📦 natpate/ninfer-windows

Ninfer Windows: Run Big AI Locally

Run massive artificial intelligence models directly on your single graphics card without any cloud connection. That is exactly what ninfer windows provides. It is a specialized engine built specifically for Windows users who want serious performance. Instead of relying on generic, slow software, this project includes hand crafted code that squeezes maximum speed out of your specific hardware. It supports popular, large language models and handles the complex math in a way that is incredibly efficient. If you have been waiting for a way to run heavy AI locally without a data center, this is the tool you need.

🆔 @hackernewsgithubprojects

📦 agentsafelabs/safelabs-eval

Safelabs Eval: How To Stress Test Your AI Agent

Safelabs Eval lets you stress test your AI agent to find hidden weaknesses before they cause real problems. It works by firing thirty specific adversarial prompts at your system to see if it leaks data or obeys malicious commands. The coolest part is that it scores the answers using simple pattern matching instead of calling expensive AI models, so it runs fast and costs you nothing. You get a clear report on exactly where your agent failed, without needing to rewrite your code. It is a straightforward way to check if your bot is actually safe. Try it out to see if your agent can handle a bit of pressure.

📰 https://news.ycombinator.com/item?id=48297782

🆔 @hackernewsgithubprojects

📦 danialkhilji/homeos

HomeOS: The Kitchen Screen That Runs Your House

Turn an old laptop into the brain of your house with HomeOS. It runs on a Linux mini PC and gives every family member a shared touchscreen to manage chores, shopping, and prayer times without pulling out a phone. The coolest part is how it handles weekly task rotation; it automatically resets chores every Monday so nobody has to remember to clear the board. You get a calendar, live weather, and a store-based grocery list that actually looks like a real app. Since it is self-hosted, your family data stays on your own machine, not in some cloud.

📰 https://news.ycombinator.com/item?id=49535594

🆔 @hackernewsgithubprojects

📦 ryzqi/learn-agent

Learn-Agent: Build Real AI Agents

Learn-agent is the Chinese tutorial series that actually teaches you to build production-grade AI agents. Instead of just reading about theory, you get twenty chapters of runnable TypeScript code that shows you exactly how to wire up tools, permissions, and memory systems. It solves the common problem of agents getting stuck or making mistakes by giving you a clear, step-by-step architecture that you can run on your own machine. Whether you want to understand how to keep an agent safe or how to make it work on complex tasks, this repo gives you the practical skills to do it.

🆔 @hackernewsgithubprojects

📦 haohervchb/sglang-v100

Run Qwen 3.8 on Four V100s

SGLang for V100 lets you run massive AI models on four aging graphics cards from 2017. Most people assume you need brand-new, expensive hardware to handle large language models, but this project proves otherwise. It takes a powerful inference engine and tweaks it specifically for the unique architecture of these older V100 chips. By using clever memory tricks and optimized code paths, you can actually serve modern models like Qwen 3.8 Flash on this legacy hardware. It is a fascinating look at how software engineering can bridge the gap between old hardware and new AI capabilities.

🆔 @hackernewsgithubprojects

📦 bradgroux/veritas-kanban

Veritas Kanban: The AI Truth Teller

Your AI agents are lying to you about progress, and Veritas Kanban is the antidote. It’s a local-first task board where every agent action is bound to a verifiable audit trail, meaning you get the unfiltered truth about where your project actually stands instead of just a cheerful status update. This lightweight orchestration harness gives you a visual board to track work while letting you plug in agents as optional layers, so you can see exactly what they did without trusting their word. It’s a clever way to keep AI honest while staying in control. If you’re running agents, this is the reality check you need.

🆔 @hackernewsgithubprojects

📦 ningbainb/deepseek-harness-desktop

DeepSeek Harness Desktop: The Zero-Setup AI Workspace

Install a fully functional artificial intelligence workspace on your Windows PC without touching a single command line. DeepSeek Harness Desktop bundles the entire coding environment, including terminal access and mobile remote controls, into a single installer that requires no prior setup. It solves the tedious problem of manually configuring complex AI agents by handing you a ready-to-run application with built-in plugins and visual themes. This means you can start building with advanced language models immediately, bypassing the usual technical hurdles. If you want to test powerful AI tools without the configuration headache, this is the straightforward way to begin your next project.

