📦 10xchengtu/harness-engineering

Stop Blaming Your AI for Bad Code

Stop blaming your AI for messy code. This project teaches your coding assistant to actually work. Think of it as building an operating system for your bot. It gives the agent clear rules, checks its own work, and remembers what it learned. Most bad output comes from missing context, not a dumb model. This tool fixes that. It turns a chatty helper into a reliable engineer that follows your rules. Give your AI a proper harness. You will get better results with less babysitting. Stop fighting your tools. Start giving them the structure they need to succeed. Your future self will thank you for this small setup.

📰 https://news.ycombinator.com/item?id=47616335

🆔 @hackernewsgithubprojects

📦 robert-dean/deadair

Deadair: The AI Director For Radio Stations

Deadair acts as an automated director that keeps your radio station on air, even when the schedule runs out. This tool manages the running order by shuffling tracks and filling gaps with generated content, ensuring a continuous broadcast. It solves the critical problem of dead air, which is the total loss of signal or audio. By handling the playout logic, it allows broadcasters to maintain a professional presence without constant manual intervention. The system integrates with your catalog to keep the stream flowing smoothly. You get a reliable, hands-off solution for managing your broadcast schedule. It is a practical way to automate the technical side of radio.

🆔 @hackernewsgithubprojects

📦 qwenlm/qwen-image-2.1

Qwen Image 2.1: The AI Model That Creates Transparent Images

Qwen Image 2.1 is the open-source image model that finally handles transparent backgrounds without a single extra step. It takes a simple text prompt or an existing photo and builds a high-quality picture where the background is actually gone, not just covered in white. This is huge because designers and developers can now drop these images directly onto websites or apps without needing to manually cut out shapes in expensive design software. The model uses a compact structure to run efficiently while still producing sharp details and realistic textures. It turns a blank idea into a ready-to-use asset in one go.

🆔 @hackernewsgithubprojects

📦 gylive/ccodex-sleep-state

ccodex-sleep-state: The Local Fix for AI Throttling

A single Go service now sits between your browser and AI models to stop them from getting dumber or cutting you off mid-thought. Codex Sleep State runs entirely on your machine, using a web panel to let you pick which server your requests go through without touching your global computer settings. It takes over your local AI configuration automatically, backing up your files first so you can always revert if things go wrong. The goal is simple: keep your connection stable and your answers consistent, no matter what the server is doing.

🆔 @hackernewsgithubprojects

📦 to3akarin/mac-computer-use

mac-computer-use

Hand over your Mac to an AI that can actually see your screen and move the mouse for you. This tool gives large language models a real set of hands to click buttons, type text, and open windows, effectively turning your computer into an obedient assistant. The coolest part is the proof: it built a three-dimensional model of a museum exhibit from a single photo, with no human touching the keyboard for twelve minutes. It works with almost any agent you already use, like Claude or Cursor, and verifies every step to make sure it did not just guess. Your next project might be one prompt away from finishing itself.

🆔 @hackernewsgithubprojects

📦 context-labs/whip

Why Whip Is The Fastest Coding Agent

Whip is the Go-based coding agent that finally makes open-source models feel as fast as the cloud giants. While most coding assistants rely on slow, heavy runtimes that bog down your terminal, Whip compiles into a single, lightweight binary with zero external dependencies. This means it boots instantly and streams results to your screen without the usual lag. It was built specifically to keep pace with the rapid growth of local, open-weight models, giving developers a native, high-speed way to run complex coding tasks directly from their terminal.

🆔 @hackernewsgithubprojects

📦 hanyang9/umr

How UMR moves human motion to robots

Take a human dance clip and make a humanoid robot perform the exact same moves without manual tweaking. UMR handles this by looking at the surface of the body rather than counting joints. It learns how points on a moving human surface match points on a robot surface. Then it uses that match to guide the robot motion. This lets the same system work across many motion sources and many robot bodies. That is the key idea. You get a repeatable path from human movement to robot movement. The result is a tool for anyone who wants to move human motion onto humanoid robots in a consistent way.

