📦 nota-america/forgecat-agent-profiles

ForgeCat Agent Profiles

Grab production-ready setups for your favorite coding bots in seconds. ForgeCat Agent Profiles takes messy, scattered AI agent configurations and turns them into clean, installable packages you can grab with one simple command. It solves the headache of copying files between tools by giving you a single catalog that works across Claude Code, Cursor, and Codex. You get real, tested workflows from trusted sources, complete with proper licensing so you always know exactly what you are running. No more guessing if a setup actually works. Just search, install, and get your AI agents working exactly how you want, instantly.

🆔 @hackernewsgithubprojects

📦 anthropics/claude-tag-plugins

Claude Tag Plugins: The Ultimate AI Connector

Claude Tag Plugins is the open-source toolkit that lets Anthropic’s AI agents talk directly to the software your team already uses. Instead of writing complex code to connect Claude to every single tool, you simply install the specific plugins you need. The real surprise here is how modular it is. You can give Claude the ability to read your Jira tickets, query your Snowflake database, or even manage your on-call rotations without touching any manual setup. It works like a plug-and-play system, where each service acts as its own distinct package. This means your AI assistant only knows what it needs to know for your specific workflow.

🆔 @hackernewsgithubprojects

📦 0xsero/glm-5.3-flash-dgx-spark

Run GLM 5.3 Flash on a Single Spark

Run a three hundred and twenty billion parameter AI model on a single desktop computer, powered by just one GPU, using a simple container command. The GLM 5.3 Flash DGX Spark repository achieves this by compressing the massive model into a compact format that fits entirely in one hundred and twenty eight gigabytes of memory. This eliminates the need for expensive multi-server clusters, allowing you to serve high-quality, long-context AI responses locally. The project includes verified benchmarks proving the compressed model retains nearly all the original reasoning and vision capabilities. If you want to run frontier-level AI hardware on a single Spark device, this is the exact recipe to use.

🆔 @hackernewsgithubprojects

📦 ruc-datalab/evoontology

EvoOntology: The Self-Evolving Data Brain

Teaches your AI agent to understand complex databases by building a living map of your data instead of forcing it to guess every time. EvoOntology solves the problem where smart code tools stumble over messy tables because they lack context. It builds a clear guide that explains exactly what every column means, then updates that guide automatically as your agent learns. The surprising part is that the knowledge layer improves itself by watching how the agent works, making it smarter with every task. If you want your tools to stop making errors and start truly understanding your business data, this is the layer that makes it happen.

🆔 @hackernewsgithubprojects

📦 derpyu520/qq-bridge

qq-bridge: The AI That Joins Your Group Chat

qq-bridge is the messaging bridge that turns a standard AI assistant into a believable friend inside your group chat. It connects your favorite instant messenger directly to a local AI agent, allowing the computer to read, think, and reply to messages just like a real person would. The most surprising part is how it learns to fit in; the system actively studies the specific slang and internet expressions used by your friends, only adopting them after you confirm they fit the group's vibe. This creates a natural, low-key presence where the AI can observe the room, decide when to speak, and even stay quiet to let things breathe.

🆔 @hackernewsgithubprojects

📦 alesha-pro/tools

Tools: Make AI Video With Sound On Your Own PC

Tools is the open source collection that finally makes home video generation with sound actually practical. Built by someone running four high end graphics cards, it packages everything needed to run the MiniMax H3 model locally. The standout feature is synchronized audio. Instead of generating silent clips and bolting sound on later, this model creates the visuals and the matching voice and music in one single pass. It also includes smart workflows that split the heavy lifting across multiple graphics cards to cut render times from nineteen minutes down to under seven.

🆔 @hackernewsgithubprojects

📦 vinnylarouge/jevlike

Jevlike: The One-Pass Scorer

Skip the slow word-by-word generation and score every option in a single lightning-fast pass. Jevlike is a small, independent model that reads a prompt and a changing list of choices, then instantly assigns a probability to each one. It solves the problem of needing a dedicated, closed-source system for multiple choice tasks by giving you a clear, open starter. The coolest part? The same brain that picks text answers can actually control video game characters. It played Doom and chess by looking at the screen and choosing from a shared list of controller buttons. It is a clever, lightweight way to build models that make decisions in one shot.

