📦 behzade/farcaster

Farcaster: The Keyboard-First Workspace for Coding Agents

Farcaster is the native desktop app that puts your keyboard at the center of the workflow. Instead of juggling browser tabs or scattered terminal windows, it gives you a single, focused interface to run and coordinate multiple coding agents at once. The coolest part? You can drop directly into an embedded terminal or Neovim editor without losing your place. It’s built for speed, letting you manage complex tasks across different agent harnesses without ever touching a mouse. If you live in your code and want to stop context-switching, this is the tool that keeps your hands on the keys and your mind on the work.

🆔 @hackernewsgithubprojects

📦 jerryjliu/fly_ocr

A Fly Brain Reads Text

Recognizing printed characters using a frozen fruit fly brain is the core of Fly OCR. This project runs a fixed neural connectome of one hundred sixty-six thousand neurons on a standard Mac. It does not use a large language model or a massive cloud server. Instead, a tiny trained decoder reads the electrical spikes from the fly circuit to identify letters and numbers. The result is a local, transparent tool that shows exactly which pixels and neural activities lead to each prediction. It is not a claim that real flies can read, but it is a fascinating experiment in biological-inspired computing.

🆔 @hackernewsgithubprojects

📦 nandometzger/metal-gauss

Metal Gauss: 3D Reconstruction on Apple Silicon

Train 3D Gaussian Splatting directly on your Mac, without ever touching CUDA. Metal Gauss replaces the usual Nvidia hardware requirement with native Metal kernels that compile right on your Apple Silicon chip, running entirely through PyTorch. This means you can process photos into detailed three-dimensional scenes using only standard command line tools, skipping the heavy setup of Xcode. The project focuses on speed and quality, proving that you can achieve high-resolution reconstructions in minutes, not hours, by leveraging the built-in GPU. It is a clean, open-source way to experiment with neural rendering on your laptop.

🆔 @hackernewsgithubprojects

📦 jonashertner/opencaselaw

OpenCaseLaw: 150 Years of Swiss Court Decisions for AI

Connect any AI assistant to the entire history of Swiss justice. This project scrapes over a million court rulings dating back to eighteen seventy five, along with thousands of federal laws, and maps how every single ruling references the ones before it. The real breakthrough is the citation graph: a map of ten million links showing exactly how one decision builds on another. It’s released under a public license so anyone can run this data locally or query it through a live server. You get the full, unfiltered legal record, not just a summary.

🆔 @hackernewsgithubprojects

📦 rh1tech/frank-386

Frank 386: A Retro PC Inside a Tiny Board

Run Windows 95 on a microcontroller the size of a coin. Frank 386 is a tiny firmware that turns a Raspberry Pi Pico Two board into a full retro computer. It emulates an old Intel chip and classic hardware so you can boot real operating systems like DOS or Windows 95 directly from an SD card. The coolest part is that it generates video and sound in real time without a screen or speaker attached to the chip itself. You just plug in a keyboard and monitor.

📰 https://news.ycombinator.com/item?id=49693613

🆔 @hackernewsgithubprojects

📦 kunchenguid/backpass

Train Your AI Memory

Instead of manually rewriting your agent rules after every failure, Backpass treats your coding assistant’s memory like a neural network. It scans the logs your AI agents leave behind, identifies exactly where instructions failed, and uses that data to propose precise, evidence-backed edits to your configuration files. It is like gradient descent, but for your software rules, ensuring every change is backed by real quotes from actual sessions. This turns your agent’s mistakes into a learning signal rather than a manual cleanup chore. You do not just document the problem; you let the system analyze its own performance to fix it.

📰 https://news.ycombinator.com/item?id=49703919

🆔 @hackernewsgithubprojects

📦 aseiel/videohighlighter

VideoHighlighter: The AI that shows its work

VideoHighlighter is a local desktop tool that finds the best moments in your videos without ever uploading a single frame to the cloud. The wild part? It doesn’t just guess what’s cool; it shows you the math. For every clip it picks, it generates a report detailing exactly why that moment scored high, distinguishing between what the camera actually saw and what a transcript claimed. It’s the open-source, offline alternative to paid cloud editors, running entirely on your machine. This means you get a highlight reel and a full breakdown of the logic behind every cut. No data leaks, no mystery algorithms, just your footage analyzed locally.

