📦 trentorch/trentorch

TrenTorch: Build Your Own PyTorch

TrenTorch is the coding curriculum that forces you to build your own machine learning framework instead of just importing one. Most people treat neural networks like a black box, but this project makes you write every single line of code on plain NumPy to recreate the actual machinery of PyTorch. It breaks the work into twenty progressive steps, guiding you from basic math all the way to building a real language model from scratch. The coolest part is that you are not just playing with toy problems. You are actually recreating historic milestones, like the original perceptron and the modern transformer architecture.

🆔 @hackernewsgithubprojects

📦 elliothux/open-compute

open-compute: The Self-Hosted Cloud

open-compute is the self-hosted platform that finally lets you run your Cloudflare Workers code on your own hardware without the usual infrastructure headache. It bundles the entire execution environment, database, and scheduling system into a single binary file, meaning you skip the need for Kubernetes or complex server clusters entirely. The standout feature is its perfect compatibility; your existing code runs unchanged with millisecond startup times, giving you full control over your data while maintaining the speed you expect from the cloud. If you want the power of edge computing without the vendor lock-in or the monthly bill, this is the tool to try.

🆔 @hackernewsgithubprojects

📦 tonyd2wild/qwen3.8-flash-next-nvfp4-dgx-spark

Running a 125B Model on One Box

A 125 billion parameter model now runs on a single, desktop-sized computer. The trick is that this repo keeps a massive lookup table on the hard drive, only pulling the specific pieces it needs for each word as it goes. It fits the rest of the brain into the machine’s memory without any compression tricks, serving text at speeds that rival much larger systems. This proves you can run the latest heavy language models on consumer hardware if you handle the memory smartly. It is a practical, measured guide for developers who want to test these big models without buying a data center.

🆔 @hackernewsgithubprojects

📦 vortrix5/sifty

Sifty: The Windows Cleaner That Actually Cares About Your Files

Sifty is the Windows maintenance tool that refuses to delete anything you cannot get back. While most cleaners send files to the void, this project routes every single removal through the Recycle Bin, meaning you can instantly undo your last cleanup if you change your mind. It also defaults to a dry-run mode, showing you exactly what it plans to touch before you give the go-ahead. This makes it the safest way to reclaim space from browser caches, old build files, and forgotten startup items. The interface is built for people who want control, not surprises, ensuring your system stays tidy without the risk of accidental data loss.

📰 https://news.ycombinator.com/item?id=48506948

🆔 @hackernewsgithubprojects

📦 amarnath3003/iot-claw

IoT-Claw: Your AI Smart Home

IoT-Claw is the local-first smart home platform that lets an AI agent actually fix your own devices. Most smart home setups break when a sensor goes offline or a network hiccups, but IoT-Claw has a built-in sandbox where the AI writes, tests, and debugs its own code to solve problems without you touching a wire. It connects to your cheap microcontrollers over Wi-Fi, so if your light stops working, the system figures out why, patches the script, and pushes the fix directly to the hardware. It’s basically a self-healing smart home that runs entirely on your own network, keeping your data private while it handles the tedious troubleshooting for you.

🆔 @hackernewsgithubprojects

📦 amarnath3003/dataset-creator-app

Dataset Creator App: Your All-In-One AI Lab

Dataset Creator App is the local workspace that finally puts dataset creation, model training, and testing under one roof. Most developers feel scattered across different tools for every step of the AI pipeline, but this project solves that by connecting the whole workflow in a single place. You feed it documents, and it automatically builds the training data, runs the fine-tuning on your own hardware, and then lets you chat with the result to see if it actually works. It even runs completely offline so your data stays private. If you want to build custom AI without touching the cloud, this is your all-in-one starter kit.

