📦 sgl-project/sglang-torchtpu

SGLang TorchTPU: PyTorch LLM Serving on TPU

SGLang TorchTPU is the PyTorch-native plugin that lets you run large language models on Google TPUs without leaving the PyTorch ecosystem. This project bridges the gap between familiar PyTorch development workflows and high-performance TPU hardware, specifically targeting the latest Ironwood chips. By integrating directly with the SGLang serving engine, it provides an OpenAI-compatible API that handles complex tasks like prefix caching and speculative decoding under the hood. This means you can use standard PyTorch tools for debugging and compilation while deploying to specialized silicon. It is a powerful way to serve models at scale without abandoning your existing Python setup.

🆔 @hackernewsgithubprojects

📦 tetsuo-ai/agenc-core

Agenc Core: The Cheap Terminal Coder

Agenc Core hands you a coding agent that lives right inside your terminal, ready to write, fix, and test your code. The real kicker? It scored seventy-six percent on a major benchmark for just thirty-four dollars, while every other tool on the leaderboard cost at least one hundred thirty-four. It works with any model you like, including local ones on your own machine. So if you want a smart, affordable coding partner that actually runs in your command line, give it a try and see the difference.

🆔 @hackernewsgithubprojects

📦 lnsgroup/ms-human-700

MS-Human-700: 700 Muscles in Code

MS-Human-700 is a digital human model that includes seven hundred individual muscle actuators, giving researchers an incredibly detailed way to simulate how the body actually moves. It is built on the MuJoCo physics engine and comes in different versions, like a full-body version for walking or a focused version for hand tasks. This matters because simple robot models often look wrong, but this one captures the complex way human muscles pull on bones to make motion feel natural. It is used to train AI agents that can walk, run, and dance with high precision.

🆔 @hackernewsgithubprojects

📦 lacs-project/sysknife

The AI That Can't Break Your Linux

SysKnife lets an AI manage your Linux system without ever handing it the keys to a raw shell. Instead of generating risky strings that might accidentally break things, the AI selects from a fixed list of pre-defined, typed actions. The software then builds the actual command itself, ensuring the AI can never slip in something dangerous. You review the plan, explicitly approve each step in your terminal, and the system executes it with a permanent, cryptographic audit trail. This removes the fear of an AI agent making a catastrophic mistake. It turns unpredictable automation into a safe, verifiable process you can actually trust.

🆔 @hackernewsgithubprojects

📦 helpin-ai/helpin

How Helpin Connects Your Entire Team With AI

Helpin is an open source workspace that puts your support chat, project tasks, customer records, and meeting notes into one single shared view. The most surprising part is how the built in artificial intelligence agents actually do the work. They look at the full history of a customer conversation, then draft replies, plan next steps, and even help write code changes before a human approves them. This solves the messy problem of teams losing context when passing work between departments. Instead of repeating yourself, the next person or agent sees exactly what happened and what still needs doing.

🆔 @hackernewsgithubprojects

📦 juivyy/legoocc

LegoOcc: 3D Maps from One Photo

LegoOcc is the computer vision tool that turns a single photo into a full three-dimensional map of a room. Most systems need expensive laser sensors to understand space, but this one uses just a regular camera to build a detailed grid showing exactly where the walls, floor, and furniture are. The most surprising part is that you can ask it to find almost any object you can name, without needing to retrain the model. You simply tell it what you are looking for, and it identifies those items inside the virtual space. It solves the problem of making smart environments aware of their surroundings without bulky hardware.

🆔 @hackernewsgithubprojects

📦 openbmb/minicpm-ocr

MiniCPM OCR: The AI That Grades Its Own Homework

MiniCPM OCR is an open-source tool that not only reads your documents but actually checks its own work. While most scanning software just guesses what it sees, this system includes a built-in quality judge that flags any text, table, or formula it isn’t confident about. It handles messy, real-world inputs like curved pages and blurry photos, turning them into clean, usable text. You don't have to trust every single word blindly, because the model tells you exactly where you need a human to double-check. It is a practical, self-correcting way to digitize paperwork without the usual manual headache.