🆔 @hackernewsgithubprojects

📦 git-pkgs/brief

Brief: The Ultimate Project X-Ray Tool

Turn any codebase into a clear, instant report with Brief. It scans a project to find the language, package manager, test runner, and linter, then outputs a structured summary. This is built specifically for AI coding agents and new contributors who need to understand a project's conventions without reading every file. The most impressive part is the outline command, which packs your source tree into a single, compact document. It strips out function bodies and keeps only the signatures and imports, giving an LLM a perfect, token-friendly map of your code. It also tells you exactly which tools are missing from your setup.

🆔 @hackernewsgithubprojects

📦 wfzyx/von

Von: Instant AI Decisions

Classify complex support tickets in under fifteen milliseconds without waiting for a giant language model to type out its thoughts. Von is a local, open-source tool that makes instant, probabilistic decisions by analyzing your data in a single pass rather than generating text word by word. It handles tricky tasks like routing urgent issues or detecting secret leaks with high accuracy, all while running right on your own machine. The best part is that it answers multiple distinct questions about the same input simultaneously, saving you from slow back-and-forth calls. If you need rapid, reliable logic for your apps, this is the lightweight engine you have been looking for.

🆔 @hackernewsgithubprojects

📦 himovo/movo

MOVO

Movo is the self-hosted platform that turns a single AI agent into a managed enterprise tool. Instead of just chatting with a model, it gives your team a shared workspace where they can search internal documents, draft reports, and generate files. The surprising part is how it handles real files. It does not just read text. It understands charts and tables inside PDFs and spreadsheets, then uses that context to write new documents back. You can run this entire system on your own hardware with a single command. It keeps your data private and gives admins full control over who sees what.

🆔 @hackernewsgithubprojects

📦 video-rebirth/hyperflow

HyperFlow: 8-Step Video Magic

HyperFlow is the video accelerator that cuts generation time by nearly thirty percent without swapping out your base model. Usually, creating video and audio with MiniMax requires forty-nine slow passes to refine every frame, which takes forever. HyperFlow drops that to just eight passes using a lightweight adapter that teaches the original model to work smarter, not harder. It keeps the official quality but makes rendering three times faster on standard hardware. If you are generating video clips, this is the shortcut you need to stop waiting around for results.

📰 https://news.ycombinator.com/item?id=49756840

🆔 @hackernewsgithubprojects

📦 0xsero/deepseek-v4.1-flash-4x-rtx-pro-6000

DeepSeek V4.1 Flash: One PC, Four GPUs, Four Million Tokens

DeepSeek V4.1 Flash runs on a single Linux box with four RTX Pro 6000 Blackwell GPUs, handling over four million tokens without a second server. It keeps the heavy compute on the graphics cards while shunting the massive memory tables to a local drive, grabbing data only when needed. You get eight simultaneous chat sessions with vision support and no cloud dependency. It is experimental, but it proves you can run frontier-grade models locally on one machine. If you want to run big AI without renting a data center, this is your proof of concept. One machine, real speed, no cloud.

🆔 @hackernewsgithubprojects

📦 adonis-singh/re4

re4: The GameCube Classic, Fully Rebuilt

re4 is the decompilation that turns a GameCube classic back into readable code. Most games stay locked inside their binaries forever, but this project reverses the process, translating the compiled machine instructions back into the original human-written source. The result is byte-identical, meaning the recompiled version behaves exactly like the original, down to the last bit. It is a massive feat of reverse engineering, peeling back the layers of a complex title to reveal how the camera, combat, and audio actually work under the hood. You are not just looking at a patch; you are seeing the full skeleton of a commercial product laid bare.

