📦 zjunlp/mechanist

Mechanist: The AI Researcher That Proves Itself

Run a full scientific investigation into how large language models actually think, without writing a single line of code. Mechanist is an autonomous agent that takes a vague research question and turns it into a verified, evidence-backed discovery. It hunts through thousands of academic papers, formulates a specific hypothesis, writes and executes the necessary experiments on your hardware, and then cross-checks its own conclusions with an independent model to ensure the results are solid. It handles the entire messy workflow from literature review to final verification, acting like a tireless lab assistant that never gets tired of checking its own work.

🆔 @hackernewsgithubprojects

📦 ucsandman/discovery-loop

How Discovery Loop Broke World Records for $28

Discovery Loop recently broke ten world records in the circle packing problem for under twenty-eight dollars. It is a local research lab that uses your existing AI subscriptions to evolve optimization solvers. Instead of trusting the models blindly, it runs them in isolated environments, demands independent mathematical checks, and requires human approval before anything counts. This rigorous loop found better solutions than previous human experts for specific packing benchmarks. It proves you can run serious scientific research from your own machine without massive API bills. If you want to build AI systems that actually verify their work, this project is the blueprint.

🆔 @hackernewsgithubprojects

📦 alibabaresearch/cut-craft-bench

CutCraft: The Video Editing Benchmark

Current AI video generators look cinematic, but they completely ignore how real film editors actually cut scenes. CutCraft is the first benchmark built specifically to test if these models understand professional editing techniques like J-cuts and transition timing. It stops relying on simple visual checks and instead measures if a model can execute specific editorial instructions, like keeping audio and video cuts in sync or building a proper montage. The project also includes an agentic baseline that breaks generation down into planning, shot synthesis, and composition to actually realize these edits.

🆔 @hackernewsgithubprojects

📦 roy-tong/agentmeasure

AgentMeasure: The Truth About AI Agent Metrics

AgentMeasure treats missing data as a valid answer rather than a hidden zero. As artificial intelligence agents start handling real work, their success rates often look inflated because a single failed task might be counted as two separate attempts. This open source standard fixes that by distinguishing between one logical goal and the multiple technical tries required to finish it. It helps you see exactly where your software fails and retries, so you can trust your numbers before customers pay for the outcome. This project turns vague activity into clear, verifiable facts.

🆔 @hackernewsgithubprojects

📦 penpot/penpot-ai-kit

Penpot AI Kit: Let AI Design Real Files

Penpot AI Kit lets you tell an AI assistant to build real, editable designs directly inside your Penpot file, rather than just generating flat images. You simply ask for a dashboard or a button component, and the kit guides the assistant to use your existing colors and shapes. The coolest part is that the AI grades its own work for quality and accessibility before showing you the result. It pauses to ask permission before making big changes, so you stay in control. This is perfect for designers who want to speed up their workflow without losing the polish of a human touch.

🆔 @hackernewsgithubprojects

📦 blacktwist/social-media-skills

Social Media Skills: Your AI Content Brain

Social Media Skills is the AI toolkit that turns your creative chaos into a clear, data-backed plan. Instead of guessing what posts will land, this collection of instructions teaches your AI assistant to analyze your past performance and identify exactly which topics and formats are driving real growth. It helps you see the difference between content that just gets likes and content that actually builds an audience. By connecting your creative output to hard follower data, it removes the guesswork from your content strategy. You get a personalized roadmap based on your specific numbers, making every post you write a calculated step toward your goals.

🆔 @hackernewsgithubprojects

📦 iml1s/resume-skills

Resume Skills: Your Coding Agent Memory Bank

Your coding assistant can forget its own work history, but resume skills fixes this by reading local conversation logs and packaging them into a safe, offline summary. This tool copies that context into a brand-new session so your agent remembers exactly what you were building. It works entirely on your computer without sending data to the cloud or restarting your old chat. The biggest benefit is that it supports almost every major coding assistant, letting you switch between tools without losing your place. It is not a magic save file that reboots a dead process, but a clean handoff of notes.