🆔 @hackernewsgithubprojects

📦 mcpfyy/mcpfy

mcpfy: One Line for Your AI Tools

Turn your AI models into powerful apps with a single line of code. mcpfy is a developer kit that takes the messy setup of connecting artificial intelligence to your data and makes it simple. Instead of writing hundreds of lines of boilerplate just to get started, you define your tools in plain TypeScript and run a single command to see them working. It handles the hard networking parts automatically, so you can focus on building the actual features your users need. If you have ever wanted to give ChatGPT or Claude hands that can fetch live weather or draw diagrams, this is the fastest way to do it.

🆔 @hackernewsgithubprojects

📦 liangmianya/dsh-synapse

Synapse: Your AI Chat Map

Synapse is the visual workspace that turns tangled AI chats into a navigable map. Instead of scrolling through endless text, you see every conversation, follow-up, and branch as a card on a single, draggable canvas. It plugs directly into DeepSeek Harness, syncing in real-time so your history stays perfect while you explore new ideas. You can even copy specific parts of a reply to spawn a new branch instantly. It’s like a whiteboard for your AI, helping you see how your thoughts connect without losing track. If you get lost in long chats, this is your shortcut back to clarity.

🆔 @hackernewsgithubprojects

📦 beamivalice/ponyexl3

Run 27B LLMs on Your Mac

Run a twenty-seven billion parameter language model directly on your Mac without a dedicated graphics card. PonyExl3 squeezes massive models into low bit formats that your Apple chip can process natively. The surprising part is that it does not just store these compressed weights; it decodes them on the fly inside the math engine. This means you get near full quality answers while using a fraction of the memory, fitting huge models into standard laptop hardware. It turns your everyday computer into a capable local inference station. If you want to test large language models locally without buying extra hardware, this is the tool to try.

🆔 @hackernewsgithubprojects

📦 hannay001/lockkeeper

Lockkeeper: The Firewall for Your AI Agents

Lockkeeper is the capability router that stops your AI coding agents from drowning in unnecessary tools. The problem is simple: modern agents install hundreds of skills and plugins, but most of that context is just noise that slows them down. Lockkeeper solves this by acting like a bouncer for your agent’s brain, picking only the eight or ten specific tools actually needed for your current task. The truly surprising part is its built-in firewall. It scans every new skill you install to catch hidden malicious instructions before they ever touch your code. It’s the security layer you didn’t know you needed for your AI toolbox.

🆔 @hackernewsgithubprojects

📦 amarnath3003/quickliquid

QuickLiquid: Real Glass for the Web

QuickLiquid calculates real optical refraction to bend the actual pixels behind your interface elements, creating genuine Apple-style liquid glass on the web. Instead of using a flat translucent layer that just looks fake, this library uses math to simulate how light physically bends through a curved surface. It handles everything from dynamic rim lighting to chromatic aberration, making buttons and cards feel like tangible objects made of solid material. Because it is heavily optimized with a smart caching system, you get this premium depth without sacrificing speed.

🆔 @hackernewsgithubprojects

📦 seeed-studio/kicad-mcp-server

KiCad MCP Server

Hand your AI assistant direct access to your KiCad hardware designs. This project connects your local circuit schematics and printed circuit boards to large language models through the Model Context Protocol. Instead of manually digging through files or guessing how parts connect, you can simply ask your chatbot to trace pin level connections or run design rule checks. It parses your files directly to find errors and validate your work. It is the perfect bridge between human creativity and machine precision. You draw the board, the AI does the heavy lifting.

🆔 @hackernewsgithubprojects

📦 vinzdg/codenotch

Codenotch: See Your AI Limits

Stop guessing if your AI is about to cut you off. Codenotch pins a tiny, sleek black bar to the edge of your Mac screen that instantly shows exactly how much of your usage limit you have burned for Claude, Cursor, and other coding tools. It works in the background without you needing to sign in again, just reading the data your apps already have. The coolest part is that it tells you if your AI is still working or if it’s stuck waiting for your input, so you never stare at a frozen screen wondering what went wrong.