📰 https://news.ycombinator.com/item?id=49778022

🆔 @hackernewsgithubprojects

📦 typesafe-ai/skills

The AI primitive

TypeSafe just handed your codebase a brain that actually understands context. This repository gives AI coding agents a new superpower called skills. Instead of asking a chatbot to guess what to do, you get tiny, focused units of intelligence. These units read your data and return structured answers with confidence scores. Your code stays in control, but the AI supplies the common sense. It handles routing, extraction, and verification without the messy back and forth of traditional prompts. This turns semantic understanding into a clean programming primitive. It is a clever way to build smarter apps that feel less like magic and more like engineering.

🆔 @hackernewsgithubprojects

📦 samuelcardillo/glm-5.3-flash-2x-rtx-pro-6000-blackwell

Run One Million Token AI Locally

The glm five point three flash recipe is the local setup that finally handles one million tokens without crashing. It solves the problem of massive context limits by teaching two high-end graphics cards to work together on a single huge document. Instead of just storing data, it uses a smart prediction trick to generate new words faster. You can feed it sixteen images and a novel-length text in one go. The repository provides exact instructions to build this powerful, private server. You get enterprise-grade memory handling without sending your data to the cloud. It is a reliable way to run top-tier artificial intelligence on your own desk.

🆔 @hackernewsgithubprojects

📦 rolecraft-sh/rolecraft

RoleCraft: The Security-First Skill Manager

RoleCraft scans every single AI skill you install for hidden threats like prompt injection and credential theft before it touches your system. Most developers install these skills blindly, trusting that a popular repository is safe, but RoleCraft changes that by running a static analysis check on every download. It gives each skill a security score from zero to one hundred. If the score is high, it installs. If it’s low, it blocks the install entirely. This makes it a crucial safety net for anyone using tools like Cursor or Claude Code, ensuring you never accidentally run malicious code. It’s a zero-dependency tool that just works.

🆔 @hackernewsgithubprojects

📦 cgallic/kai-cmo-harness

Kai CMO: Your AI Marketing Team

Kai CMO is a full marketing department that runs entirely inside your code editor. It plugs directly into Claude Code to handle every part of your growth, from writing blog posts and running ad campaigns to analyzing your website traffic. The coolest part is that it acts as a living expert, constantly checking its own work against the latest marketing trends and platform rules to make sure it isn't giving you outdated advice. If you are a developer who wants to handle your own marketing without hiring a whole agency, this is your answer.

🆔 @hackernewsgithubprojects

📦 zju-real/sdar

SDAR: The AI That Learns By Watching Itself

SDAR lets a language model teach itself how to act by learning from its own past mistakes, which is a big deal because it means the AI does not need a human to constantly correct it. Instead, it uses a trick called self-distillation, where the model reviews what it did before and uses that experience to make better choices next time. This is useful because it helps the AI get much better at doing things like searching for answers, shopping online, or navigating a virtual house, all by practicing on its own.

🆔 @hackernewsgithubprojects

📦 openma-ai/open-managed-agents

Open Managed Agents: Your Own AI Control Center

The single most surprising thing about this project is that it lets you run a full, multi-agent AI platform directly on your own server, completely free of vendor lock-in. Open Managed Agents is an open-source toolkit that gives you total control over how artificial intelligence executes tasks. Instead of relying on a single cloud provider, you choose the specific models, execution environments, and security boundaries for every job. It solves the problem of dependency by allowing you to self-host the entire infrastructure, from the agent logic to the secure sandboxes where code runs. This means you own the data and the workflow, not the platform.

🆔 @hackernewsgithubprojects

📦 wuyoscar/jev-skill

Jev Skill: Teach Your AI Agent to Decide

Your coding agent can now make complex decisions about which tool to use, how to triage a bug, or which file to inspect before it writes a single line of code. Jev Skill is a collection of specialized instructions that you can drop into your coding agent. Instead of guessing, the agent follows a specific workflow to analyze the task, check the context, and pick the right action. It turns a general purpose AI into a focused expert that knows exactly when to ask you for help and when to proceed on its own. You stop micromanaging every step and let the agent handle the routine judgments.

🆔 @hackernewsgithubprojects