📦 aiha-lab/tangram

Tangram: Efficient Multi-Turn LLM Serving

Tangram is the large language model serving engine that finally makes non-uniform cache compression practical for multi-turn conversations. Built directly on top of vLLM, it solves a major bottleneck in AI applications where storing previous conversation history completely floods your system memory. Tangram intelligently analyzes which attention heads and layers are actually important to the context, and it dynamically frees up memory from the less critical ones. By reclaiming this physical memory in real time with zero runtime scheduling overhead, it converts raw memory savings into massive serving throughput. This means you can handle longer conversations and serve way more users simultaneously without upgrading your hardware.

🆔 @hackernewsgithubprojects

📦 visualbruno/comfyui-trellis2

comfyui-trellis2

The ComfyUI node pack that finally makes high-quality image-to-3D generation fast and accessible right inside your local workflow. ComfyUI-Trellis2 wraps Microsoft's powerful 3D generation technology into custom nodes, letting you turn flat images into incredibly detailed 3D meshes and textures. What makes this version so special is how it solves the messy post-processing puzzle, packing in smart features to fill holes in your meshes, weld vertices, and smooth out normals. You can even generate custom camera views and bake beautiful textures directly onto low-poly models. It is the ultimate tool for turning quick 2D concepts into clean, usable 3D assets.

🆔 @hackernewsgithubprojects

📦 dcdmllm/instructsam

InstructSAM

InstructSAM is the image editor assistant that actually understands what you want to isolate using everyday language. Instead of spending hours clicking and dragging anchors or typing weirdly specific code, you just tell it what to crop in plain English. You can ask it to isolate a category, describe a specific item in detail, or even give it logical reasoning clues to find the right object. It instantly pinpoints and isolates those exact boundaries in one go without laggy back-and-forth loops. It is a brilliant way to make complex computer vision feel as simple as texting a friend.

🆔 @hackernewsgithubprojects

📦 marc-shade/world-intel-mcp

World Intelligence MCP Server

Feed live global intelligence directly to your artificial intelligence agents using over one hundred specialized tools. The world-intel-mcp project connects your AI models to real-time data spanning thirty different domains, including financial markets, military flights, natural disasters, and cyber threats, all gathered from free public sources. It structures this massive stream of information so your digital assistants can instantly analyze complex global events. Even better, it continuously saves and indexes everything inside a local vector database, allowing your AI to run semantic searches on historical events to spot emerging trends or security risks. Use it to give your AI assistants real-world awareness.

🆔 @hackernewsgithubprojects

📦 tyedalwaves/hugginghack

HuggingHack

HuggingHack is the local-first hub manager that finally gives you total control over how you download and store open-source artificial intelligence models. Instead of dealing with massive, bloated downloads that eat up your hard drive, this clever tool lets you browse the live Hugging Face catalog and surgically select only the specific files you actually need, like a single GGUF or safe tensors weights. It runs right on your home computer or network attached storage, saving files in a clean, portable folder structure that plays nicely with runtimes like Ollama and vLLM. Grab HuggingHack to build your ultimate local AI library today.

🆔 @hackernewsgithubprojects

📦 vincentwei1021/video-shotcraft

Video Shotcraft

video-shotcraft is the motion design toolkit that turns AI coding assistants into fully automated video editing suites. Instead of wrestling with complex animation keyframes, you just point your AI agent at this library, and it instantly builds professional, high-end product promos using the Remotion framework. The project is loaded with over a hundred cinematic shot recipes, motion previews, and a complete ready-to-use video template featuring smooth camera sweeps and perfectly timed sound effects. It is like having a professional motion designer living right inside your terminal, ready to ship stunning videos on demand. Clone it to start making jaw-dropping product demos today.

🆔 @hackernewsgithubprojects

📦 drumih/turbo-fieldfare

TurboFieldfare

You can now run a massive twenty-six-billion-parameter artificial intelligence model on an eight-gigabyte Apple Silicon Mac using only about two gigabytes of system memory. This is possible because turbo-fieldfare is a custom Swift and Metal engine built specifically to stream the model's massive neural experts directly from your solid-state drive only when they are needed for a token. By using a clever caching strategy and custom Apple graphics kernels, it keeps the core system memory footprint tiny while maintaining surprisingly fluid text generation speeds on standard consumer hardware. It is the perfect playground for anyone wanting to run heavy local AI without buying expensive hardware.