🆔 @hackernewsgithubprojects

📦 databricks-solutions/consort

Consort: The Music Ensemble For AI Code

Stop letting your AI code generator guess. Consort fixes this by acting like a music ensemble where every agent plays one specific instrument. It uses a strict conductor to enforce real tests against live database branches, so the code is actually correct. This means you get genuine engineering discipline instead of just lucky prompts. The result is a clean, verified application backend. Trust the music, not the model.

🆔 @hackernewsgithubprojects

📦 rohasnagpal/legal-ai-skills

Rohas Legal AI: 160+ Skills in One Click

Rohas Legal AI is the all-in-one toolkit that turns any chatbot into a legal assistant. It bundles over one hundred and sixty specific skills, covering everything from drafting contracts to analyzing arbitration awards, all wrapped in a single easy-to-install plugin. The coolest part is how it works: you simply describe your legal task in plain everyday language, and the system automatically picks the right skill from the library to handle it. You never have to remember a specific command or search for the right tool. It is built for lawyers, law students, and business owners who want accurate, structured legal help without the complexity of managing individual prompts.

🆔 @hackernewsgithubprojects

📦 quome-cloud/openworld

openworld

Building a working model of a complex system usually requires massive datasets, but this project claims you can do it in under two minutes. Openworld functions like a universal toolkit for creating these internal simulations, allowing you to test how code behaves before you ever deploy it. It solves the problem of needing expensive training data by constructing a logical map of the environment from the rules alone. You essentially build a safe, digital playground where you can run your software through hundreds of different scenarios instantly. This is useful because it turns guesswork into a verifiable process, letting you catch errors before they break your production systems.

🆔 @hackernewsgithubprojects

📦 jingxiao10/gua

GUA: The Missing Link in Neural Network Training

Here is the reality: when you train these networks, the optimizer can secretly ruin your carefully calculated steps. The project called GUA fixes this. Think of it like this: your navigation app tells you to turn left, but your car’s steering wheel is jammed, so you turn right anyway. GUA is the mechanic who checks the steering wheel before you start driving. It ensures the actual movement matches your plan, even when the math gets messy. This matters because physics problems are tricky, and small errors add up fast. By keeping things aligned, GUA helps these models learn correctly.

🆔 @hackernewsgithubprojects

📦 surge-ai/gdp-pdf

gdp-pdf: The Benchmark That Actually Reads PDFs

GDP pdf is the evaluation tool that measures whether large language models can genuinely understand complex documents. Built by Surge AI, this harness tests AI against a hundred real world tasks where models must read attached PDFs and answer specific questions without using any external search tools. The core strength here is strict, automated grading. Instead of relying on vague human opinions, the system uses another AI as a judge to check every single answer against a detailed rubric. This makes it incredibly useful for developers who need to know if their model is actually reading the data or just guessing.

🆔 @hackernewsgithubprojects

📦 otodock/oto-dock

OtoDock: Your AI Agent Company

OtoDock lets you build a self-hosted company of AI agents that answer real phone calls. It’s not just a chatbot; it’s a full operating system for your digital employees. You pick your own AI engine, like Claude or Codex, and plug it into this platform running on your server. These agents have persistent memory, can access your files, and even run while you sleep. The coolest part? They can literally pick up the phone. If you want your AI to feel like a real teammate who handles tasks autonomously and securely, this is the blueprint you need.

📰 https://news.ycombinator.com/item?id=49630606

🆔 @hackernewsgithubprojects

📦 yicheng47/runner

Runner: The App That Makes AI Coding Agents Work Like a Team

Runner lets you run multiple AI coding agents side by side inside a single native desktop app, turning scattered terminal windows into a coordinated crew. Instead of juggling separate sessions, you assign specific roles to each agent, like a coder and a reviewer, and they communicate through a shared event log to work on a shared goal. The most surprising part is that you can set up one agent to manage other agents, creating a self-organizing team that handles the entire workflow from planning to code review. It works locally on your machine, so every session is a real, persistent process you can watch and control.

