📦 rynfar/meridian

Meridian

You can now run your favorite developer agents using a standard Claude Max subscription without paying for expensive API keys. A brilliant local proxy called meridian bridges the official Claude Agent software development kit to standard Anthropic and OpenAI API protocols. Instead of dealing with sketchy browser automation or binary hacks, it leverages documented SDK calls under the hood to handle your sessions, streaming, and prompt caching natively. By translating Claude's official output into standard API formats, you can instantly plug the model into tools like Aider, Cline, or Open WebUI. It is the perfect local bridge for developers who want subscription-grade AI in their favorite coding interfaces.

🆔 @hackernewsgithubprojects

📦 h3nock/remux

Remux

remux is the native iOS client that finally brings a mobile-first interface to your remote terminal sessions. Built on top of the Ghostty core, it turns your complex terminal workspace into an intuitive touch experience. You can swipe between windows, pick active panes from a live preview sheet, and split or close them with a tap. The absolute best part is how it handles server data. You can long-press a file path or a localhost web address to preview documents, images, and live development servers directly inside the app. It connects straight to your server over secure shell with zero middleman relays.

📰 https://news.ycombinator.com/item?id=49022284

🆔 @hackernewsgithubprojects

📦 neural-maze/production-ocr-course

production-ocr-course

Most text-recognition tools struggle when documents get messy, but this production-grade OCR course shows you how to build a self-scaling pipeline that reads charts and handwriting just like a human. Instead of using bloated cloud models, the system deploys a compact four-billion parameter model called Qwen three point five. It uses a high-speed Rust gateway built with Axum to ingest files, saves processing time using shared memory, and coordinates tasks using Redis. Best of all, it automatically scales graphics processing units up and down on Kubernetes to handle heavy loads. It is a complete masterclass in running optical character recognition at scale.

📰 https://news.ycombinator.com/item?id=49037050

🆔 @hackernewsgithubprojects

📦 powerycy/goutoujunshi

Goutoujunshi

Goutoujunshi is the romance adviser database that prioritizes emotional validation before analyzing relationships and generating actionable strategies. Instead of offering generic advice or manipulative tactics, this project acts as a structured knowledge base blending psychology, law, sociology, and ethics to support diverse relationships safely. It translates complex concepts like adult attachment styles and boundary setting into practical, everyday conversation guides for navigating conflict, active dating, and emotional support. By structuring real psychological evidence into testable, bite-sized social guides, it helps you understand connection without the mind games. Check out goutoujunshi to build healthier, more conscious relationships today.

🆔 @hackernewsgithubprojects

📦 vilmire/adhdev

adhdev

AI coding agents have evolved into long-running background workers, but babysitting their terminals all day is exhausting. adhdev fixes this by giving you a unified, self-hosted dashboard to monitor, control, and steer multiple agent sessions from your browser or phone. Instead of watching paint dry, you dispatch tasks across your local machines and let its clever engine verify, merge, and clean up the git branches automatically. It even handles tricky git submodules without breaking your repository. It is like having a command center for your digital workforce, so you can focus on writing code while they handle the busywork.

🆔 @hackernewsgithubprojects

📦 andrewyng/openworker

OpenWorker

Get actual work finished on your desktop instead of just chatting with an artificial intelligence. OpenWorker is a local first assistant that runs directly on your machine, orchestrating your files, terminal, and over twenty-five integrations like Slack, GitHub, and Jira. You can ask it to draft a customer brief, tidy up your calendar, or check a release status across platforms. It breaks down the task, builds the files, and crucially, always asks for your explicit approval before sending any messages or running shell commands. It is completely model agnostic, letting you plug in anything from Claude to local models using Ollama.

🆔 @hackernewsgithubprojects

📦 nyblnet/bento

Bento

Build and run complete presentation decks out of a single, portable file. Bento is an alternative to presentation software that packs its own editor, viewer, and presenter directly inside one five hundred and sixty kilobyte file. You just open it in any web browser, customize your slides, and save. The file itself rewrites its own code to keep your changes safe. Even better, it features local first collaboration with end to end encryption built right in. Your data stays in clean, open readable code, giving you complete ownership of your presentations forever.

