📦 sunblaze-ucb/exploitgym

exploitgym

exploitgym is the benchmarking suite that finally puts artificial intelligence agents to the test in the high-stakes world of cybersecurity. Instead of using artificial toy problems, it challenges automated models with a massive collection of real-world vulnerabilities sourced straight from Linux kernel bug reports and capture-the-flag competitions. The platform evaluates whether these intelligent agents can actually analyze a security flaw, write code to trigger it, and successfully demonstrate the vulnerability. By providing a strict, realistic environment, it helps researchers measure how close AI is to autonomously finding and fixing complex software bugs. Try exploitgym to see if modern AI is ready for real security work.

📰 https://news.ycombinator.com/item?id=49094493

🆔 @hackernewsgithubprojects

📦 gair-lab/light-mer

Light-MER

You can now run full multimodal emotion recognition using eleven times fewer floating point operations by shrinking a massive eight billion parameter teacher network down to a tiny sub-one billion parameter student model. The light-mer repository achieves this using a clever geometric distillation process to align hidden-state representations across different models. This means a lightweight model can process face cropped visuals, audio, and text with minimal memory while keeping the exact same expressive output interface. It is a practical way to run complex emotional understanding tasks on normal computer hardware. Try downloading their stage-one student checkpoint to see how fast on-device emotional intelligence can really run.

🆔 @hackernewsgithubprojects

📦 milkevich/routeveil

Routeveil

You can now add cinematic, full-screen transitions and shared-element animations to your React Router applications without cluttering your page components with complex animation logic. Routeveil coordinates the complete navigation lifecycle, seamlessly fading, sliding, or dissolving views while holding the router from committing the new page until the outgoing animation finishes. It includes twenty built-in effects like retro film-grain dissolves, circular iris wipes, and rotating venetian blinds. Developers get fully typed options and the ability to preview animations without changing the URL. Routeveil makes web navigation feel like a polished native application with zero configuration clutter.

🆔 @hackernewsgithubprojects

📦 gavamedia/deltafin

Deltafin: Running a 2.8 Trillion Parameter AI on a Single Mac

Deltafin is the local inference engine that lets you run a massive two point eight trillion parameter artificial intelligence model on a single Apple Silicon Mac or Linux workstation. Normally, a model this large requires an entire data center, but Deltafin makes it possible on home hardware by loading only the tiny fraction of the model needed for each word. While the core engine stays in memory, it dynamically streams the rest of the model's massive expert layers from the web directly into a local disk cache only when they are called. It even sets up a local network server compatible with popular coding and chat tools, giving you a...

📰 https://news.ycombinator.com/item?id=49090233

🆔 @hackernewsgithubprojects

📦 nolabs-ai/nono

nono

You can now run any AI agent on your computer while keeping your private files, SSH keys, and cloud credentials completely invisible to it. A security tool called nono lets you spin up zero-latency sandboxes for popular AI coding assistants in seconds with no containers, no virtual machines, and absolutely zero setup. It works by enforcing strict, least-privilege security boundaries directly through your operating system. What makes nono incredibly clever is that it does not just sandbox the agent itself, but also dynamically isolates the individual terminal tools the agent tries to run, like Git or curl, giving each one its own tailored security policy. It is a brilliant, lightweight...

🆔 @hackernewsgithubprojects

📦 deepcoldy/botmux

Botmux

You can now control powerful terminal-based AI coding assistants directly from your team chat. An open-source project called botmux bridges chat apps like Feishu and Lark straight to command-line developers like Claude Code and Gemini. Instead of wrestling with terminal windows, every direct message or group topic automatically spins up its own live-streaming terminal session. This means your team can collaborate with autonomous coding agents inside your existing chat workspace, watching them write and execute code in real time. It solves the headache of isolating and managing terminal instances, turning a solitary command-line experience into a shared multiplayer playground.

🆔 @hackernewsgithubprojects

📦 nizos/probity

Probity

Force your artificial intelligence coding assistants to actually follow your development rules with Probity. Instead of letting an agent write messy production code first, this tool hooks directly into your session to check every file write and shell command before it even happens. It acts as a lightweight policy engine, reading your agent's recent activity transcript to block actions that break your guidelines, like writing code without a failing test. It tells the agent exactly why it was blocked so it can immediately correct course. It works across different platforms, giving you complete control over your automated workflow.