📰 https://news.ycombinator.com/item?id=48236283

🆔 @hackernewsgithubprojects

📦 deepseek-ai/deepjit

DeepJIT: The GPU Compiler That Caches Everything

DeepJIT is the lightweight C++ library that lets you compile GPU kernels right at runtime without the usual boilerplate. It handles the messy details of finding your compiler, caching the results, and loading the code onto the device so you can just focus on writing the math. The coolest part is its shared cache system, which lets your entire team or cluster reuse compiled kernels from a single folder, meaning nobody has to wait around for slow compilation if the code hasn’t changed. It works seamlessly with both NVIDIA GPUs and Huawei Ascend chips, giving you a uniform way to handle device code.

📰 https://news.ycombinator.com/item?id=49641406

🆔 @hackernewsgithubprojects

📦 dicklesworthstone/franken_markdown

franken_markdown

Turn your Markdown into polished, self-contained HTML and tiny, optimized PDFs with just one single command. Franken Markdown is a pure Rust tool that does the heavy lifting for you, handling everything from syntax highlighting to font embedding without needing a browser or a massive software stack. The coolest part is that it generates real vector graphics for your diagrams directly in the PDF, so your charts stay perfectly crisp instead of looking like blurry screenshots. It is built to be fast and reliable, giving you professional looking documents from a simple text file.

📰 https://news.ycombinator.com/item?id=48826265

🆔 @hackernewsgithubprojects

📦 akitaonrails/ai-usagebar

The AI Usage Tracker That Actually Works

ai-usagebar is the desktop widget that finally tells you exactly how much you have left on your AI subscriptions. It was built because constantly checking your browser to see if you hit your rate limit is annoying, so it just grabs that info and puts it right on your screen. The coolest part is it works with almost everything, like Claude, ChatGPT, and Cursor, all in one place. It even has a timer that counts down when your quota resets, so you stop guessing. It is lightweight and built to be fast.

🆔 @hackernewsgithubprojects

📦 yilewang/llm-for-zotero

Your Zotero Library, But Make It An Agent

A browser extension just gave your personal research library a brain. L L M for Z o t e r o is an open-source research agent that lives directly inside your Z o t e r o library, reading your papers, notes, and PDFs to actually help you do research instead of just storing it. The most striking part is that it doesn't just chat; it acts. You can ask it to audit your library for missing metadata, auto-tag hundreds of papers, organize unfiled items, or even write and save notes back into your collection.

🆔 @hackernewsgithubprojects

📦 remysucre/prela

Prela: The Query Language You Control

Prela is the embedded query language that gives you total control over how data is processed. It works right inside your existing code, letting you mix queries with regular logic without switching contexts. The coolest part? It runs exactly as you write it, skipping the hidden optimizer that usually slows things down. That means you decide the speed, not a black box. It is fast, compact, and surprisingly simple to use. If you want to know exactly what your database is doing, this is the tool to check out.

📰 https://news.ycombinator.com/item?id=48356563

🆔 @hackernewsgithubprojects

📦 typstify/typstify

Typstify: The Desktop Editor for Typst

Edit professional documents right on your desktop with Typstify. This cross-platform tool gives you the polished power of LaTeX but swaps out the confusing, old-school syntax for a modern, intuitive interface. Built in Go and powered by the Gio UI framework, it runs smoothly on Windows, macOS, and Linux. You just need to drop in the Typst executables, and you are ready to go. It is not just a text editor; it is a full development environment that handles syntax highlighting, live previews, and file management out of the box.

📰 https://news.ycombinator.com/item?id=49127765

🆔 @hackernewsgithubprojects