📦 brayonpi/hexstellar

HexStellar: The AI Solver That Actually Proves Its Answers

Your AI chatbot can talk about the perfect delivery route, but HexStellar actually solves it. This tool hands complex optimization puzzles to a specialized engine that returns a mathematically verified answer, not just a guess. You get a clear receipt proving the result is optimal, which is huge when you are making real financial decisions. It is the difference between asking a friend for advice and having a mathematician sign your work. It is fast, honest, and built for developers who need certainty. Stop letting your AI hallucinate the solution, and start verifying it instead. This is how you get real results from your agents.

🆔 @hackernewsgithubprojects

📦 lisitskyaa/comfyui-dlss5-nr

ComfyUI Gets Neural Rendering

Nvidia's newest neural rendering engine now runs directly inside ComfyUI without launching separate programs or saving temporary files. This experimental project bridges a gap by letting users apply advanced AI-based image enhancement right within their existing workflow. It solves the problem of complex setup by handling the heavy lifting internally, which makes testing new visual styles much faster. You just plug in your own compatible runtime file and start rendering. This is useful for anyone wanting to push the limits of local image processing. The key takeaway is that you can now experiment with next-gen rendering tools without leaving your familiar interface.

🆔 @hackernewsgithubprojects

📦 git-agni/prod-farm-ios-core

Phone Farm iOS: Control Real iPhones From Your Mac

Phone Farm iOS is the self-hosted tool that turns a pile of physical iPhones into a programmable farm running directly on your Mac. It solves the annoying problem of automating real phones by giving you live video and remote tap control, backed by a database scheduler that keeps track of every task. The standout feature is the built-in TikTok plugin, which lets you schedule actions like scrolling or posting on a recurring basis. It is interesting because you get genuine device automation without renting expensive cloud services. If you want to test apps or run bots on hardware you already own, this is the place to start.

🆔 @hackernewsgithubprojects

📦 co-l/cache-hunter

Cache Hunter: The LLM Proxy That Catches Cache Invalidation

Cache Hunter is the LLM proxy that finally catches cache invalidation before it blows up your costs. It sits quietly between your app and the model, logging every request to a local database so you can see exactly when the cache breaks. The coolest part? It builds a hash grid to visualize context coherence, letting you spot broken conversation chains at a glance instead of guessing. If you are running local models or debugging latency spikes, this tool turns invisible behavior into clear data. It is the missing piece for anyone trying to understand why their inference is slower than it should be.

🆔 @hackernewsgithubprojects

📦 openlair/dr-claw

Dr. Claw: Your AI Research Lab

Dr. Claw is the AI-powered research environment that turns your computer into a full-scale digital laboratory. Instead of just chatting with an assistant, this tool gives you a team of specialized AI agents that handle the heavy lifting of academic work. It can scan thousands of papers, write your code, run your experiments, and even draft your grant proposals or rebuttal letters. The coolest part is that it understands the entire research lifecycle, so you spend less time on admin tasks and more time on actual discovery. It’s like having a tireless postdoc who never sleeps and never forgets a citation.

🆔 @hackernewsgithubprojects

📦 openclaw-easy/viralmint

ViralMint: Your Local Video Factory

ViralMint is an open-source pipeline that runs entirely on your own computer, meaning your scripts and footage never leave your machine. It handles the entire workflow from scouting trends across YouTube and TikTok to actually uploading your finished videos. The coolest part is the clip studio, which lets you drag handles on a real timeline to cut long videos into shorts. You can even chat with the AI through WhatsApp or Discord to control the whole process. It is a self-contained tool that gives you total control without any cloud dependency. If you are tired of juggling ten different apps, this is the one tool that does it all.

📰 https://news.ycombinator.com/item?id=48221229

🆔 @hackernewsgithubprojects

📦 kodustech/agent-readiness

Agent Readiness: Is Your Code Ready for AI?

Agent Readiness instantly scores your project to see if it can actually handle AI coding assistants. It runs thirty-nine automated checks right on your own machine to measure how well your codebase supports autonomous tools like Claude Code or Copilot. Instead of guessing, you get a clear maturity level from one to five, revealing exactly where you stand. The coolest part is the interactive dashboard that opens in your browser, giving you a visual radar chart of your strengths and weak spots across seven key pillars like testing, documentation, and security.

