📦 mit-spark/daaam

Daaam

Build 3D Dynamic Scene Graphs on the fly with real-time semantic descriptions for robots. By combining SAM segmentation with BotSort tracking, this system continuously labels objects in 3D space. It solves the massive computational bottleneck of querying heavy multimodal models for every single camera frame. Instead of redundant calls, its assignment service runs a mixed-integer program to select only the most salient, non-overlapping frames before prompting a Describe Anything Model. You get highly accurate spatial-temporal memory without choking your hardware. Grab a ROS 2 workspace, hook up a depth camera, and let your robot map and understand its environment dynamically.

🆔 @hackernewsgithubprojects

📦 alisa0808/vibe-creating-skill

vibe-creating-skill

vibe-creating-skill is the open-source prompt-engineering skill that finally stops you from over-specifying technical camera parameters for video generation. Rather than cramming your prompts with useless lens sizes and shutter speeds, this portable agent skill translates your rough stories into model-ready prompts that focus entirely on emotion, tone, and visual anchors. It evaluates your input on three distinct axes, outputting a consistent four-part response containing a judgment, action, result, and notes. Compatible with agents like Claude Code, it prepares your concepts perfectly for text-to-video engines like Seedance. Try feeding your raw ideas to vibe-creating-skill and let the AI handle the heavy visual rendering.

🆔 @hackernewsgithubprojects

📦 walter-grace/mac-code

mac-code

You can now run a thirty-five billion parameter AI agent on a standard sixteen gigabyte Mac mini at thirty tokens per second without paying a dime in API fees. The mac-code project achieves this by running local models like Qwen three point five entirely on Apple Silicon. It features an innovative flash streaming architecture that pins critical attention weights in RAM while streaming massive feed-forward network weights directly from your SSD on a layer-by-layer basis. By bypassing the macOS page cache and utilizing expert-level streaming for mixture-of-experts models, it keeps memory usage completely flat. This gives you a private, zero-cost coding assistant right in your terminal.

🆔 @hackernewsgithubprojects

📦 agricidaniel/claude-repurpose

Automate Content Repurposing with Claude Code

Claude Repurpose is a specialized engine for Claude Code that transforms a single piece of content into over ten platform-optimized posts in seconds. By automatically extracting key insights, quotes, and statistics from YouTube videos, blog posts, or local files, it distributes your message across six parallel AI agents to ensure every output feels native to platforms like LinkedIn, Twitter, and TikTok. It goes beyond simple cross-posting by adapting character limits, formatting, and hooks based on current algorithmic benchmarks. This tool helps creators, marketers, and solopreneurs scale their reach efficiently, saving hours of manual labor while maintaining professional, high-engagement content standards.

🆔 @hackernewsgithubprojects

📦 kyutai-labs/pocket-tts

pocket-tts

This lightweight text-to-speech model runs entirely on your CPU and starts generating audio in just two hundred milliseconds. Called pocket-tts, it squeezes a high-quality speech synthesizer into a tiny one-hundred-million parameter footprint that uses only two CPU cores. Instead of relies on heavy GPUs or slow web APIs, you can run voice cloning and streaming audio locally at six times real-time speed on a basic MacBook Air. Under the hood, it pairs a transformer-based flow language model with the Mimi neural audio codec to synthesize natural speech. It is a brilliant piece of local-first engineering that makes on-device speech synthesis incredibly accessible.

🆔 @hackernewsgithubprojects

📦 bayramannakov/claude-reflect

claude-reflect

Your Claude Code CLI can now automatically learn from its mistakes and remember your preferences forever. The claude-reflect plugin uses a hybrid system of real-time regex matching and semantic AI validation to capture your corrections, positive feedback, and workflow patterns directly from your terminal sessions. Instead of repeating instructions, these captured learnings are automatically queued, reviewed, and synced into your CLAUDE.md files and memory rules. It even analyzes your session history to discover repeating tasks and draft custom reusable commands for you. Install claude-reflect to build a permanently smarter terminal assistant that adapts specifically to your development habits.

🆔 @hackernewsgithubprojects

📦 kageroumado/adrafinil

Adrafinil

Your Mac can bypass clamshell sleep and run tasks with the lid closed only when an active AI agent is actually working. Standard tools like caffeinate keep your machine wired forever, draining your battery in your bag, but adrafinil is a clever eugeroic for macOS. It uses a tiny root helper to toggle system sleep rules dynamically when agents like Claude Code or Cursor start a task, then completely steps aside the moment they finish. To keep your laptop safe, it features a built-in thermal cutout that forces sleep if your machine starts cooking. Grab this tool to keep your background tasks running safely.

