📦 notasithlord/peerd

Peerd: The Browser-Native AI Agent Harness

Peerd is the browser-native AI agent harness that finally brings autonomous task execution directly into your existing web sessions. As a Chrome and Firefox extension, it runs a full agent loop inside your browser, allowing you to drive active tabs, execute shell commands in sandboxed WebAssembly Linux virtual machines, and manage client-side JavaScript notebooks. By utilizing your own provider keys and browser-level security, it operates without backends or telemetry. It even enables peer-to-peer sharing of agent-built applications using WebRTC. This experimental tool turns your browser into a powerful, privacy-focused engine for automated workflows and decentralized collaboration.

📰 https://news.ycombinator.com/item?id=48646165

🆔 @hackernewsgithubprojects

📦 adrianhajdin/roomify

Roomify: AI-Powered Architectural Rendering

Roomify is the architectural visualization platform that turns flat floor plans into photorealistic 3D renders with ease. Built with React and the Puter cloud platform, it leverages advanced AI models like Claude and Gemini to process your designs. You can generate stunning 3D imagery, compare your original sketches against the final result, and manage your projects with persistent metadata and high-performance storage. The system also includes a global community feed to showcase your architectural creations. Dive into this project to master serverless workflows and AI-driven visualization for your next design task or creative digital portfolio project.

🆔 @hackernewsgithubprojects

📦 facebook/astryx

Astryx: An Agent-Ready Design System

Astryx is an open-source design system that enables both developers and AI agents to build consistent, accessible interfaces using the same shared tooling. Born from years of internal development at Meta, it provides a library of over ninety React-based components that support features like automatic spacing compensation and flexible, themeable CSS variables. Its architecture is specifically designed to be agent-ready, offering JSDoc annotations and a command-line interface that allows AI assistants to scaffold, browse, and document components effectively. By offering transparent, composable primitives, this system helps teams streamline development and maintain high-quality UI standards across diverse application environments.

📰 https://news.ycombinator.com/item?id=48682464

🆔 @hackernewsgithubprojects

📦 modelplaneai/modelplane

Orchestrate AI Models Anywhere with Modelplane

Modelplane is the open source control plane that finally simplifies managing AI model infrastructure across any cloud or local environment. Built on Crossplane, it functions as an active reconciliation engine that lets you declaratively define and orchestrate your entire inference stack. It handles complex tasks like provisioning inference clusters, scheduling models, scaling replicas, and caching weights across multi-node setups without locking you into a single provider. Whether you are running a single GPU or massive distributed deployments, Modelplane gives you the tools to consistently automate your infrastructure, ensuring your AI systems remain reliable and scalable regardless of where they run.

📰 https://news.ycombinator.com/item?id=48646721

🆔 @hackernewsgithubprojects

📦 vasuki-re/istanpdf

IStanPdf: The Offline PDF and DOCX Utility

IStanPdf is the Android utility that handles all your document tasks without needing an internet connection. It serves as a completely offline alternative to online conversion services, letting you merge, split, reorder, or remove pages from PDFs and DOCX files directly on your phone. By bundling LibreOffice binaries to power its core operations, the app keeps your files secure and local while eliminating paywalls and upload limits. It is a practical tool for anyone who wants to manage their documents efficiently on the go. Try it out to take full control of your document editing workflow today.

🆔 @hackernewsgithubprojects

📦 jinyangwu/opid

OPID: On-Policy Skill Distillation for Agentic RL

Enhance agentic reinforcement learning performance by transforming completed episode trajectories into hierarchical hindsight skills. OPID functions as an on-policy skill distillation framework that routes both episode-level and step-level skills during training. By providing dense token-level supervision, it eliminates the need for external skill retrievers or privileged context during inference. This approach proves effective across complex environments like ALFWorld, search-based QA, and WebShop, consistently outperforming outcome-only reinforcement learning and standard skill-distillation baselines. Adopting this framework allows developers to build more efficient, context-aware agents that leverage hindsight to guide decision-making effectively throughout their training process.

🆔 @hackernewsgithubprojects

📦 opendatabox/awesome-agent-memory

Master AI Agent Memory Systems

Navigate the complex landscape of AI agent memory by using this curated taxonomy that organizes various systems across four distinct axes. This resource helps developers and researchers understand the evolution from simple sequential context to advanced topological structures and multi-paradigm hybrid containers. It provides a structured overview of essential reference baselines and critical benchmarks used to evaluate memory performance, retrieval fidelity, and operational stability. Whether you are building long-term memory for intelligent agents or studying the latest academic surveys, this repository offers a centralized foundation to master how agents remember, store, and utilize information effectively.