🆔 @hackernewsgithubprojects

📦 omdsh-dev/dsh-mnemon

DSH Mnemon: The Missing Memory Layer

DSH Mnemon gives your coding agent a proper long-term brain so it actually remembers what it learned. Instead of forgetting everything between sessions, it splits knowledge into three clear buckets: quick notes for right now, full documents for deep dives, and persistent facts for everything else. The coolest part is that it lets you plug in almost any memory backend you like, so you are never locked into one specific database. It turns a chatbot that keeps losing context into a reliable partner that recalls decisions you made months ago. If you are tired of repeating yourself to your AI, this is the layer that finally fixes it.

🆔 @hackernewsgithubprojects

📦 saurabhav88/enviouswispr

EnviousWispr: Private AI Dictation for Mac

EnviousWispr turns your Mac into a private dictation machine that never sends your voice to the cloud. It runs entirely on your Apple Silicon chip, using local models to transcribe speech in under a second. The standout feature is their custom AI polish model, which cleans up grammar and filler words right on your device, so your audio never leaves your computer. It’s free, open source, and works fully offline. If you want to type faster without sacrificing your privacy, this is the tool to watch.

📰 https://news.ycombinator.com/item?id=48963964

🆔 @hackernewsgithubprojects

📦 kunchenguid/vision

Vision: The AI Skill That Finds Your Project's Soul

Vision is the coding assistant that finally writes down what your project actually stands for by reading your past work. Instead of asking you to list vague values, it studies your real history of accepted and rejected code to draft a concrete vision document. The surprising part is how it tests that draft by throwing tough, hypothetical scenarios at you on a simple review board. You decide whether a new idea fits your project, and the tool updates the writing based on your answers. This creates a clear rulebook that helps humans and machines agree on what the project is and is not.

🆔 @hackernewsgithubprojects

📦 sylvinus/agent-vm

agent-vm

agent-vm lets you run autonomous A.I. coding agents inside a safe, isolated Linux virtual machine while keeping your actual computer completely untouched. The tool creates a lightweight Debian environment that mounts only your current project folder, meaning the A.I. can execute any command, install dependencies, or run containers without ever exposing your personal files, cloud credentials, or system keys to potential supply-chain attacks. You get a fully equipped workspace with Docker and a headless browser ready to go, so you can let the agent work with full permission without worrying about accidental damage or malicious code.

📰 https://news.ycombinator.com/item?id=47301380

🆔 @hackernewsgithubprojects

📦 dgillis/emacs-codex-ide

Codex IDE for Emacs

Codex IDE for Emacs renders artificial intelligence code suggestions directly into your editor as native text buffers instead of hiding them in a separate terminal window. This project solves the problem of context switching by turning AI output into clickable, syntax highlighted code blocks that behave exactly like your regular files. It uses Emacs’ built in diff engine to show changes clearly, so you can review edits with the same familiar tools you already use. You can even let the AI see what you are currently editing to give it better context for its suggestions.

📰 https://news.ycombinator.com/item?id=47664195

🆔 @hackernewsgithubprojects

📦 oh-my-dsh/dsh-plugin-upgrade-skill

dsh-plugin-upgrade-skill: The Upgrade That Tests Itself

dsh-plugin-upgrade-skill is the maintenance tool that actually proves your plugin survives the update. Most upgrades break things silently, but this project includes a massive benchmark suite of tricky test scenarios called traps. It runs your plugin through these real-world headaches to see if it still works after a version bump. If it fails, you know exactly why before your users do. It’s a safety net for developers who want to ship updates with confidence. This is the upgrade process that checks its own homework. Your plugins will finally stay in sync without the guesswork.

🆔 @hackernewsgithubprojects

📦 robbyant-research/dsaquant

DSAQuant: Fast Video Generation Explained

DSAQuant is the video generation tool that makes massive AI models run faster on standard graphics cards. It solves the problem of slow rendering by shrinking the math inside the model, letting it create videos using far less computer memory. The most surprising part is that it keeps the picture quality sharp even after cutting the data size in half. This means your local computer can handle heavy video creation without needing a data center. If you build visual apps, this is how you get professional results on everyday hardware.