🆔 @hackernewsgithubprojects

📦 cslsolow/repo0

Repo0: Code from Requirements

Repo0 is the design-driven generator that builds entire software projects from a single text file. It reads a plain-language description of what an app should do, then uses artificial intelligence to plan the structure and write all the code. The system checks its own work to ensure every part fits together logically before outputting a complete, usable program. You feed it the goal, and it returns the finished product. It turns a simple description into a working digital reality.

🆔 @hackernewsgithubprojects

📦 goposta/posta

Posta: Self-Hosted Email Delivery Platform

Point your app at a single HTTP endpoint and Posta takes over the entire email journey, handling delivery, templates, and tracking so you never touch SMTP settings again. This self-hosted platform replaces paid services like SendGrid by letting you own your data and infrastructure completely. It accepts plain JSON requests, processes them through a secure queue, and manages bounces, unsubscribes, and analytics automatically. The standout feature is inbound email parsing, where Posta receives messages, extracts attachments, and forwards clean data to your app via webhooks. It is a single binary that scales with your traffic, giving developers full control without the complexity of managing mail servers manually.

🆔 @hackernewsgithubprojects

📦 sebbbi/nographicsapi

NoGraphicsAPI: Pointers, Not Objects

NoGraphicsAPI lets your graphics shaders use real sixty-four-bit memory pointers to fetch data directly, completely bypassing the usual object bindings and descriptor sets that clutter modern graphics programming. This experimental Vulkan implementation strips away the traditional layer of abstraction, allowing your code to treat GPU memory like regular system memory. It is a fascinating look at how much of the standard graphics interface actually disappears when you give the hardware direct addresses. The project proves you can build a working renderer without the heavy object management overhead, offering a cleaner, more direct path to the silicon.

📰 https://news.ycombinator.com/item?id=49580105

🆔 @hackernewsgithubprojects

📦 shaheer-00/claude-adhd

claude-adhd

Your coding assistant can now quietly track all the half-finished tasks and ideas you completely forgot about between coding sessions. A local plugin called claude-adhd reads your past session transcripts to catch commitments you mentioned in passing but never actually finished. It operates like a gentle external memory, showing a quick digest of stale threads when you start up, and offering custom reminders when you need them. There is even a timed focus mode that redirects you if your conversation starts to drift, and an energy manager that serves up low-effort tasks when you tell the system you are fried.

🆔 @hackernewsgithubprojects

📦 vicoa-ai/vicoa

Vicoa

Vicoa is the open-source agent development environment that lets you orchestrate a fleet of coding AI assistants right from your phone. Think of it as a command center for tools like Claude Code, Codex, and Gemini, allowing you to run them in parallel on separate git worktrees so they never step on each other. You can spin up a task on your desktop, walk away to grab coffee, and monitor live git diffs or direct the session using native mobile apps with push notifications and dictation. It is a self-hostable, multi-device developer cockpit that turns coding agents into a true collaborative team.

🆔 @hackernewsgithubprojects

📦 mvrl/terradit

TerraDiT Satellite Image Synthesis

Imagine being able to generate highly detailed satellite imagery of any place on earth simply by feeding a computer a rough map layout. A research team has built terradit, an advanced artificial intelligence system that does exactly that. By blending language models with spatial transformers, it lets you construct realistic satellite views using simple markers like points, lines, or polygons representing rivers, roads, and buildings. It translates coordinate data and text prompts into matching, top-down geographic imagery with incredible structural accuracy. It is a massive step forward for virtual map design, urban planning simulation, and synthetic dataset creation.

🆔 @hackernewsgithubprojects

📦 muscriptor/muscriptor

MuScriptor: Audio to Sheet Music

MuScriptor turns any recording into readable sheet music, a capability that was previously restricted to expensive commercial software. Developed by Kyutai and Mirelo, this open-source model listens to a song and identifies every individual instrument, such as the guitar, bass, or drums. It then converts that audio into a MIDI file and engraves it into a full musical score. This solves the problem of manually transcribing complex music by hand, which is tedious and prone to error. You can run this tool on your own computer or use their hosted version. It provides a practical way for musicians and developers to extract notation from audio.

🆔 @hackernewsgithubprojects