🆔 @hackernewsgithubprojects

📦 hellzerg/optimizernxt

Automate Windows Tweaks with OptimizerNXT

Automate complex Windows configurations using a powerful command-line interface designed for system administrators and power users. This tool streamlines system management by executing YAML-driven scripts to handle registry keys, DNS settings, startup programs, and service controls efficiently. It enhances privacy and performance by allowing you to batch uninstall unwanted applications and disable telemetry or unnecessary background services safely. For added security, the system enforces cryptographic signatures on all configuration files to ensure only verified operations run on your machine. Leverage this robust utility to gain full control over your Windows environment and optimize your daily workflow.

🆔 @hackernewsgithubprojects

📦 ika-rwth-aachen/ros2-depth-anything-v3-trt

Real-Time Metric Depth Estimation with ROS 2 and Depth Anything V3

This ROS 2 node achieves fifty frames per second performance by leveraging TensorRT acceleration for the Depth Anything V3 model. It transforms standard camera streams into dense metric depth maps and generates real-time three-dimensional point clouds. By subscribing to camera image and calibration topics, the system performs specialized post-processing, including focal length scaling and sky classification, to produce accurate depth data for robotics applications. Users can configure precision settings and debug visualizations through parameter files, making it a robust tool for developers working on automated driving or perception tasks. This implementation bridges advanced computer vision with high-performance robotics infrastructure.

🆔 @hackernewsgithubprojects

📦 haxl0p4/l0p4map

Advanced Network Mapping with L0p4Map

Visualize your network infrastructure with precision using this professional-grade monitoring tool. It combines high-speed ARP discovery and robust nmap integration into a modern interactive dashboard built with PyQt6. You can automatically fingerprint devices, map hierarchical network topologies, and analyze live traffic patterns with ease. The software excels at identifying host roles and detecting potential vulnerabilities, providing security researchers with deep actionable insights into their local environments. By automating complex scanning tasks and displaying them through intuitive graph layouts, this tool transforms raw network data into a clear and manageable map. Gain full visibility into your security posture today.

🆔 @hackernewsgithubprojects

📦 benchflow-ai/awesome-evals

Stop Guessing: A Non-BS Guide to Evaluating AI Agents

Build better AI agents by moving past basic accuracy metrics with a curated, practitioner-focused library of evaluation resources. This repository provides a high-signal, non-BS collection of papers, tools, and benchmarks, specifically designed to help developers bridge the gap between demo prototypes and production-ready reliability. You get access to a practical playbook featuring real, runnable code patterns for techniques like human-aligned LLM-as-judge scoring, pass-at-k estimation, and robust error analysis. By focusing on concrete workflows and verified, annotated resources rather than generic listicles, this project helps you ship software with confidence and gain clarity on how your agents actually perform.

📰 https://news.ycombinator.com/item?id=48683312

🆔 @hackernewsgithubprojects

📦 yanghuisong/proc-sam3

ProC-SAM3: Boosting Remote Sensing Segmentation

ProC-SAM3 achieves state-of-the-art results in remote sensing segmentation by improving how models interpret category prompts. This training-free framework solves issues with noisy activations and limited semantic coverage by using an offline prompt pool refined through category-specific constraints. By caching embeddings and implementing presence-guided residual fusion, the system effectively gates unreliable instance responses while preserving fine-grained details for small objects. Tested across eight diverse benchmarks, this approach outperforms previous methods by nearly four percentage points. The project demonstrates a sophisticated way to adapt powerful foundation models for complex geospatial analysis without needing additional training cycles.

🆔 @hackernewsgithubprojects

📦 inkeep/open-knowledge

Build an AI-Powered Wiki with OpenKnowledge

Manage your knowledge base with a beautiful, local-first markdown editor designed to function as an intelligent LLM wiki. This project simplifies technical documentation and project management by offering a familiar, Notion-like editing experience that supports collaborative AI agents through seamless integrations with tools like Claude and Cursor. By using git and GitHub for synchronization, it provides robust team sharing and versioning capabilities while keeping your data under your control. Whether you are building an agent second brain or managing engineering specifications, this tool helps you organize information effectively. Install the desktop app or CLI today to transform your markdown workflow.