📰 https://news.ycombinator.com/item?id=47378675

🆔 @hackernewsgithubprojects

📦 tencent/megastyle

MegaStyle: AI Style Transfer That Actually Works

Tencent just dropped a tool called MegaStyle that lets you take any picture and repaint it to match a specific artistic style, without losing the actual content of the image. It’s not just a filter; it uses a massive dataset of eight million images to teach the model exactly what a consistent style looks like. The coolest part is that it plugs directly into ComfyUI, so you can drop it into your existing workflow and start generating in minutes. If you’ve ever wanted to turn your photos into digital paintings or match a specific aesthetic, this is the tool that makes it actually happen.

🆔 @hackernewsgithubprojects

📦 cogevol/cogevol-4b

Cogevol-4B: Offline Course Generator

Cogevol-4B is the on-device model that turns a one-line course brief into a complete interactive classroom. It takes a simple request like 'teach me pendulum motion' and instantly generates structured slides and a working HTML simulation with adjustable sliders. The coolest part is that it runs entirely offline on your laptop, meaning you can kill your Wi-Fi and still build a full, functional physics lab in minutes. No cloud servers, no API keys, just your computer doing the heavy lifting. It is a tiny model that punches way above its weight, proving you do not need a massive data center to build real learning tools.

🆔 @hackernewsgithubprojects

📦 instituto-ia-lncc/manaca-1b-base

Manacá: The Open Brazilian AI Brain

Manacá is the open language model that proves you can build serious Brazilian Portuguese intelligence without a giant data center. Developed by a team at Brazil’s National Computing Center with partners from Japan, this project took the entire recipe for training an AI from scratch and made it completely transparent. Instead of just handing you a black box, they share the exact data, the code, and the steps needed to recreate the whole thing. It’s a rare case of true scientific open source in the AI world, where anyone can verify how the model learned to speak.

🆔 @hackernewsgithubprojects

📦 joeynyc/hermes-hud

Hermes HUD: Watch Your AI Think

Hermes HUD is a terminal dashboard that lets you watch an AI agent’s internal state in real-time. It reads directly from your local agent data to display memory usage, acquired skills, and a log of every mistake it corrected. Think of it as a flight recorder for an AI, showing you exactly how it learned and grew since yesterday. If you run Hermes, this gives you a clear, human-readable window into what your assistant actually knows. It turns abstract AI memory into visible, trackable progress. No more guessing if your agent is retaining information; you see it live.

🆔 @hackernewsgithubprojects

📦 badlogic/pi-share-hf

pi-share-hf: Publishing Safe AI Data

Publishing real AI coding traces for research without leaking your secrets sounds impossible, but pi-share-hf makes it possible. This tool takes messy coding agent logs and runs them through a strict safety pipeline. It scrubs out exact passwords, uses a specialized scanner to catch anything it misses, and has an AI double-check the content for privacy issues. Only then does it upload the clean data to a public dataset. This is the single most interesting part: it turns private, messy work into safe, shareable training data automatically. You get a reliable way to contribute real-world examples to the open-source community without risking a data breach.

🆔 @hackernewsgithubprojects

📦 zju-real/papergym

PaperGym: The Low-Leakage Research Plan Benchmark

PaperGym turns academic papers into a clean, low-leakage training environment for research-plan generation. The key win: it builds 20,000 instances across five domains with only 3.7% criterion leakage—far below the 12% to 34% in existing benchmarks. Using a two-stage train-then-reward scheme, it produces fine-tuned models that beat larger systems on ResearchQA while avoiding answer leakage. If you want reliable, high-quality plans without memorization, PaperGym delivers exactly that.

🆔 @hackernewsgithubprojects

📦 icoretech/codex-pooler

Codex Pooler: Self-Hosted AI Gateway

Codex Pooler turns your own server into a private gateway that routes every AI request through a shared pool of accounts. Instead of juggling separate logins, you connect your tools to one central endpoint that manages the traffic automatically. It tracks usage, handles failures, and keeps your data in-house. The project solves the chaos of managing multiple subscriptions by giving you a single, organized control panel for your team's AI access. It is a practical tool for anyone who wants to run their own infrastructure without the overhead. Give it a look if you want to take charge of your own AI pipeline.

