📦 lok-i/orcs

ORCS: Train Humanoid Robots to Move Like Humans

ORCS lets researchers teach humanoid robots to perform complex physical tasks like dodging a ball or walking over rough terrain without falling. Instead of building a brain from scratch, it takes a pre-trained robot controller and fine-tunes it with extra sensory information, such as scanning the ground or tracking an object. This approach is surprisingly effective because it allows the robot to adapt its movements to real-world conditions while keeping the core balance stable. The toolkit handles everything from data preparation to training, making it much easier to experiment with advanced robot behaviors.

🆔 @hackernewsgithubprojects

📦 itsssssjack/slopmonster

SlopMonster: The Linter That Catches AI Slop

SlopMonster is the linter that finally catches AI slop before it ships. You know that distinct, robotic flavor in machine-written copy, the kind that makes readers roll their eyes? This project treats that flavor like a syntax error. It scores your text out of five and fails your build if the score is below perfect. It doesn't just flag words. It hunts for the specific sentence shapes and rhythm tricks that models love, like the rule of three or excessive dashes. Here is the coolest part. It forces you to clean the copy with a rival AI model. Why? Because a model can never hear its own accent.

🆔 @hackernewsgithubprojects

📦 nvidia-bionemo/bionemo-inference-runtime

BioNeMo Inference Runtime: The Protein Folding Speed Demon

BioNeMo Inference Runtime is the GPU pipeline that turns protein folding models into lightning-fast structure predictions without the usual compilation headache. Instead of wrestling with complex engine builds, you simply drop in standard neural network modules and let a five-stage system handle the heavy lifting. It takes raw sequence data and spits out 3D structures with confidence scores, all while running significantly faster and using less memory than standard PyTorch setups. This is perfect for anyone who wants to run AlphaFold-style models on modern hardware without turning their computer into a server farm. If you are working on protein design, this is the easy button that actually works.

📰 https://news.ycombinator.com/item?id=49672745

🆔 @hackernewsgithubprojects

📦 xqy2006/modeltrace

ModelTrace: Know Your AI Model

ModelTrace is the local tool that finally verifies which AI model is actually answering you. It solves a major trust gap by actively fingerprinting the backend, so you know if a cheaper, different model is silently swapping in behind the scenes. The project runs entirely on your machine using a browser-based dashboard where you can feed challenges to the AI and compare its unique linguistic traits against a stored fingerprint bank. This is brilliant because it gives developers a real, active way to audit their infrastructure without sending sensitive data to a third party.

🆔 @hackernewsgithubprojects

📦 tripogrowthlab/awesome-astra-prompts

Awesome Astra Prompts: 3D Games from Text

Awesome Astra Prompts is the curated collection that turns text into playable three-dimensional worlds. It gathers over two hundred prompts for GPT-6 Astra, an AI model that generates interactive games and scenes directly from written instructions. You do not just see static images here; you find real, explorable environments like particle colliders and anatomical explorers, complete with source code links. The repository updates twice daily and translates every entry into fourteen languages, making it a rare, accessible resource for creators. Whether you want to see what the model can do or find inspiration for your own projects, this is the place to start.

🆔 @hackernewsgithubprojects

📦 edonadei/caliper

Caliper: The Agent Skill Stress Test

Caliper doesn't just run your AI agent skills; it scientifically proves they actually work. You can test your specific skill by running it ten times with the skill active, and then ten more times with it completely removed. Caliper compares the two results to show you exactly how much smarter the agent became and how many tokens you saved. It is the only tool that isolates a single variable to tell you if your prompt engineering is real or just luck. If you build agent skills, you need this proof. Stop guessing if your code works. Measure it. Caliper gives you the data. Trust your agent because you measured it.

📰 https://news.ycombinator.com/item?id=48709606

🆔 @hackernewsgithubprojects

📦 skforecast/skforecast-ai

skforecast-ai: AI Forecasting Made Simple

skforecast-ai returns the exact, runnable code that produced your forecast. This AI assistant handles the complex work of profiling data and choosing models so you get a result in one call. The real surprise is the transparency: you don't just get a prediction, you get the script that created it. This means you can inspect, version, or run the code yourself later. It turns a black box into a clear, auditable process. You get the forecast and the proof of how it got there. That’s the takeaway: a forecast you can actually verify.