📰 https://news.ycombinator.com/item?id=48675435

🆔 @hackernewsgithubprojects

📦 bymedion/macos-wine-steam

Run Windows Steam Games on Apple Silicon Mac Easily

Merlot is the one-click solution that simplifies playing Windows Steam games on your Apple Silicon Mac. By automating the complex setup of Wine and DXMT, it handles the entire installation process, including downloading the necessary components and configuring the environment for optimal performance. You can use its dedicated app launchers to manage specific game presets directly from your Applications folder or use the generic launcher for a broader setup. It streamlines the technical heavy lifting, allowing you to focus on your games instead of managing backend compatibility layers, providing a smooth path to enjoying your library on macOS.

🆔 @hackernewsgithubprojects

📦 sam1am/backlogia

Organize Your Entire Gaming Library with Backlogia

Consolidate your fragmented digital gaming collection into one unified library by using Backlogia. This local application aggregates games from major platforms like Steam, Epic, GOG, and Xbox into a single interface with smart deduplication. It enriches your collection with high-quality metadata, ratings, and screenshots from IGDB, while providing advanced filtering and sorting options to help you manage your backlog. You can create custom themed collections and even discover new titles based on your unique play history. By running entirely on your own machine, Backlogia keeps your library private and perfectly organized, making it easy to decide what to play next.

🆔 @hackernewsgithubprojects

📦 marcellm01/tinysearch

Local LLM Web Research with TinySearch

TinySearch empowers local large language models by providing a lightweight, source-grounded web research engine that functions entirely through local infrastructure. It replaces bulky search setups by automating the process of performing web searches, reranking the most relevant results, and intelligently crawling pages to extract key data. By transforming raw web content into structured prompts, it ensures that your AI agents receive concise, evidence-based context while keeping citation URLs attached to every piece of information. This tool is designed for developers who want to give their local agents efficient, reliable research capabilities without needing enterprise-scale systems.

📰 https://news.ycombinator.com/item?id=48153340

🆔 @hackernewsgithubprojects

📦 arcosoph/nanowakeword

Build Custom Wake Word Models Easily with Nanowakeword

Nanowakeword provides a complete end-to-end ecosystem for training high-accuracy, custom wake word models that deploy seamlessly from tiny edge devices to powerful cloud servers. This framework simplifies the entire lifecycle by automating data engineering, including phonetic adversarial negative generation to reduce false positives, and advanced model training using techniques like importance sampling based on loss. With support for diverse architectures—ranging from lightweight DNNs to state-of-the-art Transformers and Conformers—it allows for precise optimization of speed and noise robustness. This tool makes creating responsive, reliable voice triggers accessible for any project while keeping deployment efficient and flexible.

📰 https://news.ycombinator.com/item?id=47838289

🆔 @hackernewsgithubprojects

📦 bigstationw/comfyui-untwisting-rope

Master Training-Free Style Transfer with Untwisting RoPE for ComfyUI

ComfyUi-Untwisting-RoPE is the custom node suite that enables high-quality, training-free style transfer for DiT models directly within your existing workflow. By patching model attention layers and rescaling rotational position embedding frequencies, this tool allows you to accurately transfer styles between images without the need for additional model training. It includes powerful features like RF Inversion to maintain noise-matched trajectories and various gating mechanisms to refine the stylistic output. Whether you are using Flux or Z-Image models, these nodes provide precise control over how your target images inherit style from reference inputs, giving you professional results instantly.

🆔 @hackernewsgithubprojects

📦 plantpotatoonmoon/hpsv3-plusplus

HPSv3++: Scaling Reward Models for AI Image Generation

HPSv3++ is the advanced reward model that enables consistent image preference scoring across diverse generative capabilities and reinforcement learning stages. This repository provides a two-stage training pipeline and evaluation framework for a reward model that uses a capability encoder and feature-wise linear modulation to adapt to different generator qualities. It addresses the challenge of maintaining accurate preference signals throughout the optimization process of diffusion models. Users can utilize the provided scripts and weights to score images based on text-following and aesthetic quality, ensuring reliable feedback during development. Incorporate this tool to refine your AI image generation workflows.

🆔 @hackernewsgithubprojects