📰 https://news.ycombinator.com/item?id=48701512

🆔 @hackernewsgithubprojects

📦 fakechris/obsidian_vault_pipeline

Obsidian Vault Pipeline

Your local Obsidian vault is no longer just a collection of static files. obsidian_vault_pipeline transforms your notes into an auditable knowledge state runtime using local intelligence. Instead of just searching your files, this system runs six structured pipeline stages: it ingests raw inputs like Pinboard links, GitHub repositories, and web clippings, compiles them into structured claims, evidence, and relations, and resolves internal contradictions. It then projects this processed knowledge into dynamic reader atlas pages, interactive graph views, and a local web interface. It makes your notes active, connected, and completely traceable. Check out this clever local pipeline to make your vault think.

🆔 @hackernewsgithubprojects

📦 nirdiamant/agent_memory_techniques

Agent Memory Techniques

Build smarter LLM agents that actually remember past conversations by exploring thirty runnable Jupyter notebooks filled with real Python code. The agent_memory_techniques repository solves the persistent problem of AI agents forgetting user context between sessions. It walks you through short-term conversation buffers, long-term vector stores, and cognitive structures like hierarchical memory layers. You can run code for advanced self-reflection patterns, episodic-to-semantic knowledge graph extraction with Graphiti, and frameworks like Mem0 or Letta. It is a highly practical playbook for anyone wanting to build personalized, long-lasting AI assistants that learn from every single interaction.

🆔 @hackernewsgithubprojects

📦 prithvi-web/treemap

TreeMap

Visualize all of your disk space to reclaim gigabytes of storage using interactive, color-coded squarified treemaps. This self-contained application runs locally on macOS, Windows, and Linux without any trackers or external dependencies. It features an eight-way concurrent disk scanner that maps out files and folders while offering smart cleanup suggestions for leftover system caches and build outputs. It even tracks historical snapshots to show exactly which directories grew since your last scan. Best of all, it deletes files safely through your native operating system trash so you can recover them at any time. Download the app today to master your machine's storage.

📰 https://news.ycombinator.com/item?id=48548995

🆔 @hackernewsgithubprojects

📦 vukrosic/llm-research-kit

Build and Train Your Own LLM with This Research Kit

LLM Research Kit is the high-performance codebase that finally provides a modular foundation for training and optimizing language models. Designed for researchers, this repository includes everything needed to test new transformer architectures, experiment with advanced optimizers like Muon, and conduct full pretraining runs on one or more GPUs. The toolkit features a flexible configuration system, pre-built training scripts, and a robust benchmark suite for evaluating models on commonsense reasoning tasks. By simplifying the pipeline from data download to performance analysis, this project gives you the structured environment necessary to iterate quickly and master modern AI research.

🆔 @hackernewsgithubprojects

📦 sybil-solutions/local-studio

Local Studio

Launch, manage, and run your self-hosted LLMs right from your own machine with a unified control panel. local-studio bridges the gap between raw inference engines and a clean user experience by giving you a centralized Bun and Hono controller that directly coordinates backends like llama.cpp, vLLM, SGLang, and MLX. You can monitor live GPU metrics, spin up a Pi coding agent session, and deploy models locally or on a remote GPU server. With its Next.js frontend, terminal TUI, and desktop Electron shell, you get complete authority over your self-hosted AI stack. Grab your models and run them your way.

🆔 @hackernewsgithubprojects

📦 metalogico/issued

Host Your Own Personal Comic Library Server

Issued acts as a fast, self-hosted server that transforms any folder of digital comics into an organized library accessible from your web browser or mobile reader. It supports common formats like CBZ, CBR, and PDF while automatically extracting metadata and generating thumbnail previews for your collection. You can either use the built-in web reader to browse and read directly in your browser or connect your favorite OPDS-compatible mobile app to read on the go. The server handles automatic file monitoring for new issues, giving you a centralized way to track your reading progress and manage ongoing series.