🆔 @hackernewsgithubprojects

📦 dongnguyenvie/bashcut

BashCut: The Video Editor Your AI Can Drive

Your AI agent can now cut video. BashCut is a native macOS editor where Claude Code and Codex sit right next to your timeline, driving every edit through the same commands you use. This means your coding assistant doesn’t just write text; it literally manipulates footage, adjusts timing, and bakes in captions using the exact same undoable steps as a human editor. It turns your AI subscription into a real pair of hands for post-production. You get a professional-grade editing suite that is fully scriptable, meaning you can automate entire workflows or have your agent fix awkward cuts while you sleep. It is a clever bridge between code and creativity.

🆔 @hackernewsgithubprojects

📦 sebastianhaba/book-to-course-skill

Book to Course: The AI Skill That Turns Books Into Interactive Courses

Book to Course is the AI skill that turns any book into an interactive course you can actually study. It reads your document, whether it is a PDF, an ebook, or plain text, and builds a custom learning experience in your browser. The coolest part is that it does not just summarize the text; it creates working coding exercises that automatically check your answers. You write code in your own editor, click a button on the course page, and it runs the tests to tell you if you got it right. It even packages the whole thing as a mobile app for your phone.

🆔 @hackernewsgithubprojects

📦 ggml-org/llama-windows

Run AI Locally with Llama for Windows

Run large AI models directly on your computer without any internet connection. Llama for Windows brings the power of local language models to your system tray, letting you chat privately from anywhere. The coolest part is a hidden overlay that pops up over any application. Just press Alt and Space to instantly start a conversation with your loaded model, and it stays hidden in the background so you can pick up right where you left off. Everything runs locally, meaning your prompts never leave your machine. It is a simple, private way to use artificial intelligence on Windows.

🆔 @hackernewsgithubprojects

📦 covai-labs/ace

AI Chat Exporter: Your AI Conversations, On Your Terms

Your AI chat history stays on your computer while you move it between different platforms in one click. AI Chat Exporter is a privacy-focused browser extension that solves a common annoyance: having your valuable conversations locked inside specific websites. It scans your active chat in ChatGPT, Claude, or Gemini, then converts that text into clean files like Markdown or PDF entirely on your device. No servers, no tracking, just your data. The standout feature is cross-model continuation, which takes your current conversation context and injects it directly into a different AI platform, letting you pick up exactly where you left off without manual copy-pasting.

🆔 @hackernewsgithubprojects

📦 akivet/mica

Mica: The AI that answers, not chats

Mica is the decision model that turns messy data into clear probabilities without writing a single word of text. Unlike standard chat bots that guess and type out paragraphs, Mica reads your specific rules and data once, then instantly gives you a confident percentage for every possible answer. This makes it perfect for sorting customer tickets or validating complex forms where speed and certainty matter more than conversation. It even tells you when it is unsure, so you never have to rely on a blind guess. If you need software that makes split-second calls on hard questions, Mica is the tool that finally brings real clarity to automated logic.

🆔 @hackernewsgithubprojects

📦 matt0975/vory

Vory: Your AI Remote

Turn your phone into a full remote control for your own private AI. Vory connects straight to the Hermes Agent running on your machine, so you can chat, approve actions, and manage your bots without touching a web browser. The coolest part? It streams every word in real time, right into a sleek, glassy interface that feels like it came straight from Apple. No cloud, no middlemen, just you and your agent. If you run your AI at home, this is how you finally get the mobile experience you deserve.

🆔 @hackernewsgithubprojects

📦 tencentcloud/octop-memory

octop-memory: Give your AI a brain

octop-memory is the memory system that gives your AI agent a long-term brain. Most chatbots forget everything the moment a session ends, but this tool stores your preferences, decisions, and context so your assistant actually remembers you. It works like a portable notebook for your AI, letting you save facts in one place and move them to different agents as needed. You can start with a simple local file or scale up to a server database, and it even handles the messy work of organizing memories for you. The best part is that your AI stops repeating the same questions and starts feeling like a real colleague.

🆔 @hackernewsgithubprojects