📦 withlang-dev/open-jai

OpenJai: The Self-Hosting Language Compiler

This project is building a compiler that eventually writes its own source code in the language it compiles. OpenJai is a new systems language designed for speed and developer control, but right now it is just a bootstrap tool written in Zig. It takes your code, breaks it down into steps, and hands it to LLVM to turn into a runnable program. The crazy part? The team is writing the actual compiler logic in the new language itself, right next to the old one. It is a race to see if the new compiler can build itself without help. If it works, you get a fully independent toolchain from scratch.

📰 https://news.ycombinator.com/item?id=48116458

🆔 @hackernewsgithubprojects

📦 cseti007/comfyui-crossviewwarp

CrossView Warp: The 3D Camera Orbit for AI Video

Take a standard video and physically move the camera around the subject without re-shooting. CrossView Warp is a ComfyUI node that acts as a visual depth-warp engine, letting you drag a marker on a 3D sphere to choose a new angle. It reprojects your clip into that new view, filling in the gaps where the original lens never looked. The coolest part is the live preview: you orbit the scene in real time, seeing exactly what the AI will generate before you even start. It turns a flat recording into a dynamic, multi-angle shot, solving the stiff, locked-camera problem that plagues current AI video tools.

🆔 @hackernewsgithubprojects

📦 kingmaniya/comfyui-llm-text-processor

LLM Text Processor for ComfyUI

Feed images straight into local language models inside ComfyUI without leaving your node graph. This plugin, called the LLM Text Processor, lets you drop in GGUF models like Qwen or Gemma to automatically write prompts, translate text, or generate image captions. The real kicker is that it handles multimodal tasks right on your own hardware. You can plug in a batch of photos and let the AI look at them, all while running completely offline. It even separates the model’s internal reasoning from its final answer, so you see exactly how it thought. If you are building automated image workflows, this gives you a powerful, private brain for every step.

🆔 @hackernewsgithubprojects

📦 yifanzhang-pro/klpo

KLPO: The No-Frills AI Trainer

KLPO is the reinforcement learning method that cuts the noise out of AI training. Most algorithms need a whole group of answers to figure out what is right, but this one gets by with just a single attempt. It solves a real problem where you need to teach an AI to handle tasks, but you do not want to waste time calculating the value of every step. It works by using simple math to estimate the score without needing a separate judge model. This is fascinating because it makes complex training much more direct and easier to run on your own computer.

🆔 @hackernewsgithubprojects

📦 pacesalab/bindcraft2

BC2: The Tool That Designs Proteins From Scratch

BC2 designs entire protein binders from scratch, creating new molecules that lock onto specific biological targets without needing a pre-existing template. It solves the problem of finding the right molecular key for a lock by generating and testing thousands of candidate designs in one go. You simply describe your target and pick a binder style, like a small antibody or a cyclic peptide, and it handles the complex math of folding and binding. This is useful because it turns a months-long trial-and-error process into a structured workflow that spits out ranked, testable sequences. Instead of guessing what works, you get a list of predicted complexes ready for the lab.

🆔 @hackernewsgithubprojects

📦 saganaki22/comfyui-hyperflow

Generate Video and Sound in Eight Steps

A single plugin lets you generate video and synchronized audio using only eight distinct steps, a massive speedup for standard video models. This tool, called ComfyUI Hyperflow, adapts a specific training method to work directly inside the ComfyUI environment without modifying its core files. It achieves this by adding a unique timing condition that guides how the model integrates data at every single stage. This approach preserves the distinct relationship between time and the generation process, ensuring the output matches the original high-quality results. The result is a faster, integrated workflow for creators who want to produce both moving pictures and matching sound without needing complex external setups.

🆔 @hackernewsgithubprojects

📦 ghuntley/preflight

Preflight: Stop Leaking Secrets to AI

Preflight is the security guard that stops your API keys from accidentally leaking to AI models. It sits quietly between your applications and the cloud, scanning every request and attached file for sensitive data like passwords before the model ever sees them. If it finds a secret, it blocks the transmission or redacts the sensitive text on the spot. This simple check prevents catastrophic data breaches caused by copy-pasting code or uploading documents that contain hidden credentials, making it an essential safety net for anyone building with large language models.