🆔 @hackernewsgithubprojects

📦 dashdogy/rtx40mfg-unlock

RTX 40 MFG Unlock: Pushing DLSS Limits

RTX 40 MFG Unlock is the experimental mod that lets you push DLSS Multi Frame Generation beyond the default limits on your graphics card. Normally, game developers cap how many extra frames your chip can create, but this tool overrides that ceiling to expose up to six times more frames in compatible titles. It works by quietly intercepting the frame generation process before it starts, ensuring your card runs as fast as possible without touching any of the original game files. This gives you noticeably smoother motion in heavy scenes, though the author warns it is still research software that can cause glitches if things go wrong.

🆔 @hackernewsgithubprojects

📦 evomap/autoresearch

AutoResearch: AI That Writes Its Own Evidence

AutoResearch doesn't just give you a chatbot answer; it runs actual experiments and writes a full evidence package you can turn into a paper. It starts by scanning recent papers and code trends to find gaps, then uses multiple different AI models to brainstorm and debate ideas so one model can't just rubber-stamp its own work. It plans a pilot experiment, runs it, and if the idea fails, it honestly records the negative result instead of spinning it into a success story. Everything is saved to disk so you can inspect the code, logs, and reviews later. It is basically a tireless research assistant that shows its homework.

🆔 @hackernewsgithubprojects

📦 shootthesound/comfyui-mesh

ComfyUI Mesh: Split AI Models Across GPUs

ComfyUI Mesh lets you run massive image generators by splitting the work across two separate graphics cards, even if they are on different computers. The tricky part is that these cards need to constantly send huge amounts of data back and forth, which usually slows everything down. This project solves that by treating that raw data like a video stream. It uses a specific chip inside your graphics card, called N V Enc, that is usually sitting idle, to compress the data right before it leaves the card.

🆔 @hackernewsgithubprojects

📦 intellisensing/geochrono

How Computers See the Past

GeoChrono can ask a satellite image what happened last Tuesday, not just what is there right now. Existing tools treat space photos as single snapshots, missing the story of how places change over years. This project fixes that by giving a vision model the ability to recall and reason about long-term evolution. It uses a special encoder to track how a single location transforms over time, rather than just comparing two dates. The result is a system that actually understands the history behind the pixels.

🆔 @hackernewsgithubprojects

📦 theagentplane/tokenops

TokenOps

TokenOps can cut your AI agent’s wasted spending by sixty-five percent just by checking the bill before every single step. It treats an entire multi-agent workflow as one single job with one shared budget, not a bunch of separate calls. Instead of letting a complex task quietly burn through your credits, it watches the total spend in real time and stops the run the moment it hits the limit. This prevents that sneaky scenario where dozens of tiny, cheap actions add up to a surprisingly huge final invoice. It is basically a guardrail for your AI’s money.

📰 https://news.ycombinator.com/item?id=49574349

🆔 @hackernewsgithubprojects

📦 shiphrahx/ai-for-engineering-leaders

AI for Engineering Leaders: Your New Copilot for Management

There is a specific, practical playbook on GitHub called AI for Engineering Leaders that turns the chaotic art of managing teams into a structured, copy-and-paste workflow. It isn't just a vague list of tips; it is a library of over one hundred ready-made prompts and skills designed for the messy reality of engineering leadership. If you are a manager, director, or VP, this repo hands you the exact words to use for your next executive status update, a tricky performance review, or a critical incident postmortem.

🆔 @hackernewsgithubprojects

📦 johndpope/llama-cpp-turboquant

llama-cpp-turboquant

llama-cpp-turboquant is the local inference engine that squeezes massive language models onto your regular computer hardware. Most people think running big A.I. models requires expensive cloud servers, but this project lets you run them directly on your laptop or desktop without sending your data anywhere. It uses clever math called quantization to shrink the model files, which means you get faster speeds and less memory usage. If you care about privacy or want to play with A.I. offline, this is the tool that makes it possible right now. Stop renting your A.I. and start owning it. Your next project runs locally, and it is actually quite impressive.

🆔 @hackernewsgithubprojects