🆔 @hackernewsgithubprojects

📦 comfy-org/comfy-aimdo

Comfy Aimdo

Your graphics card does not have to crash when running giant artificial intelligence models. Comfy Aimdo is a dynamic virtual memory allocator for PyTorch that keeps your hardware running smoothly even under heavy load. Instead of stuffing every single neural network weight into your graphics memory all at once, this clever tool reserves virtual address space for free and only maps physical memory right when a specific layer needs it. If your card starts running out of space, it automatically offloads lower priority weights to keep the system stable. It is a brilliant way to run massive workflows without constantly hitting memory limits.

🆔 @hackernewsgithubprojects

📦 codewithmukesh/dotnet-claude-kit

dotnet-claude-kit

dotnet-claude-kit is the development plugin that finally transforms Claude Code into a senior dot net developer. Instead of letting your AI assistant guess your architectural patterns or waste massive amounts of tokens reading entire files, this kit sits between Claude and your code with ten specialized agents and dozens of pre-configured developer skills. Its defining superpower is surgical code analysis, using a custom Roslyn-powered model context protocol server to fetch exact code symbols and structures in tiny, token-efficient queries. It handles everything from verifying your clean architecture boundaries to instantly scaffolding features with built-in tests, keeping your context window clear and your dot net code immaculate.

🆔 @hackernewsgithubprojects

📦 backnotprop/orchestrator

Orchestrator AI Agent Coordinator

Your main artificial intelligence agent can now automatically delegate tasks to a whole team of specialized sub-agents working in the background. A clever tool called orchestrator turns your primary AI assistant into a manager that coordinates other popular tools like Claude Code, Codex, and Grok. Instead of manually running different models for different tasks, you just tell your main agent what you want to achieve, and it spins up, monitors, and stops background workers as needed. It handles all the process supervision, letting you run parallel coding and design sessions using a simple, human-friendly command language.

📰 https://news.ycombinator.com/item?id=48853202

🆔 @hackernewsgithubprojects

📦 laravel/lsp

Laravel LSP

Your code editor is actually blind to half of your Laravel project, but this official language server completely changes that by making your editor fully framework-aware. Instead of just treating your code like generic PHP, it deeply understands your custom Blade directives, Inertia pages, Livewire components, and even your route names. It runs in the background to serve real-time autocompletion, instant hovers, and quick fixes right inside editors like VS Code, Neovim, and Zed. It even hooks into your local environment to automatically keep your Pest test helpers updated. It is the ultimate way to make your favorite editor behave like a custom-built Laravel IDE.

📰 https://news.ycombinator.com/item?id=49097178

🆔 @hackernewsgithubprojects

📦 six-ddc/ccbot

ccbot

Control your Claude Code sessions right from your phone by turning Telegram into a remote cockpit for your terminal. Instead of wrapping the API and isolating your sessions, this clever tool hooks directly into tmux on your machine. When you step away from your desk, your AI coding tasks keep running, and you can monitor, reply, or approve permissions on the go. If you want to jump back to your computer, simply attach to the tmux session and keep typing with your full history intact. It is the perfect way to keep your terminal as the source of truth.

🆔 @hackernewsgithubprojects

📦 libredb/libredb-studio

LibreDB Studio

Manage all your databases directly from your browser using a single lightweight tool. This project is a modern, web-based database editor that runs anywhere, giving you a full development environment without bloat. Instead of wrestling with heavy desktop clients, you can write queries, view execution plans, and manage structures from your laptop or phone. The coolest part is its automatic schema visualizer, which instantly maps out your tables and relationships into interactive diagrams. With built-in support for major databases, it makes exploring, editing, and monitoring your data incredibly simple. Give it a spin next time you need to query on the go.