📰 https://news.ycombinator.com/item?id=49523988

🆔 @hackernewsgithubprojects

📦 lynnreal-ai/lynnreal-omni

LynnReal-Omni: One Model for All Video Tasks

LynnReal-Omni uses a single thirty-two billion parameter model to handle everything from making videos out of text to fixing blurry old footage, all in just a few seconds. Instead of juggling five different tools, this one framework takes a simple sentence, a sketch, or even a hand gesture, and turns it into a moving picture. It works so fast that it can start streaming video in real time, which makes it perfect for people who want to experiment without waiting around. The result is a flexible toolkit that lets creators mix and match inputs, like a photo and a pose clip, to get exactly the video they imagine.

🆔 @hackernewsgithubprojects

📦 tkasperczyk/thunderbird-mcp

Thunderbird MCP: Give AI Control Over Your Inbox

Thunderbird MCP lets you hand your AI assistant the keys to your entire digital life. This open-source extension plugs directly into the desktop mail client, granting large language models access to forty distinct tools for searching, drafting, and organizing mail, contacts, and calendars. The real kicker? It enforces a strict safety layer that forces you to manually review and approve any outgoing message or new calendar event before it actually goes out. No more guessing if your bot messed up a reply; you stay in the loop for every single send.

🆔 @hackernewsgithubprojects

📦 go-gui-org/go-gui

Build Native Apps in Pure Go

Go-Gui lets developers build modern, cross-platform desktop applications using only Go code, completely eliminating the need for JavaScript, HTML, or CSS. Instead of relying on a browser engine, this framework renders your user interface directly to the native graphics card, delivering fast, smooth performance on Windows, Linux, and macOS. It handles everything from simple buttons to complex data tables with GPU acceleration, ensuring your application looks and feels like a true native program. This approach solves the problem of heavy, slow web-based wrappers by keeping your entire application logic and state in one simple language.

📰 https://news.ycombinator.com/item?id=49223756

🆔 @hackernewsgithubprojects

📦 onmyway133/claude-code-tips-tricks

Claude Code Tips and Tricks

Claude Code Tips and Tricks is the collection that finally stops you from guessing how to use the tool. Instead of scattered forum posts, this repository curates specific, working workflows for the terminal-based coding agent, breaking down complex features into simple, actionable steps. The standout value here is practical context management, teaching you exactly how to keep your AI focused so it doesn't lose track of your project goals. You learn how to save and resume sessions like code branches, control the effort level for different tasks, and even track your token spending without leaving your machine.

🆔 @hackernewsgithubprojects

📦 mpociot/claude-siri-ai

Claude for Siri

Your Mac can now answer to Claude directly from Spotlight, bypassing the need for a separate browser tab or app switch. The repository named Claude for Siri turns your signed in account into a native provider, letting you ask the model right inside the system search bar. It works by running a small local bridge that connects your existing login to the operating system's new delegation framework. This setup is experimental and requires disabling security protections on a development machine, but it proves that third party models can plug into Apple's interface.

🆔 @hackernewsgithubprojects

📦 robinvdvleuten/xprsn

xprsn: Safe expressions without eval

Evaluate rules written by your users without ever executing raw JavaScript code. The xprsn library turns plain text conditions into safe functions that work perfectly under strict browser security policies. It parses expressions into simple closures, so it completely avoids the dangerous shortcuts that usually break secure apps. This makes it ideal for building form filters or spreadsheet formulas where anyone needs to write logic. The result is a tiny, fast tool that keeps your app secure while letting non-programmers create complex checks. If you need user-defined logic, this is the safe way to do it.

🆔 @hackernewsgithubprojects

📦 bilikaz/qwen38-flash-next-cluster-recipe

Two Sparks, One Model: The Qwen Recipe

Spin up a high-speed AI server on just two DGX Spark boxes with the Qwen Three Eight Flash Next cluster recipe. This kit automatically wires your hardware together, downloads the model, and serves it over a high-speed connection without you touching a single config file. The coolest part? It squeezes peak speeds of over one hundred tokens per second from hardware that usually chokes on that load, all while keeping memory tight. If you want a local AI setup that actually feels fast, this is the blueprint.

