📦 bespokelabsai/nimble

Nimble: Fast AI Decisions Without Reasoning

Nimble is the local AI model that answers complex questions instantly without writing out any reasoning. Instead of generating a long paragraph to explain its thought process, it reads your text once and simply picks the correct answer from a list you provide, making it blazing fast for tasks like routing requests or checking conditions. The real trick lies in how the team trained it using contrastive data curation, where they create nearly identical examples that differ by just one fact to teach the model exactly which details matter.

🆔 @hackernewsgithubprojects

📦 webprodigies-org/flute

Turn Your React UI Into a 3D Movie

Transform your live React interface into a cinematic three-dimensional video using Flute by Web Prodigies. This open source tool lets your coding agent compose camera angles, depth, and motion directly from your existing components, solving the problem of creating high-end promotional assets without rebuilding your app in a separate 3D engine. You keep your original styles and data while exporting high frame rate MP4 files entirely locally. It is a fascinating way to visualize your actual product in motion. If you build React apps, this is a powerful way to make your work feel dynamic.

🆔 @hackernewsgithubprojects

📦 mah3sec/forgeguardian

ForgeGuardian: The Offline Supply Chain Scanner

ForgeGuardian scans your software dependencies entirely on your own machine, meaning your code never leaves your computer to get checked for threats. It solves the problem of trusting the thousands of external packages you download to build your apps, catching hidden malware and fake packages before they sneak into your final product. The coolest part is that it runs offline, so you do not need to upload your secrets to a cloud server to stay safe. It gives you full control over your security without sacrificing speed. You get peace of mind knowing exactly what is inside your code. Keep your data private and secure. Try it today.

📰 https://news.ycombinator.com/item?id=49574041

🆔 @hackernewsgithubprojects

📦 moritzkremb/jev-voice-browser

The browser that acts before you finish the sentence

The browser starts typing your search query before you even finish saying the name of the person you are looking for. This project, called Jev Voice Browser, connects a specialized AI model to a real web browser. The model does not write code or plan actions. It answers a series of simple yes or no questions in about three hundred milliseconds. If it is confident you are done speaking, the browser acts immediately. If you are still talking, it waits. The text you search for is never generated by the AI. It is cut directly from your spoken transcript. This means the system is fast, cheap, and highly reliable.

🆔 @hackernewsgithubprojects

📦 devmortimer/pi-warden

pi-warden: The AI Safety Net You Didn't Know You Needed

Pi-warden prevents AI coding agents from making irreversible mistakes by checking every action against your project rules before execution. Instead of interrupting your workflow with constant approval dialogs, it quietly judges if a command is safe or off-track, then nudges the agent to fix its own work. In one hundred fifty controlled tests, projects with this guard active broke zero rules, while those without it failed six times. It acts like a seatbelt, catching dangerous deletions or unverified claims without stopping your momentum. You keep working while the agent self-corrects.

🆔 @hackernewsgithubprojects

📦 unicodef1wn/grokbot-field-notes

Grok Bot Field Notes: 72 Hours of Agent Failures

Grok Bot Field Notes is the practical handbook that turns seventy-two hours of live software chaos into reusable rules for your own coding agents. Three engineers from the xAI Grok Bot team built a product from scratch on stream, documenting every failure, from broken layouts to security leaks, to create a clear set of house rules you can drop directly into your repository. The core insight here is simple: verification is the real job, not just writing code, and every agent needs a specific way to prove its work before it is finished.

🆔 @hackernewsgithubprojects

📦 romanslack/jev-drone

Jev Drone: How a Small AI Model Pilots a Quadcopter

Fly a real quadcopter through a gauntlet of obstacles using just its onboard camera. Jev Drone shows how a lightweight judgment model can make high-level tactical decisions while simple code handles the twitchy, fast-moving safety reflexes. The coolest part is that without this small model, the drone gets stuck at the first wall because it cannot figure out how to fly over it. With the model in the loop, the drone climbs over barriers, times sweeping gates, and even re-acquires its target after losing sight in a dense cluster of pillars.