📰 https://news.ycombinator.com/item?id=49704396

🆔 @hackernewsgithubprojects

📦 nickelsec/bough

bough: See what your AI coding agent actually built

bough turns your AI coding session logs into a visual map of your actual work. It reads your local history to draw a diagram where squares represent days, smaller boxes show specific tasks, and circles mark every prompt you typed. The most surprising part is how it highlights the struggle; failed attempts appear in rust color, making it obvious where you got stuck versus what succeeded. It even checks your local git history to confirm which tasks resulted in real commits. Because it runs as a single binary on your machine and never sends data to a server, it keeps your private code completely private.

🆔 @hackernewsgithubprojects

📦 mpiton/tauri-pilot

Give AI Eyes for Your Desktop Apps

Tauri Pilot gives artificial intelligence the ability to actually see and touch your desktop applications. Since Tauri apps run inside the system's own web browser component rather than a standard one, tools like Playwright simply cannot reach them, leaving developers without a way to let AI agents test their software. This project solves that gap by embedding a lightweight plugin directly into your app, which opens a secure channel for an external command line tool to inspect the screen. It lets an AI read the layout, click buttons, type into fields, and check for errors in real time, just like a human user would.

🆔 @hackernewsgithubprojects

📦 purshow/purshow_notes

Purshow Notes: The Ultimate AI Architecture Guide

Purshow Notes is the bilingual research repository that finally demystifies the complex attention architectures powering modern AI models. It solves the problem of fragmented technical documentation by providing meticulously organized notes on mechanisms like Multi-Head Latent Attention and Mixture of Experts. The truly fascinating aspect is how it breaks down why removing the encoder is actually better for infrastructure, a nuance often lost in high-level overviews. Each note is compiled into polished, beautiful PDFs, making it effortless to study deep technical concepts without wading through raw code.

🆔 @hackernewsgithubprojects

📦 prokuon/cardputeradv_cap_tft_v2

Cap TFT V2: A Retro Screen for Your Cardputer

Cap TFT V2 is the display expansion that brings a chunky, retro-futuristic vibe to your Cardputer ADV. This hardware kit gives the tiny developer device a second, larger screen using a two point eight inch panel. It is a complete package with a 3D printed shell and a small power regulator, so you can plug it in and go. The most interesting part is the firmware, which runs a retro operating system styled after the handheld decks seen in the TV series Loki. It turns a simple tool into a piece of cyberpunk art.

🆔 @hackernewsgithubprojects

📦 sakanaai/pc-alm

PC-ALM Fixes a Core AI Math Problem

PC-ALM is the neural network method that finally closes the gap between local and global learning. Traditional deep learning relies on backpropagation, a process that demands the network see the entire picture before it can adjust any single part. This approach breaks down in scenarios where that global view is impossible to compute or transmit. PC-ALM solves this by letting each individual layer track and fix its own errors locally, using a specific mathematical trick to ensure these local corrections align perfectly with the overall goal. Instead of needing the whole system to agree at once, each part learns independently yet accurately.

🆔 @hackernewsgithubprojects

📦 made-in-basement/ember

Ember: The OS Built From Scratch

Ember lets you run classic games on a modern laptop without touching the operating system you actually use. It is a complete operating system built entirely in assembly language that boots directly from a USB stick. The coolest part is that it fakes a Sound Blaster card for old programs on a computer that has no such hardware, letting you play Doom with full music and sound effects. Since it runs on raw hardware, it skips the bloat of modern systems. You get a custom desktop with a paint program, word processor, and a file manager, all working right on the machine.

🆔 @hackernewsgithubprojects

📦 remorses/lintcn

Lintcn: Stop AI from Writing Messy TypeScript

Stop letting your AI coding tools clutter your TypeScript with messy type assertions and redundant checks. Lintcn gives you a powerful way to control this by letting you own custom lint rules that actually understand your code’s types. Unlike standard linters that only look at surface syntax, this tool uses the real TypeScript type checker to spot deep issues like unsafe unknowns or floating promises. You simply add the rules you want, customize them freely, and let the system catch the exact patterns your AI tends to overuse.