🆔 @hackernewsgithubprojects

📦 rightnow-ai/picolm

PicoLM

picolm is the minimal local inference engine that runs a one-billion parameter language model on a ten-dollar hardware board with only forty-five megabytes of runtime RAM. By memory-mapping a quantized GGUF model file directly from disk and streaming it one layer at a time, this zero-dependency engine bypasses the typical memory constraints of embedded systems. It packs an entire neural network forward pass, a custom tokenizer, and even a grammar-constrained JSON output mode for reliable tool calling into just twenty-five hundred lines of pure C. It is a brilliant way to build completely offline, private AI hardware without any cloud API bills.

🆔 @hackernewsgithubprojects

📦 baairon/torlink

torlink

torlink is the terminal-native torrent finder that lets you search and download directly from your command line with absolutely zero configuration. It streams live results from a curated catalog of trusted sources, prioritizing safety by limiting executable game downloads to highly verified repackers while keeping movies, anime, and TV to plain video files. You can paste a magnet link, search with text, and track your active downloads with live speeds and progress bars. Once finished, downloads seed automatically in the background, and everything resumes perfectly if interrupted. Run npx torlnk to start downloading your media completely server-free.

🆔 @hackernewsgithubprojects

📦 om-ai-lab/vlx-seek

VLX-Seek: Smarter Visual Localization for Edge AI

VLX-Seek is the fine-grained vision-language model that finally replaces fragile coordinate generation with intuitive region referencing for edge-side embodied systems. Instead of forcing models to output long, prone-to-error numeric bounding boxes, it treats candidate visual regions as addressable tokens. This allows the model to naturally select and refer to specific objects by ID during reasoning. By using a hybrid encoder to connect local details with semantic understanding, it delivers faster, more stable performance for tasks like open-vocabulary detection and object counting. This efficient approach keeps your visual AI grounded and accurate while maintaining high responsiveness for real-world interactions.

🆔 @hackernewsgithubprojects

📦 rennf93/roboco

RoboCo

Run a complete virtual software development company right from your own hardware. This system deploys a structured, self-hosted organization of twenty-five distinct AI agents—including developers, quality assurance testers, and project managers—who coordinate through a formal task lifecycle. You act as the human CEO, feeding requirements to a board that breaks them into subtasks, delegates them to specialized backend, frontend, or design cells, and runs the code through a strict git-based build, test, and documentation pipeline. It lets you orchestrate complex, multi-project development securely using a custom control panel, leaving you in sole control of what finally merges.

🆔 @hackernewsgithubprojects

📦 seeed-projects/ai_robotics_academy

Master Robotics Engineering with the Seeed AI Robotics Academy

Seeed AI Robotics Academy is the comprehensive learning portal that finally bridges the gap between theoretical robotics and hands-on hardware implementation. This repository offers a structured curriculum covering everything from ROS1 and ROS2 basics to advanced simulation environments like NVIDIA Isaac Sim and MuJoCo. It features dedicated modules for mastering wheeled-legged robot dynamics, including PID-based pitch self-balancing and terrain adaptation, alongside practical tutorials for the ReachyMini desktop platform. Whether you are building kinematics models or deploying reinforcement learning, these resources provide the engineering workflows and deep technical insights needed to develop high-performance, autonomous robotic systems.

🆔 @hackernewsgithubprojects

📦 areshamohanad/bughunter-ai

Automate Penetration Testing with BugHunter-AI

Streamline your security assessments with BugHunter-AI, an automated penetration testing framework designed to manage complex scanning tasks while keeping your system stable. By leveraging a resource-aware task queue, the tool monitors CPU and RAM usage to execute security tools efficiently without causing system overload. The project features a futuristic cyberpunk-themed graphical interface where you can configure concurrent execution limits, track real-time tool progress, and view live system metrics. With its integrated AI analysis capabilities, it processes scan reports to suggest follow-up actions, ensuring a structured approach to your security testing workflow. Take control of your automated security audits today.

🆔 @hackernewsgithubprojects

📦 ethanfel/comfyui-krea2textencoder

Better Krea2 Prompts with Vision-Aware Encoding

ComfyUI-Krea2TextEncoder is the custom node that finally provides proper vision-aware conditioning for the Krea2 model. It solves issues with the default encoder by forcing the correct Krea2 descriptor template and removing unnecessary VAE inputs that the model ignores. This tool allows you to feed multiple reference images and optional masks into the Qwen3-VL-4B vision path to influence your generations effectively. It supports auto-growing input slots and configurable padding for precise image cropping, ensuring your prompts interact correctly with reference visuals. Upgrade your ComfyUI workflow to achieve more consistent and visually responsive output for your creative projects.

🆔 @hackernewsgithubprojects