🆔 @hackernewsgithubprojects

📦 nekuda-ai/windtunnel

WindTunnel: The Browser Agent Benchmark

WindTunnel settles the debate on how AIs should actually control your web browser. It pits the new WebMCP standard against older methods, like having an AI stare at screenshots and click pixel by pixel. The project runs real, open-source websites in isolated containers and throws forty-nine specific tasks at different AI models. It then checks the actual database to see if the job got done. The surprising result? Giving the AI direct commands is dramatically cheaper and faster than letting it guess by looking at images. If you are building agents, this data tells you exactly which path saves you money and time. Stop guessing. Test it.

🆔 @hackernewsgithubprojects

📦 coldteadotai/abide

Abide: The Rule Enforcer for AI Coders

Stop letting your AI coder ignore your strict project guidelines. Abide hooks directly into tools like Claude Code to audit every single line of code your agent writes. It uses a specialized decision model to check your specific rules, like 'use Yup for validation,' in just three hundred milliseconds. If the agent breaks a rule, it gets an instant, automatic prompt to fix the error before moving on. This tool turns your text-based instructions into hard, technical enforcement, catching mistakes that standard linters miss. It is the difference between hoping your AI follows instructions and actually making it obey them.

🆔 @hackernewsgithubprojects

📦 openai/tunnel-client

Tunnel Client: Local AI, Private Access

Tunnel Client is the secure bridge that lets private AI tools talk to ChatGPT without touching the internet. Most local AI servers are locked behind firewalls, making them invisible to cloud assistants. This tool fixes that by creating a hidden, encrypted tunnel from your private machine to OpenAI's servers. You can now use your own local data and code with ChatGPT or Codex while keeping everything safely offline. It is the perfect way to enjoy powerful AI features without compromising your data privacy. If you want to connect your local AI to the cloud, this is your safe, private path.

🆔 @hackernewsgithubprojects

📦 mcncarl/yichen-skills

Yichen Skills: The AI Toolkit for Creators

Yichen Skills is the skill collection that turns messy content creation into a clean, automated pipeline. Maintained by a serial entrepreneur named Yichen, this open-source toolkit connects AI agents to real-world tools like Obsidian, X, and WeChat. It handles the heavy lifting, such as uploading long-form articles directly to X drafts, exporting private WeChat archives into usable data, and slicing posts into ready-to-post video clips. Instead of manually copying text and formatting images, the system automates the entire workflow from writing to publishing. This is a powerful way to reclaim hours of manual work. If you create digital content, this repository is your shortcut to a smarter, faster daily routine.

🆔 @hackernewsgithubprojects

📦 penglonghuang/chinese-novelist-skill

chinese-novelist-skill

Generate a complete, multi-chapter Chinese novel from scratch. This tool solves the hardest part of writing, which is actually finishing the book. It starts by asking a few simple questions about your story idea, then automatically builds a plot outline and character profiles. The most surprising part is that it writes the entire book for you, checking every single chapter to make sure it is long enough and makes sense. If a chapter is weak, it automatically rewrites it until it is good. It is not just a text generator; it is a full production workflow that remembers your writing style for next time.

🆔 @hackernewsgithubprojects

📦 continuum-ai-corp/orcabonsai-27b-uncensored

OrcaBonsai-27B-Uncensored

OrcaBonsai-27B-Uncensored is the runtime ablation technique that removes safety filters from compressed language models without touching the original file. Traditional uncensoring tools alter the model's internal math, which breaks the heavy compression that makes these tiny models run on your phone. This project instead applies a correction at the exact moment the model generates each word, leaving the downloaded weights completely untouched and bit-identical. It is a fascinating look at how you can change a model's personality on the fly while keeping the data efficient. The takeaway is that behavior is just a runtime adjustment, not a permanent edit to the file itself.

🆔 @hackernewsgithubprojects

📦 razorback16/openjev

OpenJev: Fast AI Decisions

Get fast, reliable yes-no answers from an AI model without any generated text. OpenJev is an open-source decision server that reads probabilities directly from a diffusion model to answer typed questions in milliseconds. It solves the hallucination problem by never generating free text, so answers cannot drift off script. This is great for routing support tickets or scoring customer sentiment instantly. The coolest part is that it speaks the same API as TypeSafe Jev, so existing tools just work. You get calibrated confidence scores and support for images too. It is fast, open, and built to be swapped in easily.

