.

📰 https://news.ycombinator.com/item?id=49259611

🆔 @hackernewsgithubprojects

📦 haiyuwu/visreg

VISReg: Self-Supervised Visual Learning

VISReg is a Python package that teaches AI models to understand images without any manual labels. It uses a technique called Variance-Invariance-Sketching Regularization to stop neural networks from cheating during training. Most self-supervised methods eventually collapse, meaning the model learns to ignore the actual visual details. VISReg adds a mathematical check that forces the network to pay attention to differences in the data. This makes the model much more robust, especially when dealing with messy or rare images. The project includes pre-trained weights and evaluation scripts for tasks like classification and segmentation. It is a practical tool for researchers who want better visual representations without the hassle of heuristic tuning.

🆔 @hackernewsgithubprojects

📦 william-lu-stack/flawless

Flawless: AI SRE for Kubernetes

Flawless is the AI infrastructure platform that actually proves its own fixes worked. Instead of just offering vague advice, it acts as an autonomous SRE that gathers real-time evidence from your logs and metrics to diagnose failures. It then drafts a safe recovery plan, waits for your explicit human approval, and executes the fix in a controlled environment. The truly surprising part is that it does not stop at running the command; it actively verifies that the service is truly healthy before closing the ticket. This means every action is auditable, every fix is verified, and you get a complete record of exactly how the system healed itself.

🆔 @hackernewsgithubprojects

📦 codejunkie99/agentic-stack

One Brain for Every Coding Agent

Most coding tools treat your AI assistant like a goldfish with a five-second memory, wiping the slate clean every time you restart. The Agentic Stack solves this by giving you a portable brain that lives in a simple agent folder and travels with your project. You can use it with Claude, Cursor, or Windsurf, and your AI keeps its memories, skills, and preferences intact no matter which tool you switch to. It also includes a local dashboard to track costs and activity, plus a clever flywheel that turns your approved work into reusable learning material without training a new model.

🆔 @hackernewsgithubprojects

📦 haydenbleasel/blume

Blume: Zero-config docs that just work

Blume is the documentation generator that turns a folder of Markdown files into a production-ready website without touching a config file. You simply drop your notes into a directory, run one command, and it silently spins up a hidden Astro project to handle navigation, search, and mobile design for you. The real magic is how AI-friendly it is: it automatically generates special text files and a server endpoint that let coding agents search, read, and even update your documentation directly. It even includes skills to teach AI tools how to maintain the site for you.

🆔 @hackernewsgithubprojects

📦 0x0funky/audioghost-ai

AudioGhost AI

AudioGhost AI is the audio splitter that lets you surgically isolate any sound using simple text prompts. Instead of manually picking tracks, you describe exactly what you want to hear, like vocals or a dog barking, and the AI listens through your whole song to extract just that specific part. It works with audio files and even video uploads, handling the heavy processing efficiently by using a lightweight version of Meta’s powerful model that runs smoothly on standard graphics cards. This means you can clean up noise or pull out isolated instruments without needing a supercomputer or complex coding skills.

🆔 @hackernewsgithubprojects

📦 milesdeutscher/garchmethod

GARCH Method: The Volatility Targeting Skill for Claude Code

GARCH method is the volatility forecasting tool that finally answers how much you should trade instead of guessing which way the market moves. Real risk desks use this Nobel Prize winning machinery to size positions based on how violent the market is likely to be rather than trying to predict direction. It installs as a simple Claude Code skill that runs a walk forward calculation to forecast daily volatility and classifies whether conditions are calm normal or stormy. You can ask plain English questions about any asset like Bitcoin or stocks and it instantly calculates whether you should size up or down with a hard safety cap on your exposure.

🆔 @hackernewsgithubprojects

📦 danielcardonarojas/codemark

Codemark: Semantic Code Bookmarks

Codemark bookmarks code by structure instead of fragile line numbers. It uses tree-sitter to anchor your notes to named functions or blocks, so your references survive any refactor or reformatting that shifts the text. You can install a simple skill into your favorite AI coding agent, and it automatically learns to create and recall these structural markers for you. This means the context you built in one session stays alive in the next, even if the code moves around. You can then ask the agent to walk you through a specific flow, like a request lifecycle, and it will pick up exactly where you left off.

🆔 @hackernewsgithubprojects

📦 rayfish/rayfish

Rayfish: Private Mesh VPN Without a Server

Create a private network connecting your laptop, phone, and friends' computers anywhere in the world, with zero servers to host. Rayfish lets you share a simple code and instantly link machines together, using a peer-to-peer mesh that punches directly through NATs for fast, encrypted connections. It solves the headache of complex remote access by giving every device a stable name, so you can reach a friend's machine by typing just their name instead of hunting for IP addresses. The real magic happens with its built-in firewall and SSH support, which lets you securely log into devices without managing any SSH keys or certificates.

🆔 @hackernewsgithubprojects

📦 activeing123/mcptoon

Cut AI Context Waste with Mcptoon

mcptoon is a zero-dependency CLI that stops MCP servers from eating your context window. Instead of stuffing thousands of JSON tool schemas into your AI chat, it keeps them stored locally and returns ultra-compact text. It saves up to ninety-seven percent on discovery and fifty-six percent on results using a special encoding called Toon. One config file works for every agent, and your AI can even add new tools itself without you touching a JSON file. It works on Windows, Mac, and Linux with zero setup headaches. Stop letting tool definitions ruin your conversation flow.

📰 https://news.ycombinator.com/item?id=49253721

🆔 @hackernewsgithubprojects

📦 datagallery-lab/datafoundry

DataFoundry: Auditable AI Data Analysis

When teams let AI query enterprise databases, the real worry is never if the model writes SQL. It is whether it understands business definitions, avoids mutating production data, and keeps every step auditable. DataFoundry is an open-source workbench that turns natural-language questions into safe, verifiable data tasks. It connects to dozens of databases and enforces read-only boundaries, so credentials never leak into prompts. The system inspects your schema, aligns terms like GMV to real tables, and executes queries inside a governed runtime. Every step, from the first SQL statement to the final chart, is persisted and replayable. You can branch sessions, reference previous outputs, and export reusable assets.

🆔 @hackernewsgithubprojects

📦 lucid-softworks/lumen

Lumen: A Zero-Dependency JavaScript Engine in Rust

Lumen passes every single JavaScript standard test—fifty-three thousand four hundred out of fifty-three thousand four hundred—built entirely from scratch with zero external libraries. This Rust project does not rely on third-party crates for parsing, compiling, or running code, which means it stays completely self-contained and lightweight. It features three execution modes, including a fast native code compiler for Apple Silicon and Linux, and even mimics the Node.js environment so you can run real-world tools like Vite. It is a fascinating look at what happens when you strip away all dependencies to see exactly how a JavaScript engine works under the hood.

🆔 @hackernewsgithubprojects

📦 incy-dev/incy-platforms

Incy Platforms: Privacy Proxy

Manage multiple proxy servers across iOS, Android, and macOS with a single app that handles subscription updates and complex routing rules. Incy Platforms lets you split traffic by domain, block unwanted sites, or route specific requests through different servers, all while keeping your private data strictly on your device. It even supports Siri shortcuts so you can toggle your VPN with your voice or automate connections based on location. The tool is free to use, collects zero data, and supports a wide range of modern protocols for reliable, fast connections without needing to buy a service.

🆔 @hackernewsgithubprojects