📦 deep-jlu/awesome-graph-engineering

Graph Engineering: The Blueprint for AI Agents

Instead of treating AI agents as isolated chatbots, Awesome Graph Engineering maps out a structured path to system-level intelligence using graph structures. This curated repository, built by researchers at JLU, organizes papers, benchmarks, and tools into a clear hierarchy: from basic model training, to individual agent skills, up to complex system coordination. It solves the problem of scattered research by showing how explicit, dynamic graphs can organize tasks, manage memory, and let multiple agents coordinate effectively. Whether you are building software automation, scientific discovery systems, or healthcare assistants, this single resource is your map.

🆔 @hackernewsgithubprojects

📦 cyberful/cyberful

Cyberful: The Open-Source AI Red Team

Cyberful acts as an autonomous security lab that orchestrates over three hundred tools to hunt down vulnerabilities before real attackers do. It is not just a chatbot; it is a structured system where independent agents test exploit paths while a strict control layer ensures every finding is verified and evidence is preserved. This setup turns complex penetration testing into a repeatable, open-source workflow that anyone can run. It is a clever way to bring serious, multi-agent security research to your own infrastructure without hiring a full red team. Keep this in mind if you want to audit your own systems with serious, verifiable depth.

🆔 @hackernewsgithubprojects

📦 pytraveler/minimax-h3-prompt-rewriter-comfyui

MiniMax H3 Prompt Rewriter for ComfyUI

MiniMax H3 Prompt Rewriter for ComfyUI is the local workflow that turns a simple sentence into a professional video script. It takes a short prompt in any language and expands it into a structured English description with precise camera shots and synchronized audio details. This solves the problem of writing complex prompts manually, ensuring your AI video generator gets exactly the timing and sound cues it needs. The standout feature is its ability to run entirely on your own hardware, offering a lightweight version that works on modest graphics cards. You get production-ready descriptions without sending your creative ideas to external servers.

🆔 @hackernewsgithubprojects

📦 aiimaginglab/pcflow

PCFlow: Fix Blurry Photos in Just Three Steps

Clean up blurry, noisy, or damaged photos in just three steps. PCFlow is a new method that makes image restoration fast and visually stunning. It solves the annoying tradeoff where fixing pixel errors usually makes images look smooth and artificial. Instead, it uses a clever flow matching technique to jump straight to a realistic result. The coolest part is how it uses a conflict-free gradient strategy. This prevents the model from getting stuck when trying to balance sharpness and accuracy. It is lightweight and uses a simple convolutional backbone to achieve high-quality results without heavy computational costs. This is the smart way to restore images with speed and style.

🆔 @hackernewsgithubprojects

📦 llmllmllm/bpw

BPW: Fixing How We Test AI Systems

BPW is the diagnostic tool that ensures your AI system tests actually catch the problems hiding in plain sight. Traditional test compression looks for typical requests, but BPW hunts for rare, weird cases that expose deep system flaws. By preserving these specific pressure points, it creates a smaller replay suite that is statistically unique but diagnostically powerful. This framework helps engineers verify that their infrastructure can handle real world stress without wasting resources on redundant data. It is a vital step toward reliable, evidence based evaluation for large language model serving systems.

🆔 @hackernewsgithubprojects

📦 stokou/facet-terminal

FACET: How a Small AI Model Beats Big Ones on Terminal Tasks

FACET Terminal turns seven thousand raw code snippets into six thousand verified, executable computer tasks. The core problem it solves is that most AI-generated instructions are just text, often missing the actual environment needed to run them. This project builds the real computer setup first, then writes the task, solution, and checker based on that single, shared state. This ensures the challenge is actually solvable. The result is a small, fine-tuned model that scores higher on terminal benchmarks than much larger competitors. If you build AI agents, this is the framework to make them actually work.

🆔 @hackernewsgithubprojects

📦 nekodificador/comfyui-nkd-basic-tools

The ComfyUI Node Pack That Lets You Drag Handles On Your Images

NKD Basic Tools is the ComfyUI node pack that finally lets you drag handles directly on your images to pose facial expressions and grade color without leaving the canvas. This repository solves the busywork of complex workflows by bundling everyday utilities into single nodes, such as cropping a face for a model and stitching it back in seamlessly, or painting masks with a pen tool that saves editable control points. The most surprising feature is the live preview system, which allows you to adjust skin tones, blur fields, and film grain right inside the node without running the entire graph.