📰 https://news.ycombinator.com/item?id=49731282

🆔 @hackernewsgithubprojects

📦 tele-ai/telechat3

TeleChat3: China's Homegrown AI

Train a massive AI model entirely on domestic Chinese hardware without relying on foreign chips. That is the core promise of TeleChat3, a large language model developed by China Telecom's AI institute. The team built the system to run on Ascend processors and Huawei's MindSpore framework, proving that high-end artificial intelligence can be developed and deployed locally. This matters because it reduces dependency on external technology and shows that local computing power can handle complex training tasks effectively. The model handles text generation, reasoning, and coding with impressive accuracy. If you are tracking the future of AI infrastructure, this project shows how nations are building self-sufficient digital capabilities right now.

🆔 @hackernewsgithubprojects

📦 poppuppy/amfd

One Step Image Generation: AMFD

AMFD can generate high quality images in a single step. It works by post training existing visual generators using neural amortizers that learn the statistical moments of real versus generated features. This technique solves the problem of slow and complex sampling processes in diffusion models. The project provides official PyTorch code for class conditional experiments on ImageNet with JiT and pMF, as well as text to image experiments with PixelGen and FLUX. It is useful for researchers who want to improve generation speed without sacrificing quality. The code is released under an MIT license and builds upon the FD Loss framework.

🆔 @hackernewsgithubprojects

📦 gavr728/comfyui_kleintiledupscaler

Klein Tiled Upscaler for ComfyUI

This ComfyUI node lets your graphics card create new image details in small slices without running out of memory. It splits your photo into equal blocks and uses an AI model to invent realistic textures for each one, rather than just blurring the original pixels. The tool is specifically built for the Flux model family, making it a powerful option for enhancing low-resolution art on budget hardware. It includes smart settings that detect empty areas like skies and skip them to save time, ensuring the computer focuses its energy only on the busy parts of your picture.

🆔 @hackernewsgithubprojects

📦 inkeep/visimer

Visimer: The Visual Editor for Mermaid Diagrams

Visimer is the visual editor that lets you tweak Mermaid diagrams right on the canvas. Most people write these technical charts as text code, which means fixing a single label requires hunting through a file of syntax. Visimer solves this by letting you double-click any label in the rendered image and type directly. As you type, the text code updates automatically behind the scenes. This is huge for anyone using AI to generate flowcharts, because you can now polish the visuals without touching the underlying code. It turns a text-heavy task into a simple drag-and-drop experience.

📰 https://news.ycombinator.com/item?id=49335470

🆔 @hackernewsgithubprojects

📦 goccy/wasm2go

wasm2go: Your WebAssembly, But Natively Fast

Your WebAssembly binary just became a native Go program. A tool called wasm2go takes that portable bytecode and spits out real Go source code, completely stripping away the usual runtime interpreter. You stop paying the heavy price of booting up a virtual machine just to run your compiled C or C++ library. Instead, the code links directly into your Go application, letting it run at native speed with almost no startup delay. It is a clever way to keep your high-performance compiled logic while enjoying the simplicity of standard Go builds, proving you do not have to choose between portability and raw speed.

📰 https://news.ycombinator.com/item?id=49732749

🆔 @hackernewsgithubprojects

📦 awlevin/typesafe-computer-use

TypeSafe Computer Use: AI That Costs Less Than a Penny Per Step

A Mac can now navigate to a checkout page and buy tickets using just a fiftieth of a cent per step. This project, called TypeSafe Computer Use, drives your computer by reading text off your screen and asking a tiny, specialized model what to click next. Instead of sending expensive screenshots to massive general AI, it uses deterministic code to handle the heavy lifting and only calls a larger model when you need to type a free-form sentence. The result is a tool that works in about a second per step, making automation fast, cheap, and predictable.

📰 https://news.ycombinator.com/item?id=49733647

🆔 @hackernewsgithubprojects

📦 ontos-ai/knowhere

Knowhere: The Memory for AI

Knowhere is the document engine that turns messy files into clear memory for AI. It reads your PDFs and slides, organizing them into a smart map instead of just random text. This means your AI assistant doesn't just guess what is on a page. It actually knows where things are, like a librarian with perfect recall. This is huge because it stops AI from hallucinating facts about your documents. It makes the data reliable and easy to trace back to the source. If you are building agents that need to read real papers or reports, this is the tool that makes them smart. Stop feeding raw files to your models.