🆔 @hackernewsgithubprojects

📦 trungdq88/youtube-sponsor-detection

Sponsor Skip

Skip the sponsor reads on YouTube before you even hear them. This project uses a specialized model to scan video captions, pinpoint exactly where a paid promotion starts and ends, and jump straight past it. It isn't just a guess; it listens to the audio in real time to confirm the transition, ensuring you never miss a beat or lose actual content. Whether you use the Chrome extension or the simple web app, it turns those annoying ad breaks into zero seconds of waiting. It is a clever way to reclaim your time and keep your watch sessions focused purely on the video you actually wanted to see.

🆔 @hackernewsgithubprojects

📦 jerryzliu/agentplayback

AgentPlayback: Visualize Your AI Coding Agents

AgentPlayback is the local dashboard that finally shows you exactly how much your AI coding agents are costing you. If you use tools like Claude Code or Codex, you probably have a lot of hidden logs that show what happened, but you likely have no idea how much you are actually spending. This project reads those private logs and turns them into a simple visual timeline, so you can see precisely when your agents were working versus when they were just waiting for you to type back.

📰 https://news.ycombinator.com/item?id=49456949

🆔 @hackernewsgithubprojects

📦 pleaseprompto/smabar

The Dock That Builds Itself

Pin your favorite apps, folders, and websites to a sleek dock that sits right next to your system taskbar. But the real magic is that you can ask your coding agent to build a custom plugin for you, like a live bus schedule or a crypto ticker. smabar connects to your AI assistant so it can write the code, test it, and drop it straight onto your screen without you touching a terminal. It’s like having a developer friend who just works for you. If you want a desktop that adapts to exactly what you need, this is the way to do it.

🆔 @hackernewsgithubprojects

📦 dominikbucko/qwen38-flash-next-2x3090

qwen38-flash-next-2x3090

You can now run a massive two-hundred-and-fifty-six-thousand token context window on just two consumer graphics cards at a blazing peak of over eighteen-hundred tokens per second. The qwen38-flash-next-2x3090 project accomplishes this feat by pairing a specialized model checkpoint with a highly clever memory offloading system. Instead of forcing the entire model into expensive graphics memory, it strategically keeps a massive n-gram table and less-frequent experts in system memory, while keeping a dynamic, running cache of active experts right on your GPUs. By using a smart drafting system to speed up generation, it gives you lightning-fast performance without sacrificing quality.

🆔 @hackernewsgithubprojects

📦 incoai/splash

How Splash Makes Your Mac a Supercomputer

Splash runs massive AI models directly on your Mac, beating cloud servers in speed without sending data away. On a standard M5 Pro, it processes text twice as fast as the next best engine, returning your first word in less than three hundred milliseconds. The secret is that every single component is custom built for the specific model you load, creating a perfect fit that eliminates waste. You get a local tool that speaks the same language as your favorite coding agents, meaning your private projects stay on your machine while working faster than you ever thought possible.

🆔 @hackernewsgithubprojects

📦 githubnext/localjev

Run Jev Locally: The Surprising Bridge

LocalJev lets you run TypeSafe Jev decision-making entirely on your own computer. While the standard Jev service relies on a remote, proprietary backend, this open-source project bridges that gap by translating typed decision requests into a format your local hardware can process. It connects directly to local inference servers running on Macs, allowing you to make complex, structured decisions without sending sensitive data to the cloud. By running the engine locally, you gain full control over privacy and latency while maintaining compatibility with existing SDKs. It is a powerful way to experiment with autonomous decision logic right on your desktop.

📰 https://news.ycombinator.com/item?id=49766255

🆔 @hackernewsgithubprojects

📦 mxxm-t/mx-llama.cpp

mx-llama.cpp: Multi-GPU Inference in C/C++

mx-llama.cpp is a specialized fork of the famous llama.cpp project, built specifically to wring maximum performance out of multi-GPU setups using pure C and C++. While the original handles single cards well, this fork introduces a clever multi-stage tensor parallelism feature. This lets you split your model's layers across different groups of GPUs, pipelining the data so multiple cards work on different parts of the sentence simultaneously. This is huge for people running massive models like DeepSeek or Qwen locally without paying for a data center. It also includes custom GPU communication code that beats standard drivers on certain hardware.

🆔 @hackernewsgithubprojects