🆔 @hackernewsgithubprojects

📦 aws/agentcore-cli

AgentCore CLI: Your Agent in the Terminal

Spin up a full artificial intelligence agent right from your terminal. AgentCore CLI handles the heavy lifting of deploying, testing, and managing AI agents without forcing you to write infrastructure code by hand. It connects your agent to cloud resources, lets you test it locally before it goes live, and even helps you evaluate how well it performs. The real magic is that it turns complex cloud setup into simple commands, so you can focus on building your agent logic instead of wrestling with server configurations. If you want to move faster with AI development, this is your new command line companion.

🆔 @hackernewsgithubprojects

📦 miaai-lab/qwen3.8-27b-nvfp4-rtx-5090

Qwen 3.8 27B on a Single RTX 5090

A twenty-seven billion parameter language model now runs on a single consumer graphics card that still drives your desktop display. Mia AI Lab’s project, Qwen 3.8 27B on a Single RTX 5090, turns one script into a fully functional, OpenAI-compatible API server. The surprising capability here is that it achieves a full 256,000-token context window on a 32GB GPU, thanks to a heavily optimized 4-bit KV cache and a specific patch that fixes garbled output from the underlying inference engine. You get roughly 160 tokens per second for single-stream generation, with the system intelligently managing memory so it doesn’t eat your graphics card’s RAM needed for your actual desktop.

🆔 @hackernewsgithubprojects

📦 aarr-bench/aarri-bench

AARRI-Bench: The Research Intern Test

Stop asking if AI can code. AARRI-Bench asks if it can think like a scientist. This project drops AI agents into the shoes of a research intern, testing them on the messy, human parts of research that code alone can't solve. It’s not about writing algorithms; it’s about spotting a flawed experiment, reading between the lines of a lab partner’s email, and knowing when to quit a dead-end project. The goal is to expose the gap between a smart assistant and a real researcher. If you’ve ever wondered what an AI is actually missing when it tries to do science, this is the clearest way to see it.

🆔 @hackernewsgithubprojects

📦 adudeguyman/comfyui-h3-audiorefine

ComfyUI H3 Audio Refine

ComfyUI H3 Audio Refine lets you fix bad AI sound without re-rendering your video. MiniMax H3 generates video and audio together, so if the picture looks great but the audio sounds mushy, you usually have to restart the whole process. This tool freezes the finished video and runs extra cleaning steps just on the sound track, saving you from wasting hours on unnecessary re-generations. It even uses a smart memory cache to make those extra audio steps run much faster. If you hate waiting for perfect sound, this is a clever way to get quality without the penalty.

🆔 @hackernewsgithubprojects

📦 vibeinging/dsh-desktop

DSH Desktop: Your AI Workspace, Simplified

DSH Desktop turns your computer into a complete, offline AI workspace by running the official DeepSeek Harness interface right on your machine. It solves the headache of messy plugin setups by bundling everything you need into one clean, local app. The coolest part is its smart plugin system, which lets you add tools like file managers and project trackers without breaking your core setup. Because it runs locally, you get full privacy and control over your data, with no cloud dependencies slowing you down. If you want a powerful, self-contained AI environment that just works, this is the tool to try.

🆔 @hackernewsgithubprojects

📦 deeplethe/ontopilot

OntoPilot: The AI Knowledge Builder

OntoPilot is the local-first workbench that turns your messy company documents into a smart, usable knowledge map. It solves the problem of information getting stuck in PDFs by using AI to draft the structure, but it keeps humans in charge. The coolest part is that it actually learns from you. When an expert fixes a duplicate name, the system remembers that decision so it does not make the same mistake next time. It turns scattered notes into a living, version-controlled asset that your software can actually use. If your data is trapped in documents, this is how you set it free.