📰 https://news.ycombinator.com/item?id=49738595

🆔 @hackernewsgithubprojects

📦 gcwing/openbitfun

OpenBitFun: Your AI That Never Stops

Keep your AI working across your laptop, phone, and even your smartwatch without losing your place. OpenBitFun solves the annoying problem of switching devices by letting you start a task on your desktop and finish it on your phone, keeping the context alive. It is not just a chatbot; it is a full workspace where you can build custom tools directly for your daily needs. The coolest part? You can actually ask the agent to write new code to extend the app itself. If you want an AI partner that adapts to your workflow and never loses its train of thought, this is the tool to try.

🆔 @hackernewsgithubprojects

📦 jakeschincariol/linkedin-agent-skill

Linkedin agent skill

Write LinkedIn content that actually sounds like a human. This open source pack gives Claude eleven specific instructions to help you draft posts, comments, and weekly plans. The real stand out is the humanizer tool. It scans your draft to remove invisible characters and overused words that scream artificial intelligence. It then scores your text on five different checks to make sure it reads naturally before you hit send. You stay in full control, deciding exactly what goes out. If you want to sound like a person and not a bot, grab this repo and try the humanizer on your next post.

🆔 @hackernewsgithubprojects

📦 typesafeainate/dspy-typesafeify

dspy-typesafeify: One Decorator, Smarter AI

dspy-typesafeify is the Python decorator that makes your AI programs think faster by routing specific answers to a specialized logic engine instead of relying entirely on the main language model. By simply adding one line of code to your existing setup, it identifies which parts of a task need strict, reliable logic and handles them separately, while still letting the AI generate free-form text where it excels. This hybrid approach cuts processing time and costs significantly without forcing you to rewrite your application.

🆔 @hackernewsgithubprojects

📦 maurodruwel/nimstats

NIMStats: The AI Benchmark That Runs Itself

NIMStats is the free benchmark dashboard that automatically tracks real-world performance for twenty-two major AI models. It solves the problem of trusting marketing claims versus seeing actual speed and reliability data by running a simple test on every model every single hour. The coolest part is that it requires zero infrastructure because it uses free automated tasks to do all the heavy lifting and hosts the results on a static site. You get a live leaderboard showing who is fastest and who is most reliable without paying for any servers.

🆔 @hackernewsgithubprojects

📦 honda-research-institute/hrdx

How Honda's HRDX Teaches Cars to See Maps from the Sky

Honda’s research team just dropped HRDX, a massive new benchmark that helps self-driving cars build their own maps using only their cameras. The standout feature? They prove that looking down from a plane helps. By training with aerial imagery, the system learns to see the road better without needing any extra sensors on the car itself. It’s like studying the answer key to ace the test, then walking away without the cheat sheet. This dataset covers fourteen hundred kilometers of real driving, making it a huge leap forward for safe autonomous navigation.

🆔 @hackernewsgithubprojects

📦 shi3z/deepseekv4.1-a100-custom

DeepSeek V4.1 Flash: 1 Million Tokens on 4 GPUs

DeepSeek V4.1 Flash is squeezing a full million-token context onto just four A100 GPUs without crashing. This custom inference runtime is interesting because it proves you don't need a massive cluster to handle ultra-long documents on standard hardware. It works by chopping the prompt into small chunks and pipelining them across the cards, preventing the memory explosion that usually kills these runs. The result is a stable, offline server that processes massive contexts reliably. If you are building tools that need to analyze huge files, this project shows that high-end long-context performance is accessible right now.

🆔 @hackernewsgithubprojects

📦 immortalhowwl/gptheist

Ten Agents, One Veto: Inside GPTHEIST

A single market decision passes through ten named checks, and any one of them can stop the entire process. GPTHEIST is a read-only observation tool for tracking new token launches on Robinhood Chain. It simulates a strict approval chain where distinct roles verify data, liquidity, and social signals before a final coordinator renders a judgment. Crucially, the system cannot place real trades; it only records paper-only decisions. This design creates a transparent, deterministic audit trail for analyzing how multi-agent systems handle risk. It offers a clear blueprint for building safe, verifiable workflows without touching live funds, making it a valuable study in coordinated decision-making.

🆔 @hackernewsgithubprojects