🆔 @hackernewsgithubprojects

📦 duy-phamduc68/trafficlab-3d

TrafficLab 3D

The traffic visualization tool that turns flat traffic footage into a synchronized three dimensional digital twin using only standard security camera video and a Google Maps location. TrafficLab 3D links flat camera perspectives with satellite map views, showing you live three dimensional bounding boxes side by side with real time vehicle speeds and directions on the map. It handles everything from correcting lens distortion to estimating camera height and tracing lanes. This is a brilliant way to build high tech traffic demos without expensive sensors or specialized equipment. Grab your own video clips, align them to a map, and watch your local intersection come to life in three dimensions.

🆔 @hackernewsgithubprojects

📦 trycaspian/caspian-sdk

Caspian SDK

Caspian SDK is the open source library that finally lets your artificial intelligence agents talk to humans across every chat app using a single message handler. Instead of wasting weeks wrestling with separate connection pipelines for Slack, Discord, Telegram, and email, this tool unifies them into a single clean interface. You just write one incoming message function, and the system automatically handles the messy platform details, threading, and secure webhooks behind the scenes. It lets your agents seamlessly reply in the exact right thread whether the user reached out on social media or by email. Check it out and give your agent a voice everywhere.

🆔 @hackernewsgithubprojects

📦 openclaw/agent-skills

agent-skills

Your AI coding agent can now run production-grade software reviews and test compliance using shared, repeatable workflows. Instead of manually copying instruction files into every software repository, agent-skills provides a central library of plug-and-play capabilities for platforms like Codex and Claude Code. Its most impressive feature is an automated review helper that acts as a local code-quality gate before you commit changes, scanning for bugs and redacting sensitive secrets. By standardizing complex tasks like behavioral validation, prompt handoffs, and session transcript creation, it ensures your autonomous helpers always operate under the exact same safety rules and quality standards.

🆔 @hackernewsgithubprojects

📦 intisargit/aismith-3d

AISmith 3D

Turn a simple reference image into a fully rigged, animated three-dimensional asset without relying on cloud servers or subscription fees. This application brings a complete creation suite directly to your local Windows machine, letting you generate detailed geometry from a flat image using built-in artificial intelligence models. Once your shape is created, you can refine the mesh, paint textures, and even rig a complete skeleton with custom animations entirely in your web browser. It keeps your data and computing safe on your own hardware, downloading the heavy model files only when you decide to start a new design workflow.

🆔 @hackernewsgithubprojects

📦 danilo-znamerovszkij/draw-your-font

Draw Your Font

You can now turn a phone photo of your handwriting into a fully functional, installable custom font in seconds. The open source project draw-your-font lets you write out your alphabet on a plain sheet of paper, a napkin, or a spiral notebook, snap a quick picture, and generate web-ready files like TrueType or Web Open Font Format. It runs entirely on your local machine so your handwriting never leaves your computer, and you can even use it alongside artificial intelligence to automatically label your letters. It is a fantastic, free way to bring your personal handwriting into your digital designs instantly.

📰 https://news.ycombinator.com/item?id=49032751

🆔 @hackernewsgithubprojects

📦 mengto/skills

Skills

Skills is the prompt-engineering playbook that finally helps your AI agents build gorgeous web interfaces and playable games without losing their minds. Instead of relying on vague instructions and hoping for the best, this repository packages hyper-specific design systems, animations, and gameplay logic into reusable markdown files. Your AI helper can read these exact blueprints to handle complex tasks like scroll-driven storytelling, custom particle effects, or even responsive three-d web graphics. It basically gives your favorite coding assistant a masterclass in high-end design, ensuring you get polished, production-ready code on the first try.

🆔 @hackernewsgithubprojects

📦 jgravelle/jcodemunch-mcp

jcodemunch-mcp

jcodemunch-mcp is the model context protocol server that slashes artificial intelligence token costs by over ninety-five percent during code exploration. Instead of feeding entire files into your artificial intelligence assistant, this clever tool uses tree-sitter syntax trees to pull only the precise code symbols, imports, and relationships you actually need. It plugs directly into tools like Claude Code and Cursor, meaning your assistant can map out project architecture, trace dependencies, and find dead code without burning through your token budget. It keeps your context window clean and your development workflow incredibly fast.