🆔 @hackernewsgithubprojects

📦 silveroxides/convert_to_quant

convert_to_quant

The neural network optimizer that optimizes your machine learning models into tiny, hardware-friendly formats without sacrificing their performance. Convert to quant is a command line tool designed to compress heavy safetensors model weights down to ultra-low precision formats like eight-bit and four-bit representations specifically for ComfyUI. Instead of simply rounding off numbers and destroying model accuracy, it uses highly clever singular value decomposition and learned rounding optimization to dynamically adjust weights and preserve crucial details. You get incredibly fast inference speeds on consumer hardware while keeping your generated outputs razor-sharp. Just point the tool at your model, choose your precision, and watch it shrink.

🆔 @hackernewsgithubprojects

📦 izeigerman/claude-thermos

Claude Thermos

Claude Code has a secret fee that is quietly draining your wallet when you run long tasks. If your main coding agent is waiting on a slower subagent for more than five minutes, Anthropic silently expires your prompt cache, forcing a massive, expensive rewrite of your conversation history. Claude Thermos solves this by running a tiny local reverse proxy that watches your traffic. When it spots your main agent sitting idle while a subagent works, it automatically sends lightweight, cheap warm-up requests in the background to refresh the cache. It keeps your cache warm so you only pay a cheap read instead of a full rewrite.

📰 https://news.ycombinator.com/item?id=49024882

🆔 @hackernewsgithubprojects

📦 klura-ai/klura

Klura

Browser agents are incredibly slow because they navigate websites and click buttons on every single run, wasting massive amounts of time and language model tokens. Klura solves this by watching what a browser actually does during a task, capturing the underlying network traffic, and lifting those actions into a direct, reusable API call. Once it learns a workflow, it executes it in milliseconds using plain background requests instead of opening a browser UI. If a website changes or an endpoint breaks, its built-in self-healing system automatically catches the error and repairs the strategy. It essentially gives AI agents the speed of direct APIs for websites that do not have them.

🆔 @hackernewsgithubprojects

📦 21yrm/apple-pi

Apple-PI

Apple-pi is the diagnostic benchmark that finally holds modern video models accountable to the actual laws of physics. While today's models can generate videos that look visually convincing, they often fail to understand the underlying physical realities of the scenes they create. This project changes that by introducing a rigorous framework of ten classical mechanics tasks, prompting models to explicitly identify physical quantities, select the governing formulas, and generate precise trajectories. By combining machine learning metrics with advanced depth estimation and automated judging, it measures whether a model genuinely understands why things fall, roll, or bounce the way they do.

🆔 @hackernewsgithubprojects

📦 agricidaniel/claude-canvas

Claude Canvas

Claude Canvas is the Obsidian extension that lets you build and arrange visual boards entirely with voice commands and AI help. This tool turns Claude into a creative director for your canvas, letting you generate presentations, flowcharts, and mind maps in seconds. You can easily add text, structured tables, SVGs, and AI-generated images using simple terminal commands. It features six smart layout algorithms to clean up messy nodes instantly, keeping your boards perfectly aligned. If you want to build a complete slide deck or a beautiful mood board without manual dragging, this project does the heavy lifting for you.

🆔 @hackernewsgithubprojects

📦 cliffnodes/krea2-multi-character-lora-node-w-bounding-box

Krea2 Multi-Character Lora Node with Bounding Box

The ComfyUI custom node that finally stops your character models from bleeding into each other during AI image generation. This tool lets you draw precise boxes for each character, assigning a unique identity model to each region so your characters never blend together. Instead of relying on soft statistical nudges, this clever node physically masks the model's math, making it impossible for identities to leak outside their boundaries. The latest version even adds a reference lock to anchor a character's likeness across different frames, letting you compose complex, multi-character scenes with absolute control over who goes where.

🆔 @hackernewsgithubprojects

📦 ivanwng97/pixtuoid

pixtuoid

pixtuoid is the terminal visualizer that turns your active AI coding sessions into an adorable pixel-art office. Running multiple developer agents usually means constantly flipping between terminal tabs just to see who is stuck or finished. With this tool, every single agent session is represented by a tiny animated character at their own custom desk. You can watch them type while working, raise a question mark when they need your permission, and even doze off when their tasks are complete. It even features a procedurally generated lofi soundtrack and office pets you can interact with. It is the ultimate glanceable dashboard for your AI workforce.