📰 https://news.ycombinator.com/item?id=49776416

🆔 @hackernewsgithubprojects

📦 teorth/optimizationproblems

Optimization Problems: Tracking the World's Hardest Math Numbers

Optimization Problems is the collaborative math tracker that finally pinpoints the exact edges of our knowledge in pure mathematics. The repository catalogues best-known upper and lower bounds for dozens of notoriously difficult constants, turning scattered academic papers into a single, living ledger. This is useful because it lets researchers and curious minds see precisely where the frontier lies, from the moving sofa constant to the matrix multiplication exponent, and it openly invites the crowd to submit improvements. When a new proof tightens a boundary, it updates the record, creating a transparent map of progress.

🆔 @hackernewsgithubprojects

📦 acatovic/ai-game-studio

Wombo: Create Game Characters with Text

Turn plain text into playable game characters with Wombo, a local web tool that handles the heavy lifting of asset creation. Instead of drawing every frame by hand, you describe a hero in words, and the software generates a consistent character from multiple angles. It then animates that character using AI video models, automatically extracting clean, transparent frames to build a professional spritesheet ready for your game engine. This eliminates the tedious, repetitive work of manual pixel art, letting developers focus on game logic rather than tedious asset production. If you are building a two-dimensional game, this tool transforms a simple text prompt into finished, usable animation assets in minutes.

📰 https://news.ycombinator.com/item?id=48260339

🆔 @hackernewsgithubprojects

📦 ethanplusai/harvey

Harvey: The AI Sales Agent That Uses Your Subscription

Harvey runs an autonomous sales pipeline that finds local businesses, writes personalized cold emails, and handles every reply to book meetings for you. The cool part? It skips expensive API bills entirely by running on your existing Claude subscription, so you get full AI power without the per-token cost. It scans websites for specific gaps to make emails feel personal, not spammy. Plus, it acts as a copilot, waiting for your approval before sending anything. It is a clever way to automate your outreach while keeping total control. Check out the repo to see how it works.

🆔 @hackernewsgithubprojects

📦 mizorewww/laya-coreml

Laya Core ML: 5 Millisecond AI Decisions on Apple Silicon

Run a real AI agent locally on your Mac that answers questions in about five milliseconds without generating any text. This project ports open-weight models to Apple’s Neural Engine, cutting energy use by nearly three times compared to standard setups. The coolest part is a live Snake game in your terminal where the AI plays itself, showing its decision probabilities in real time. You get reproducible speed benchmarks and a fully offline experience, proving that tiny, fast models can actually do useful work right on your chip. No cloud, no lag, just lightning-fast local intelligence.

🆔 @hackernewsgithubprojects

📦 chanon/comfyui-obvpm

The Tidy Workflow Hack

ComfyUI Obsidian is the node pack that finally tames your spaghetti workflow by bundling ten wires into one. If you have ever stared at a knot of cables trying to figure out which value goes where, you know the pain. This tool lets you pack settings into a single wire, then unpack them exactly where you need them, keeping your canvas clean and logical. It is the difference between a messy mess and a system you can actually read. Your graph will be so much cleaner you might never go back.

🆔 @hackernewsgithubprojects

📦 whitecircle/halo

Halo: The LLM Training Framework

Halo is the open-source framework that makes frontier AI training accessible to everyone. Built by White Circle, it lets you train massive language and multimodal models directly on standard Hugging Face checkpoints without rewriting your entire pipeline. The coolest part? It squeezes up to 2.8 times more training speed out of your hardware than the standard tools, while keeping your models fully compatible with existing ecosystems. If you want to fine-tune or build next-generation AI without getting tangled in complex infrastructure, this is your ticket. Grab the code and start pushing boundaries today.

