📦 harbor-framework/terminal-bench

Terminal-Bench: The Ultimate Test for AI Agents

Every major AI agent builder is using Terminal-Bench to track progress and compare capabilities because it measures the frontier of agent work with a diverse, difficult, and high-quality set of tasks that evolve over time. It is a continuous benchmark published on the Harbor Hub, allowing developers to test their agents and models against real-world challenges in a sandboxing environment. The project actively seeks contributors to add new, challenging tasks, ensuring the benchmark remains a rigorous standard for evaluating AI performance.

🆔 @hackernewsgithubprojects

📦 bit-datalab/edit-banana

Turn Any Screenshot Into an Editable Diagram

Stop redrawing diagrams from scratch. Edit Banana takes a flat image of a flowchart or technical schematic and turns it into a live, editable file where every box, line, and label is separate. It uses smart vision models to find the shapes and text, then rebuilds the whole thing in a professional diagramming tool. No more tracing over PDFs. If you have ever had to rebuild a diagram because someone sent you a screenshot, this tool fixes that exact problem. It saves hours of tedious redrawing by doing the heavy lifting for you.

🆔 @hackernewsgithubprojects

📦 limecloud/lime

Lime: The AI Agent That Actually Finishes

Lime is the open-source AI agent that actually finishes the work. It is not just a chat box; it is a full-stack partner that lives in your desktop app and terminal, ready to handle real engineering tasks. While other tools stop at suggestions, Lime dives into your files, runs your tests, and executes terminal commands to build, fix, or research what you need. It can even split complex jobs across multiple agents, keeping every step visible so you can review exactly what it did. You get a powerful assistant that respects your permissions while delivering inspectable results.

🆔 @hackernewsgithubprojects

📦 xykong/flux-markdown

FluxMarkdown: The Finder Preview That Does It All

FluxMarkdown is the QuickLook extension that finally makes your Finder previews look like a polished website. Instead of seeing raw text when you press space on a document, this tool instantly renders diagrams, complex math, and interactive charts right inside the preview window. It turns a simple file peek into a full reading experience, complete with a clickable table of contents and the ability to save your work as a PDF. If you live in Markdown files, this is the upgrade you have been waiting for to make browsing your notes actually pleasant.

🆔 @hackernewsgithubprojects

📦 nvlabs/aspire

ASPIRE

ASPIRE solves a Rubik's Cube with a real robot arm, and it does not use traditional neural network weights to learn how. Instead, this system treats the robot's skills like a library of code that an artificial intelligence agent can read, test, and improve over time. The agent acts like a curious developer, running physical experiments to figure out how to grip objects or move joints, then saving the successful movements as reusable skills. This approach lets the robot adapt to new tasks without needing to be retrained on massive datasets. It is a fascinating shift from heavy computation to practical, iterative problem solving.

📰 https://news.ycombinator.com/item?id=50123746

🆔 @hackernewsgithubprojects

📦 maxmiksa/autonomous-company

autonomous-company

Spin up a digital firm that actually builds things while you sleep. Autonomous Company runs a continuous loop on your own machine, spinning up fourteen distinct AI agents to handle the heavy lifting. Each agent adopts a specific expert mindset, like a CEO using Bezos logic or a CTO focusing on failure-proof architecture. They research, code, and critique their own work, all while you just watch a local dashboard update in real time. It is the closest thing to hiring a remote team that costs nothing but your electricity. You get a self-managing workforce that works around the clock without the usual corporate bloat.

🆔 @hackernewsgithubprojects

📦 jordyzomer/lemmalog

Lemmalog: The Memory Engine for AI Agents

Lemmalog is the deductive database that turns agent memory into a verifiable, logical model rather than a fuzzy collection of text snippets. It solves the problem of AI agents forgetting or misremembering by using a logic engine to track exactly where every fact came from and how it changes over time. Instead of just storing text, it builds a structured knowledge graph that can prove its own conclusions and update efficiently as new information arrives. This means your AI assistant stops guessing and starts reasoning with confidence, giving it a reliable brain that remembers the truth, not just the last thing it read.

🆔 @hackernewsgithubprojects

📦 samuelfaj/distill

Distill: The Token Saver for Coders

Distill is a lightweight terminal tool that cuts the cost of using AI coding assistants by up to eighty-seven percent while keeping the same level of quality. It works by distilling your code context, so the AI only sees the most relevant parts of your project instead of every single file. This means you can connect your existing Grok or Codex subscription and get far more work done for far less money. It is built in Rust for speed and runs right in your terminal, making it a practical upgrade for anyone tired of high token bills.

🆔 @hackernewsgithubprojects

📦 satypardus/wowdecomp

They Rewrote World of Warcraft

Take a twenty-year-old World of Warcraft client, strip away the compiled machine code, and rewrite it all in human-readable C++ so anyone can actually read the engine. This is exactly what wowdecomp does. It reconstructs nearly every single function from the original binary, tagging each one with its exact memory address so you can verify it matches the source. The result is a complete, playable game client where the internal logic is finally open for inspection. Instead of guessing how the code works, you can just look at it. It turns a closed black box into a transparent piece of software that developers can actually study and understand.