🆔 @hackernewsgithubprojects

📦 different-ai/openwork

OpenWork

Share your AI workflows, skills, and connected services with your teammates instantly using a single tool. OpenWork acts as an open-source control center for your AI agents, allowing you to bundle custom skills, database connections, and API access into a single Model Context Protocol connection. This means you can build a workflow once and run it across different platforms like Claude Code, Cursor, or ChatGPT without rebuilding anything. It solves the headache of fragmented AI setups by giving your team a unified marketplace for custom tools. Grab the code, hook up your favorite model, and elevate your team's collective intelligence today.

🆔 @hackernewsgithubprojects

📦 max-sixty/worktrunk

Worktrunk: Git Worktrees Made Easy for AI Agents

Ever tried running five AI agents at once only to have them trip over each other's code? The tool called worktrunk solves this by completely rethinking how we use git worktrees. Instead of clunky native commands where you have to type paths and branches three times over, this rust-based tool lets you spin up, switch, and tear down isolated environments with simple branch names. It even copies ignored build caches like node modules or target folders so your agents do not suffer through slow cold starts. It is the perfect control center for parallel workflows, keeping your main directory pristine while your digital helpers build in parallel.

🆔 @hackernewsgithubprojects

📦 evolution-foundation/evolution-go

Evolution Go

Evolution Go is the high-performance messaging engine that brings lightweight, blazing-fast WhatsApp integration directly to your backend. Built from the ground up in Go, it completely bypasses resource-heavy setups to deliver minimal memory usage and rapid response times. The standout feature is its native support for passkey pairing, allowing you to link accounts locked behind modern biometric security using a simple browser helper. It easily handles real-time events through webhooks or message queues, manages media storage via compatible cloud buckets, and provides clean API endpoints for sending interactive carousels, lists, and media status updates. Experience lightning-fast, production-ready WhatsApp connectivity today.

🆔 @hackernewsgithubprojects

📦 vibepod/vibepod-cli

VibePod CLI

VibePod CLI is the terminal tool that finally makes running AI coding agents completely hassle-free. Instead of cluttering your local system with global packages or worrying about security, this utility launches popular coding assistants like Claude, Gemini, or Copilot inside isolated containers using Docker or Podman. It unifies all your favorite helper agents under a single command, automatically tracking their HTTP traffic and usage metrics. Everything gets saved directly to a local dashboard, giving you a clear, privacy-first view of what your agents are doing. It is the perfect setup for easily comparing different models on your real-world projects safely.

📰 https://news.ycombinator.com/item?id=48894261

🆔 @hackernewsgithubprojects

📦 taucetiproject/tauceti

Tau Ceti

Imagine a world where human mathematicians write the high-level roadmaps while artificial intelligence does the heavy lifting of writing and verifying the actual code. That is the reality behind Tau Ceti, an open-source mathematical library built on the Lean theorem prover. In this unique ecosystem, humans set the mathematical goals and design strict review rubrics, while AI agents write the proofs, submit code, and even conduct adversarial reviews against each other to find errors or weak arguments. It is a fascinating experiment in cooperative, machine-driven science that ensures math is verified at an incredible scale while remaining entirely free for everyone to build upon.

📰 https://news.ycombinator.com/item?id=49047064

🆔 @hackernewsgithubprojects

📦 cerebrasresearch/reap

reap MoE Compression

Researchers can now cut the size of massive mixture of experts language models in half while maintaining near-lossless performance on complex tasks. While traditional compression techniques merge neural network experts together and cause massive drop-offs in accuracy, a new tool called reap shows that intelligent pruning is actually the superior path forward. By analyzing both router gate values and average expert activation norms, reap calculates exactly which experts contribute the least to a model's output and safely removes them. On massive neural networks, this elegant approach successfully trims fifty percent of the experts with virtually no drop-off in coding or reasoning skills.

🆔 @hackernewsgithubprojects