🆔 @hackernewsgithubprojects

📦 egoist/kero

Kero Terminal Workspace

Kero is the native macOS terminal workspace that finally keeps your files, git history, and shell sessions perfectly together. Built on high-performance native Swift, it surrounds a lightning-fast terminal engine with dedicated project navigation panels like a built-in file tree, an integrated git pane, and a side-by-side diff viewer. The absolute best part about Kero is how it instantly adapts your workspace panels as you work: when you jump between git worktrees or run terminal commands, your file tree and git panes automatically follow your terminal's active directory. It is the ultimate lightweight command center for developers who want the structure of an IDE right inside their favorite terminal.

🆔 @hackernewsgithubprojects

📦 yoheinakajima/graphcon-deck

graphcon-deck

Your next presentation slide deck could actually be a single, living network visualization that rearranges itself dynamically as you speak. The graphcon-deck project turns traditional slides on their head by using nodes and edges to represent ideas instead of static pages. When you transition to a new slide, the tool pulls relevant nodes into view, pushes old ones out of frame, and centers the camera on your current topic. Because it is built as a single web visualization, you can easily reuse nodes across different parts of your presentation to show how your concepts connect. It makes dry slideshows feel like interactive, breathing maps of knowledge.

🆔 @hackernewsgithubprojects

📦 nbafrank/uvr

uvr

You can now manage your entire R data science project in under a second using a single tool. A new open-source package manager called uvr brings the incredibly fast workflow of Python's uv to the R ecosystem. Written in Rust, it completely replaces the messy combination of separate tools for tracking packages and managing R versions. It installs pre-compiled binaries instantly, handles project-isolated libraries, and locks down your exact dependencies in a simple lockfile. You can even install and pin specific R versions for individual projects without needing administrator rights. It is the ultimate speed boost for modern R development.

📰 https://news.ycombinator.com/item?id=49098984

🆔 @hackernewsgithubprojects

📦 igorfilippov3/senlo

Senlo

Senlo is the self-hosted email manager that frees you from delivery provider lock-in. Instead of tying your templates and analytics to a single service, this tool acts as a universal control center on your own servers. You can design gorgeous, responsive emails using a drag-and-drop visual editor and then send them through any major provider like Amazon SES, Resend, or Mailgun without rewriting your application code. By keeping your data in your own hands, it makes swapping email providers as simple as toggling a switch. It is a brilliant way to handle password resets, onboarding sequences, and transactional receipts independently.

🆔 @hackernewsgithubprojects

📦 abhishekbarali/speakoflow

SpeakoFlow

Your spoken words can now type themselves directly into any application on your screen without your voice ever leaving your device. SpeakoFlow is a free, local-first voice assistant for your desktop that translates, cleans up, and types your dictation in real time using your own machine's processor. By holding down a customizable hotkey, your speech instantly streams into your active text field, and starting with the phrase Hey Flow even instructs a built-in artificial intelligence to generate complete email replies or document drafts on the fly. You can host everything offline, select your own privacy-first language models, and even ask the floating assistant to answer questions about what is currently...

🆔 @hackernewsgithubprojects

📦 inclusionai/humming

Humming

Humming is the lightweight matrix multiplication library that brings extreme flexibility to quantized machine learning models. Built to speed up inference on Nvidia graphics cards, it compiles custom calculation kernels on the fly for almost any low-precision data type imaginable. While most tools lock you into standard formats, this library easily handles custom weights from eight-bit all the way down to tiny one-bit representations, across both standard and mixture of experts models. It packs all of this adaptability into a tiny package of just over one hundred kilobytes, giving developers a highly efficient way to run compressed AI models without sacrificing performance.

🆔 @hackernewsgithubprojects

📦 academiasd/academiasd_loralab-krea2

academiasd_loralab-krea2

Train your own custom image styles on a budget graphics card in under twenty minutes. This incredible software project makes training fine-tuned image styles for the Krea-two model accessible to anyone with standard computer hardware. Usually, customizing massive generative models requires high-end professional hardware and a mountain of memory. This project changes the game by lowering the hardware requirement to just eleven gigabytes of video memory and sixteen gigabytes of system memory. By making the training loop incredibly efficient, it lets you create high-quality, personalized styles right from your home computer without melting your hardware. It is the perfect gateway to affordable, local artificial intelligence creation.

🆔 @hackernewsgithubprojects