📦 calf-ai/calfkit-sdk

calfkit-sdk

Software agents can now dynamically discover each other and collaborate at runtime without you having to write any complex wiring or orchestrators. That is the magic behind calfkit-sdk, a clever framework for building decentralized, event-driven multi-agent systems. Instead of hardcoding how your agents talk to one another, you plug them into a shared communication mesh. Once they are live, a general assistant agent can automatically find a specialized finance agent and hand off tasks on the fly. It is completely decentralized, runs on Kafka-compatible streaming infrastructure, and lets you build fluid, reactive agent teams that scale independently.

🆔 @hackernewsgithubprojects

📦 jakubkrehel/skills

Skills

Instantly level up your AI agents so they can critique and polish your web designs like a seasoned pro. The skills repository is a collection of drop-in guidelines and plugins for Claude Code and Codex that turn your AI assistant into a sharp design engineer. Instead of giving vague advice, it helps the AI review your projects for precise CSS spacing, focus states, and color contrast. It covers everything from making sure screen readers work properly to picking the right mobile keyboards. Just run one quick terminal command to install these rules, and your AI is ready to audit your code.

🆔 @hackernewsgithubprojects

📦 wutishummus/truedeck

TrueDeck

Run multiple AI coding agents side-by-side in real terminal interfaces without any of the exhausting setup. TrueDeck lets you load Grok, Codex, Claude, and Cursor onto the exact same project codebase simultaneously, arranging them into split-pane terminal views. Instead of babysitting markdown files, configuring memory dashboards, or dealing with complex Docker environments, this smart terminal hub automatically syncs shared project context and memory behind the scenes. You get to interact directly with the authentic developer command-line interfaces while it handles the plumbing, giving you a powerful, unified cockpit to build software alongside your favorite digital assistants.

📰 https://news.ycombinator.com/item?id=49092280

🆔 @hackernewsgithubprojects

📦 konghayao/peri

Peri Code

This single Rust binary runs a complete artificial intelligence programming agent using only thirteen megabytes of space and fifty megabytes of memory. Called peri, it is a lightning-fast, resource-friendly terminal companion that is fully compatible with Claude Code configurations while letting you bring your own keys for DeepSeek, Qwen, or Anthropic on the fly. It features auto-compacting memory management to keep long coding sessions cheap, alongside seven specialized background sub-agents that collaborate on complex tasks. You get a gorgeous terminal interface with live markdown streaming and cross-platform support. It is the ultimate lightweight coding harness that respects your machine.

📰 https://news.ycombinator.com/item?id=49105673

🆔 @hackernewsgithubprojects

📦 perplexityai/numbat

Numbat

Numbat is the endpoint monitoring tool that finally gives you complete visibility into what AI agents are doing on your local machines. Instead of letting terminal assistants or coding agents run wild in the background, this tool hooks directly into their local processes to watch every action they take in real time. It normalizes all agent behavior into a single, clean event stream and evaluates it against smart local rules. You can use it to scan past session logs, generate clear activity timelines, and even set up custom guardrails to block risky network requests or file changes before they happen. It runs entirely on device, keeping your sensitive prompts and...

📰 https://news.ycombinator.com/item?id=49105621

🆔 @hackernewsgithubprojects

📦 fractera/agent-engineering-infrastructure

agent-engineering-infrastructure

The self-hosted deployment engine that lets you run a complete online business on your own private server using nothing but chat. Instead of paying hefty fees to cloud platforms, agent-engineering-infrastructure sets up your operating system, secure web routing, and databases on a private server in just ten minutes. From there, it coordinates a collaborative team of digital helpers to handle coding, marketing, and sales tasks. The coolest part is a clever built-in optimizer that cuts token costs by up to ninety percent, stopping your context windows from ballooning. It is the perfect setup to build and manage projects privately without cloud vendor lock-in.