📰 https://news.ycombinator.com/item?id=49039960

🆔 @hackernewsgithubprojects

📦 cliffnodes/krea2-multi-character-lora-node-w-bounding-box-by-fedor

Krea2 Regional Multi LoRA By Fedor

Put multiple distinct character identities into a single image without their features bleeding together. This ComfyUI extension lets you isolate each character's low-rank adaptation model to its own designated bounding box on the canvas. Instead of soft statistical nudges, it uses activation-delta masking to mathematically ensure that each character model is physically incapable of affecting pixels outside its box. The latest version even adds a reference lock engine, allowing you to anchor specific character likenesses to reference images right inside the node so their faces stay perfectly consistent across different generation runs.

🆔 @hackernewsgithubprojects

📦 asavvin-pixel/unslop

Unslop

Stop polishing machine-written drafts with simple word-swapping and actually make your AI text sound human. Most humanizers just trade overused words like delve for explore, but a major research study proved that detectors easily spot AI anyway because of how the paragraphs are built. Unslop is a clever tool for Claude that goes deep, fixing the structural giveaways like tidy moral-of-the-story endings, repetitive paragraph shapes, and fake details. It even calibrates to your actual voice using your own writing samples. It is a brilliant way to get natural, readable text you are actually proud to show a friend.

🆔 @hackernewsgithubprojects

📦 pullrun/pullrun

Pullrun

You can now run the exact same container image as a standard process, a secure Firecracker microVM, or a native Apple Silicon virtual machine without changing a single file or running a heavy background daemon. This magic is powered by pullrun, a tiny twenty megabyte tool that completely collapses the wall between containers and virtual machines. Instead of extracting files onto your drive, it uses a clever content addressed storage system that maps image data directly into memory. It even features peer to peer image syncing so a whole cluster can share packages at local network speeds. If you want ultra fast, secure sandboxing without the Docker bloat, pullrun is...

📰 https://news.ycombinator.com/item?id=49021880

🆔 @hackernewsgithubprojects

📦 makerspet/oomwoo

Oomwoo Open Source Robot Vacuum

Oomwoo is the open-source robot vacuum you build yourself that completely eliminates corporate cloud surveillance from your daily chores. This clever little machine combines a 3D-printed chassis with off-the-shelf parts like a Raspberry Pi, an affordable LiDAR sensor, and standard replacement vacuum motors. Because it runs locally on ROS2 and integrates directly with Home Assistant, your home maps and cleaning schedules stay private and completely cloud-free. It is highly customizable, letting you modify or upgrade its hardware and software modules exactly how you want. Gather your parts, fire up the 3D printer, and take back control of your smart home.

📰 https://news.ycombinator.com/item?id=49040684

🆔 @hackernewsgithubprojects

📦 obra/the-elements-of-style

The Elements of Style

Bring classic, razor-sharp writing principles straight into your modern coding workflow by loading a legendary style guide directly into your artificial intelligence assistant. This clever project packages William Strunk's nineteen-eighteen classic, the-elements-of-style, into a clean markdown plugin built for Claude Code. Instead of letting your AI generate wordy, passive, or cluttered documentation and user-facing text, this plugin feeds it the exact rules of composition, usage, and commonly misused words to edit prose with absolute precision. It is the ultimate way to ensure your AI writes as cleanly and clearly as a human editor. Level up your documentation today by giving your AI some actual style.

📰 https://news.ycombinator.com/item?id=49040760

🆔 @hackernewsgithubprojects

📦 nelthaarion/breeze

Breeze Go Web Framework

Imagine processing millions of web requests on a multi core server with virtually zero memory overhead. That is exactly what Breeze achieves by moving away from traditional Go routing and using an event driven architecture built on top of the ultra fast G net loop. It completely solves the performance bottleneck of memory allocation during heavy traffic by using a smart pooling system for request contexts and route parameters, meaning your server spends its time serving users instead of cleaning up memory. My favorite part is the built in real time developer dashboard that gives you instant charts on latency, database queries, and live requests over web sockets without any...

📰 https://news.ycombinator.com/item?id=49019328

🆔 @hackernewsgithubprojects