🆔 @hackernewsgithubprojects

📦 algorithmicresearchgroup/fnscribe

FnScribe: Your Voice, Never Uploaded

FnScribe is the private Mac dictation tool that keeps your voice entirely on your computer. It runs a local speech model so your audio never touches a server, which is a rare move in a market full of cloud services. You simply hold the function key, speak your thoughts, and release to drop clean text into any open app. The software even listens for words like period or new line to format your writing automatically. Because no data ever leaves your machine, there is no account to manage and no history to worry about.

📰 https://news.ycombinator.com/item?id=49475159

🆔 @hackernewsgithubprojects

📦 afari-research/wa-jepa

WA-JEPA: The Future-Seeing Car Brain

Predicting how a car should drive by mentally simulating the road ahead is what WA-JEPA does. Built by researchers at Afari Intelligent Drive, this system takes a standard video model and teaches it to forecast future scenes and vehicle paths simultaneously. Instead of guessing a single line for the car to follow, it generates a range of plausible future outcomes, making its decisions more resilient to the chaos of real traffic. This approach allows the model to plan by looking forward rather than just reacting to what is immediately in front of it.

🆔 @hackernewsgithubprojects

📦 lihaoyun6/comfyui-h3vae_trt

ComfyUI H3VAE TRT: Faster Image Generation

Make your image generation faster in ComfyUI by using a specialized version of the MiniMax H3 VAE. This tool converts the model into a format your graphics card can process much more quickly, cutting down the time it takes to generate images by up to a factor of 1.7. It’s a simple addition to your workflow that helps you get results sooner without sacrificing quality. If you use ComfyUI, this is an easy way to boost your speed. Try it out and see how much quicker your generations run. It’s a small change that makes a big difference for anyone creating images regularly.

🆔 @hackernewsgithubprojects

📦 naver-ai/omni-evaluator

Omni Evaluator: One Tool for Every AI Test

Omni Evaluator runs the same set of tests on text, images, and video so you can see exactly how an AI model handles each type of input. Instead of juggling separate tools for every format, this toolkit connects to your model and scores it on a single, unified scale. This makes it incredibly easy to compare different models fairly and spot where they might fail. It is perfect for developers who need a quick, reliable way to check their AI’s performance without setting up a complex system. Just plug in your model, run the tests, and get clear answers on its abilities.

🆔 @hackernewsgithubprojects

📦 microsoft/studentsim

StudentSim: The AI Student

StudentSim is the research tool that lets you test AI tutors against fake students. Instead of waiting for human kids to practice, it builds a digital twin for every single learner. It learns exactly how each student thinks, including their specific mistakes. When a tutor gives advice, the simulator tries the answer again. If the advice works, the score goes up. This helps researchers perfect their methods without bothering real people. It is a clever way to speed up the development of personalized education software. You get a realistic practice partner for your AI models right away.

🆔 @hackernewsgithubprojects

📦 ictnlp/hybridemo

HybridEmo: The TTS Tool for Complex Emotions

HybridEmo is the text-to-speech project that finally handles complex feelings instead of just flat reading. Most speech tools give you one single emotion per sentence, but life is messier than that. This tool lets you tell the computer exactly how a voice should change or mix emotions in real-time. You can ask it to start happy and turn fearful, or keep sadness and joy blended together the whole time. It is built on top of CosyVoice and uses a clever training method to learn these nuanced shifts. You just provide the text, a short description of the mood, and a voice sample, and it creates audio that actually sounds human.

🆔 @hackernewsgithubprojects

📦 our-ark/enoch

Enoch

The personal AI framework that lets an agent rewrite its own source code under human supervision. Enoch is a persistent software agent that separates her lasting identity and memory from the temporary model running her. While other systems simply remember your notes, Enoch identifies limitations in her own behavior and proposes tested code changes to fix them. She builds these improvements in isolated branches for your review, meaning the software itself evolves through shared experience. It is a living lineage of agent bodies that grow alongside their owners. This is not just a chatbot, it is a self-improving tool you can fork, run, and govern yourself.

🆔 @hackernewsgithubprojects