🆔 @hackernewsgithubprojects

📦 patrickmgarrity/anthropic-credited-cves

The Anthropic Cve Tracker

Keep a close eye on the Anthropic Cve Tracker, a living list that documents every security flaw publicly credited to Anthropic and its Project Glasswing research team. It isn't just a random collection; it is a massive, organized table that maps out the real-world impact of this specific AI-driven vulnerability hunt. The repository currently tracks two hundred and twenty-five distinct issues, ranging from minor bugs to critical exploits found in the wild. It highlights exactly which software components are being cracked open and who is responsible for the fixes.

🆔 @hackernewsgithubprojects

📦 jerryjliu/docjev

DocJev: The Fast Document Splitter

DocJev identifies and separates distinct documents hidden inside large multi-file PDFs. It uses natural language rules to classify content and pinpoint exact page boundaries, even when adjacent sections look nearly identical. For example, it correctly splits a fifteen-page government packet containing two separate Treasury auction results that share the same category, while keeping a dense ten-page economic report intact. The system processes real publications from the IRS and SEC, delivering precise segmentation results in under half a second. It runs locally for text extraction and uses a hosted service for final decisions.

🆔 @hackernewsgithubprojects

📦 kharmanskyi/open-steps

Open Steps

A coding agent just finished fixing a critical bug, but instead of a clear answer, it handed back a wall of technical jargon and commit hashes. This is exactly the problem Open Steps solves. Created by a product builder who doesn't read code, this project teaches AI agents to communicate in plain, human language. It replaces confusing technical details with simple reports that clearly state what is finished, what is broken, and what needs your next action. It’s like having a translator who ensures you actually understand the status of your project without needing a computer science degree.

🆔 @hackernewsgithubprojects

📦 agenta-ai/awesome-ai-agent-platforms

Awesome AI Agent Platforms

Map the sprawling world of open-source AI tools with this single, curated directory. The project, called awesome AI agent platforms, solves the confusion of finding reliable software by listing every major open-source agent tool in one place. It groups these projects into five clear categories, from digital coworkers that handle your daily tasks to frameworks developers use to build their own. The most surprising part is how it tracks the license and hosting model for every single entry. This lets you instantly know if you can self-host a tool or if you need a cloud service. It removes the guesswork from choosing an AI teammate.

🆔 @hackernewsgithubprojects

📦 smixs/creative-director-skill

Creative Director Skill: 571 Legendary Campaigns

A single project packs five hundred and seventy one of the most famous advertising campaigns in history into a practical toolkit for generating new ideas. Known as Creative Director Skill, this repository functions like a digital brain for brainstorming, combining proven thinking frameworks with a rigorous scoring system based on actual industry awards. It solves the common problem of creative blocks by providing structured methods to analyze emotions, insights, and storytelling patterns. Instead of guessing what works, users can apply the exact logic used to create iconic ads for global brands. It turns abstract creativity into a repeatable process grounded in real world success.

🆔 @hackernewsgithubprojects

📦 gefsikatsinelou/metasearchmcp

MetaSearchMCP: The Search API Built for AI Agents

MetaSearchMCP is the search gateway that speaks fluent agent instead of human. It fires off one request to dozens of sources at once, from Google to Wikipedia to Stack Overflow, then shreds all those messy results into one clean, predictable list of data. This solves the headache of writing separate scrapers for every site and worrying about which API might fail today. It’s built for your A I tools to grab facts, code, or papers without breaking their flow. If you’re building agents that need real-world answers, this is the simple way to get them.

🆔 @hackernewsgithubprojects

📦 amap-ml/dreamx-rec

DreamX Rec: The Travel AI That Predicts Your Journey

DreamX Rec is the travel research suite that finally connects the dots of where to go, how to travel, and what to visit along the way. Instead of treating recommendations as isolated clicks, this project unifies a massive dataset of real-world trips into a single system that understands the context of your journey. It predicts your destination, preferred travel mode, and departure time by analyzing the flow of your past experiences. This is fascinating because it moves beyond simple lists to model the actual logic of a trip. By connecting these different parts of travel, it offers a clearer view of how personalized systems can understand complex, multi-step human decisions.

🆔 @hackernewsgithubprojects