📰 https://news.ycombinator.com/item?id=47469510

🆔 @hackernewsgithubprojects

📦 cliffnodes/krea2-multi-character-lora-node-with-bounding-box-scene-and-outfit-edit

Krea2 Multi-Character Node

Krea2 Multi-Character Node is the ComfyUI node that finally stops character identities from blending together in one image. The biggest surprise is that you can draw a box around each person, assign a specific LoRA to that box, and the node mathematically locks that character to that exact space. This prevents the faces from merging or swapping traits, even if the characters stand close or touch. It also lets you place objects and backgrounds using the same box system, giving you total control over the scene layout without extra tools. It’s a clean, precise way to build complex, multi-person images with clear, separate identities.

🆔 @hackernewsgithubprojects

📦 oceanbase/powercontext

Powercontext: The Memory Layer for AI Agents

Powercontext is the memory layer that gives AI agents a full story, not just isolated facts. It solves the problem of context loss by capturing experiences, handoffs, and profiles into a unified system. Think of it as a long-term brain for your code assistants. When you pass a task to another agent, Powercontext ensures they remember the history and decisions, not just the current prompt. This makes your AI workflow feel continuous and aware. If you are building agents that need to learn and adapt over time, this is the infrastructure you need. It turns temporary chat sessions into persistent, evolving knowledge. Your agents stop forgetting. They start remembering.

📰 https://news.ycombinator.com/item?id=49707246

🆔 @hackernewsgithubprojects

📦 drandyhaas/kicadroutingtools

This AI Routes Your Circuit Boards

KiCad Routing Tools uses an AI assistant to automatically trace and route your entire circuit board, saving you from manually drawing every single wire. It analyzes your design, spots tricky connections like differential pairs, and even generates a step-by-step plan for how to route everything. You can watch the AI work in real time, adjust its plan, or just let it go and handle the busywork for you. It turns a tedious, error-prone task into a simple, automated workflow. This is a massive time-saver for any hobbyist or professional who wants to focus on designing rather than wiring.

📰 https://news.ycombinator.com/item?id=49707571

🆔 @hackernewsgithubprojects

📦 pku-yuangroup/openai4s

OpenAI4S: Run AI Science for Under Ten Yuan

A single API call costs nine yuan and ninety cents to run a full scientific agent that feels like a million-dollar lab in your pocket. OpenAI4S is an open-source platform that connects standard language models to real scientific workflows, allowing you to run complex tasks like protein design and genomic analysis without expensive proprietary subscriptions. It solves the problem of high entry costs for AI-assisted research by using a budget-friendly interface that mimics premium capabilities. This makes advanced computational science accessible to individual researchers who cannot afford enterprise pricing. The project proves that powerful scientific automation does not have to be locked behind a paywall.

🆔 @hackernewsgithubprojects

📦 amirivojdan/neyshekar

Neyshekar: The Persian Speech Dataset

Neyshekar is the open Persian speech dataset that finally gives researchers a clean, massive foundation for voice technology. It is not just a folder of random sounds, but a carefully curated collection of ninety-nine hours of audio recorded by native speakers through a web platform. This solves the critical problem of finding reliable data for speech recognition and text-to-speech systems in Persian. The dataset includes specific train, validation, and test splits, ensuring that experiments are reproducible and fair. By keeping the data open and free, it removes the biggest barrier to innovation for this language. If you work with speech tech, this is your new baseline for accuracy.

🆔 @hackernewsgithubprojects

📦 khansimon/field_converter

Field Converter: Soccer Player 3D Pose Estimation

Field Converter places 3D soccer players on the field with about ten centimeters of error. It takes a standard TV broadcast, finds the player’s position in 3D space, and anchors that skeleton to the actual pitch. It does not guess where the player is from scratch; it starts with a solid geometric estimate from the camera and the ground, then uses a short video clip to correct small mistakes. This means it works on real, messy broadcast footage, not just clean studio shots. It is a strong step toward tracking players for tactics, training, and video analysis from ordinary TV.

🆔 @hackernewsgithubprojects