🆔 @hackernewsgithubprojects

📦 alibaba-nlp/qqr

Train Smart AI Agents with Tournaments

Even the best artificial intelligence agents get stuck when standard scoring systems stop giving useful feedback, causing their progress to stall on complex tasks. That is where qqr comes in, an clever reinforcement learning framework that trains agents through competitive tournaments instead of static scores. By pairing agents up to battle in swiss-system or double-elimination brackets, qqr evaluates them based on relative performance, keeping the learning curve steep even in open-ended domains like deep research and travel planning. It is lightweight, integrates with standard tool environments, and offers a brilliant way to keep your agents evolving. Check it out and let your models compete their way to the top.

🆔 @hackernewsgithubprojects

📦 h-embodvis/turbovla

TurboVLA

Most robotic brain models require massive, power-hungry language processors to translate what they see into physical movement, but turbovla runs in real time at over thirty hertz on a single consumer graphics card while sipping less than one gigabyte of video memory. Instead of routing camera feeds through a giant language model to predict actions, this clever system bypasses the middleman entirely. It independently processes visual inputs and text commands, lets them talk directly to each other through lightweight interaction, and spits out smooth robotic movements. It is an incredibly smart bypass that delivers high-end dexterity without needing a room full of server racks.

🆔 @hackernewsgithubprojects

📦 evolvent-ai/rsibench-data

RSI Bench Data

Test the real-world coding and reasoning skills of artificial intelligence models in isolated sandbox environments. This project gathers rigorous test sets, including math competitions and complex software engineering problems, and evaluates how systems perform when given direct tasks. Instead of relying on simple multiple-choice questions, it runs model outputs inside safe, virtual containers to see if their code actually works and solves the problem. It is a practical playground for testing next-generation machine reasoning. Explore these datasets to see how modern models handle genuine software bugs and logical challenges, and start building more reliable systems today.

🆔 @hackernewsgithubprojects

📦 zju-real/relay-opd

Relay-OPD

Relay-OPD is the open-source machine learning tool that prevents reasoning models from wandering off-track during training. When a student model starts making a logical mistake, standard training usually lets it spin out of control, wasting massive amounts of computer processing power. Relay-OPD fixes this by introducing a clever handoff trigger that detects exactly when a student model is starting to fail. A larger teacher model then briefly steps in like a relay runner to correct the path before handing control right back. This simple relay intervention cuts total training data length in half while dramatically boosting the final model's mathematical accuracy.

🆔 @hackernewsgithubprojects

📦 taoliveaigc/aptavatar

AptAvatar

The audio-to-video generator that creates incredibly expressive speaking avatars in just two steps, AptAvatar completely changes how we generate lifelike video. Normally, turning audio into a realistic talking person takes dozens of slow computing steps, resulting in laggy, robotic movements. This clever system uses a unique distillation process to compress all that heavy rendering down to just two quick passes without losing any quality. You can even use step-by-step text prompts to direct specific hand gestures and head tilts at exact frames. It is a massive win for anyone needing high-quality, long-form presenter videos in a fraction of the time.

🆔 @hackernewsgithubprojects

📦 canonry/canonry

Canonry

Track how AI answer engines like Gemini, ChatGPT, Claude, and Perplexity cite your website so you can finally master engine optimization. Canonry is an open source, local first platform that acts as a central command center for your visibility across the modern AI landscape. It lets you monitor where you are being cited, analyze server logs to watch AI crawlers visit your pages, and diagnose your presence using real traffic data from search consoles and analytics. You can even submit direct fixes and coordinate updates. It is the ultimate toolkit to ensure your site stays visible in an AI driven world.

📰 https://news.ycombinator.com/item?id=49014880

🆔 @hackernewsgithubprojects

📦 eth-medical-ai-lab/agent-gui

Agent GUI

Watch your autonomous AI agents work right inside a virtual, scrollable office space where every task gets its own desk. With agent-gui, you can click on any desk to instantly see its live activity feed, inspect its file tree, watch the terminal, and view full debug messages. Instead of letting agents run blindly in the background, this project gives you a friendly interface to observe and steer them. You can even redirect and reassign agents mid-task without leaving the dashboard. It is a fantastic way to bring transparency and control to your multi-agent workflows.