🆔 @hackernewsgithubprojects

📦 moudrkat/brainscope

Watch Your AI Think With Brainscope

Brainscope turns a standard chat server into a live X-ray for any Hugging Face model, letting you watch the machine think in real time. Instead of just getting text back, you see exactly which words the model is planning to say before it even types them. It peeks deep into the neural layers to reveal hidden intentions and let you nudge the output live using simple sliders. This turns a black box into a transparent lab where you can see the gears turning. If you build with language models, this is how you finally stop guessing and start seeing exactly why your AI says what it says.

🆔 @hackernewsgithubprojects

📦 huskydoge/awesome-loop-models

Awesome Loop Models: The Loop Research Tracker

Awesome Loop Models is the research tracker that finally organizes the scattered world of iterative neural networks. It cuts through the noise by strictly curating papers and deep technical blogs on architectures that reuse a single block to process data repeatedly, rather than just stacking layers. The most fascinating part is how it maps these loop mechanisms to track how models refine their thinking in real time, distinguishing true iterative logic from standard prediction. It gives you a clear map of the field without the fluff.

🆔 @hackernewsgithubprojects

📦 cookiy-ai/sell-sessions-skill

Sell Your AI Chat History?

Turn your old coding chats into actual cash. Sell Sessions Skill lets you license the specific agent sessions you already ran with tools like Claude Code or Codex. It runs entirely on your computer, so nothing leaves your machine until you personally approve it. The tool automatically scrubs out secrets and private info right on your device, then estimates the value of each session. AI labs are actively buying this history because it shows how real engineers correct and guide models to ship working code. You keep full control over what gets shared, and you get paid every time a buyer licenses your work.

🆔 @hackernewsgithubprojects

📦 diqierjia/stratagate-agentmemory

StrataGate: AI Memory That Checks Its Sources

StrataGate stops A I agents from treating every past conversation as a simple, unquestioned fact. Instead of just summarizing history, it builds small cards for specific events that keep the exact time and the original source message. The most surprising part is a built in check that asks the A I if it actually has enough proof to answer a question. If the memory is vague or missing a date, the system forces the agent to dig deeper into the raw chat logs before speaking. This prevents the A I from confidently guessing based on incomplete information.

🆔 @hackernewsgithubprojects

📦 aidenai-io/aiden-firmware

Aiden Firmware: The AI Hardware That Controls Your Phone

Aiden Firmware turns a small Linux computer into a voice-controlled operator that physically drives your mobile phone. It solves the gap between cloud-based chatbots and local hardware by running a full agent loop directly on the device. The standout feature is its human handoff, which detects when the AI is stuck and passes control back to you seamlessly. This repo contains the entire C++ and Go source code needed to build this autonomous hardware. If you want to see how AI actually interacts with the physical world, this is your blueprint. Grab the code, flash the board, and let the machine do the tapping.

🆔 @hackernewsgithubprojects

📦 aetherlabsai/rsiagent

R S I Agent: How AI Learns Without Training

R S I Agent lets a computer figure out how to use new software without ever updating its brain. Instead of memorizing instructions through heavy training, three separate digital workers explore a fresh environment, test their actions, and save useful lessons into a shared notebook. This notebook stays fixed while the agent tackles real tasks, allowing it to improve performance by simply recalling past experiences. It works by first exploring broadly to gather general knowledge, then digging deeper into tricky details to solidify that memory. The result is an agent that gets better at using software just by thinking and trying, not by being retrained.

🆔 @hackernewsgithubprojects

📦 rsasaki0109/visloc-rs

The Rust Library for Drone Navigation Without GPS

This Rust library figures out exactly where a robot is by looking around, even when GPS signals are completely gone. It takes ordinary camera photos and motion sensor data to build a detailed three-dimensional map of the surroundings in real time, a process called simultaneous localization and mapping. Most tools for this job are slow or require expensive hardware, but this one runs entirely on standard computer processors with no special graphics cards needed. It also processes thousands of images faster than the industry standard, making it a powerful tool for developers building autonomous drones and ground robots.

🆔 @hackernewsgithubprojects