🆔 @hackernewsgithubprojects

📦 nuyoah-ai-works/nuyoah-xiezhen-prompt

The AI Photography Prompt Generator

One prompt engine now handles the entire workflow of professional portrait photography, from series masterboards to specific lighting physics. The project called nuyoah-xiezhen-prompt solves a real headache for AI creators who need consistent, high-quality results without guessing. It acts as a specialized agent that generates detailed Chinese instructions, specifically engineered to keep skin texture realistic and avoid that common oily, flat look from bad lighting. Instead of giving you generic text, it provides precise, ready-to-use prompts that you can hand off to any major image model. It is a clever bridge between your creative idea and the technical reality of digital photography, ensuring your portraits look human, not synthetic.

🆔 @hackernewsgithubprojects

📦 dsh-tauri-desk/deepseek-harness-desktop

The 5MB App That Runs Your AI Locally

DeepSeek Harness Desktop packs a full local AI agent environment into a five-megabyte installer that requires zero command line setup. If you want to run DeepSeek Harness on your desktop but hate juggling Node.js and Docker, this project does the heavy lifting for you. It automatically downloads the necessary runtime and core software on first launch, then runs everything locally on your machine. The interface is built with Tauri, keeping it lightweight, and it even adds a simple command to your terminal once installed.

🆔 @hackernewsgithubprojects

📦 amirh00sain/spiderpanel

SpiderPanel: The Ultimate Proxy Control Center

SpiderPanel is the web dashboard that puts total control over your network proxy setup in your hands. It solves the messy problem of managing user connections, security keys, and server configurations by bringing everything into one clean, easy-to-use interface. The most impressive part is how it handles the IP scanner: instead of using your server's bandwidth, it lets the user's own browser ping different locations directly. This means you can find the fastest connection for each person without slowing down your main system, giving everyone a faster, more reliable experience.

🆔 @hackernewsgithubprojects

📦 thenetworkchuck/n8n-terry-guide

This AI Employee Fixes Your Home Server

Run your own digital IT support that checks, diagnoses, and repairs your home network without waking you up at three in the morning. The n8n Terry Guide shows how to build this agent using plain English prompts instead of complex code, turning it into a true digital employee rather than a rigid script. The surprising part is how it handles tricky problems, like port conflicts, by investigating the issue itself and asking for your approval before making changes. You get peace of mind through simple alerts, knowing a smart assistant is keeping your systems healthy while you sleep, without needing to be a coding expert.

🆔 @hackernewsgithubprojects

📦 echomusicapp/echo-music

Echo Music: The YouTube Client That Identifies Songs

Turns your phone into a detective for any song currently playing around you. This open-source Android app, Echo Music, is a full-featured YouTube Music client that solves the problem of missing out on discovering the music in your environment. Its standout feature is built-in audio recognition, which listens to ambient sound and identifies the track for you, even without the song being streamed through the app. It then saves that track to your personal library, complete with lyrics and metadata, so you never have to wonder what a song was while walking down the street.

🆔 @hackernewsgithubprojects

📦 gigaforshovel/photo-enhancer-boost

Photo Enhancer Boost

Photo Enhancer Boost promises a no-nonsense creative workspace that actually just works, stripping away the clutter found in most heavy-duty editing suites. It is built for Windows and focuses on non-destructive editing, meaning you can tweak layers, masks, and filters without ever permanently ruining your original photo. The tool aims to keep things simple by offering precision retouching and color correction without the usual software bloat. If you are tired of wrestling with complicated interfaces, this project offers a clean, independent alternative for handling your images. It is a straightforward tool for getting the job done fast, and that is exactly why it is worth checking out.

🆔 @hackernewsgithubprojects

📦 messkan/prompt-cache

PromptCache: Stop Paying for the Same AI Answer Twice

PromptCache is the self-hosted proxy that stops you from paying for the same AI answer twice. It sits quietly between your app and your favorite AI provider, listening for questions that sound just like ones you asked before. If it finds a match, it serves the old answer instantly from a local cache instead of hitting the expensive model again. This single trick can slash your bills by up to eighty percent and cut wait times to near zero. You can even tune how strictly it matches similar questions to avoid mixing up different topics.

🆔 @hackernewsgithubprojects