🆔 @hackernewsgithubprojects

📦 musheer360/swiftslate

SwiftSlate

Instantly polish your writing, fix typos, or change your tone in any Android app without ever leaving your text field. By setting up a simple Android accessibility service, this clever tool lets you run custom AI prompts directly inside your active text box. Just type out what you want to say, add a quick trigger command like question mark formal, and watch your rough draft automatically transform into a professional message right before your eyes. It works system-wide using lightweight API integrations, giving you instant text editing power anywhere you type on your phone. Grab the app to make your mobile typing smarter, faster, and completely seamless.

🆔 @hackernewsgithubprojects

📦 hvision-nku/dreamstyle3d

DreamStyle3D

Generate high-fidelity, beautifully stylized 3D objects in less than ten seconds. DreamStyle3D is a clever framework that takes one content image and one style image, combining them into a textured 3D asset. Unlike older methods that struggle to keep the style consistent across different angles, this system uses a dual attention mechanism to separate the shape of the object from its artistic style. This means you can easily apply watercolor, ceramic, or sketch details to a 3D model without warping its original structure. It is incredibly fast, easy to run, and opens up amazing possibilities for game developers and artists looking to speed up their 3D creative workflow.

🆔 @hackernewsgithubprojects

📦 nanonets/graft

Graft

Graft is the code-mapping tool that finally stops AI coding assistants from wasting your tokens on repetitive codebase exploration. Whenever an AI agent starts a task, it blindly searches your files to build a mental map that it immediately throws away when the session ends. Graft solves this by building a permanent, local graph of your codebase once, saving it as a folder of simple markdown files right in your repository. It integrates directly with tools like Claude Code and Cursor to instantly feed them the exact context they need, making your AI runs up to three times faster and four times cheaper. Best of all, it automatically updates itself...

📰 https://news.ycombinator.com/item?id=49107075

🆔 @hackernewsgithubprojects

📦 nvlabs/cute

Understanding GPU Memory Layouts with CuTe

Design complex memory arrangements for high-performance computing without writing a single line of CUDA code. The cute repository provides a pure Python reference implementation of the hierarchical layout and tensor algebra used in Nvidia's high-performance computing libraries. It allows you to model how multi-dimensional data maps to physical memory addresses using simple, nested shapes and strides. You can experiment with complex matrix operations, visualize thread-to-value mappings, and design conflict-free memory patterns on your laptop with basic Python. It is an incredibly accessible way to master the advanced memory math that powers modern hardware acceleration.

📰 https://news.ycombinator.com/item?id=49103085

🆔 @hackernewsgithubprojects

📦 kunal12203/graperoot

GrapeRoot AI Context Engine

Slash your AI coding bills and get accurate answers on the first try by feeding your developer tools exactly the context they need. GrapeRoot runs locally to build a smart semantic map of your entire codebase, tracing how your files, classes, and imports connect. Instead of making your AI assistant blindly search through thousands of files or wasting expensive tokens uploading redundant code, this tool pre-loads the precise code snippets required for your specific question. It even remembers what you edited across a session so your savings compound over time. Just run it alongside your favorite developer tool to get faster, cheaper, and smarter programming help.

🆔 @hackernewsgithubprojects

📦 alielsokary/caskhub

CaskHub

CaskHub is the native Mac app store that finally brings a beautiful interface to your command line setup. It acts as a full graphical store for Homebrew casks, letting you search, install, update, and uninstall thousands of Mac apps without ever touching the terminal. The coolest part is its safe adoption feature, which lets you take apps you already downloaded from websites or the App Store and bring them under command line management with a single click. It is completely free, open source, and preserves full terminal compatibility, making app management incredibly satisfying.

🆔 @hackernewsgithubprojects

📦 tencent/angelspec

AngelSpec Speculative Decoding

AngelSpec is the training framework that finally unifies speculative decoding models under a single, highly efficient pipeline. Speculative decoding speeds up large language models by letting a smaller, faster draft model guess the next few tokens, which the main model then quickly verifies. Instead of forcing you to build custom setups for every new draft method, this tool brings six different major architectures together, letting you switch between them with a simple configuration tweak. It connects your training and generation processes across separate graphics chips using a clever shared memory store, so they can scale completely independently. If you want to make your language models run significantly faster in production...

🆔 @hackernewsgithubprojects