🆔 @hackernewsgithubprojects

📦 anujasid/mcp-for-blender-pro

MCP for Blender Pro: AI that builds your 3D scenes

MCP for Blender Pro lets an artificial intelligence directly manipulate your three-dimensional workspace inside Blender by looking at your scene and writing the necessary code for you. This tool solves the problem of manually scripting repetitive modeling tasks by letting you describe your goal in plain language while the assistant handles the technical construction. It is useful for artists who want to quickly prototype layouts without leaving their creative flow or writing complex scripts from scratch. You can even set up a team of agents where one plans the work, others build the objects, and a final reviewer checks the result before you approve the changes.

🆔 @hackernewsgithubprojects

📦 victorchen96/deepseek_v4_rolepaly_instruct

DeepSeek V4 Roleplay Instruct: How to Control AI Thoughts

Take direct control over how DeepSeek V4 thinks inside its roleplay mode. This repository gives you the exact prompt codes to paste into your first message to force the AI into one of two distinct modes. You can make it act like a method actor, complete with internal monologues and emotional reactions, or switch it to a cold, pure logic mode for structured analysis. It is a simple tweak that fixes the messy, inconsistent thinking you usually get out of the box. Just copy the provided text, hit send, and watch the AI's thought process change completely.

🆔 @hackernewsgithubprojects

📦 tgp482/black-xberecomp

Black: A 2006 Xbox Game Gets Modern PC Life

A 2006 Xbox shooter now runs natively on your PC with modern graphics, thanks to Black XBE Recomp. This project uses a special tool to recompile the original game code, meaning it is not an emulator running old hardware but a fresh, native program built for Windows. The standout feature is the ability to remove the old thirty frame per second cap, letting the game run at your monitor’s full speed with crisp high definition visuals. It also adds proper mouse and keyboard support, so you can play with precision instead of relying on a game controller.

🆔 @hackernewsgithubprojects

📦 lucasrosati/claude-code-memory-setup

Claude Code Memory Setup

A coding assistant can now remember everything you did yesterday without you repeating yourself. This repository turns Claude Code into an agent with long-term memory by connecting it to a local knowledge system. It solves the frustrating problem of amnesia where you have to re explain your entire project every time you open a new chat. By mapping your code structure into a queryable graph, the assistant finds answers instantly instead of scanning files. This setup saves significant time and reduces token usage by relying on persistent notes rather than temporary context. It transforms a forgetful tool into a collaborative partner that retains project history.

🆔 @hackernewsgithubprojects

📦 michelhelsdingen/ensemble

Ensemble: AI Agents That Actually Collaborate

Two different A.I. coding tools sit side by side in a single project, passing notes to each other until they solve a hard problem together. That is the core idea behind Ensemble. It treats artificial intelligence not as a solo worker, but as a team. One agent leads, another handles specific tasks, and they share their findings in real time. It is like having a junior developer and a senior reviewer working from the same desk. You can watch the entire conversation happen live in a split screen, making the hidden reasoning visible. This is useful because it shows how distinct models handle the same logic differently.

🆔 @hackernewsgithubprojects

📦 mistralai/timestral

Timestral Predicts Time Series Probabilistically

Forecast entire time series with a single model that handles multiple variables at once, giving you a full range of possible outcomes rather than just one number. This open source tool from Mistral AI runs on a plain CPU and lets you plug in extra data like known future events to sharpen the prediction. It stands out because it tops major public benchmarks without needing fine tuning, proving it understands patterns straight out of the box. You get a simple library that downloads pre trained weights and returns probability distributions for any horizon. If you need reliable forecasting without the complexity, this is the tool to try.

📰 https://news.ycombinator.com/item?id=50105524

🆔 @hackernewsgithubprojects

📦 chuiziyao/maimen-certification

MaiMen Certification Machine

Upload a random photo, and the MaiMen Certification Machine instantly analyzes it to assign you one of ten absurd McDonald's fan ranks, ranging from a humble ice cream cone to the ultimate Big Mac deity. It renders a personalized meme card featuring your unique rank and a witty tagline, while simultaneously pulling real nutritional data to recommend a specific meal combo. The project connects directly to McDonald's official digital services so you can claim genuine coupons with a single click, turning a silly social share into an actual discount. It is a playful hack for the 2026 developer contest that proves you can make serious data fun.

🆔 @hackernewsgithubprojects

📦 flybrainlabs/node

How Node Keeps Every Transaction Honest

Replay any transaction on the network with absolute certainty. This is the engine behind Flybrain Labs, handling everything from complex goal competition to private data archives. The coolest part? It acts like a digital time machine. It doesn't just store data; it reconstructs the exact state of the system at any moment. No more guessing what happened. You get a perfect, immutable replay of every action, backed by strict verification. It’s the backbone that keeps the whole platform transparent and secure. If you care about trust in distributed systems, this is the piece you need to see.

🆔 @hackernewsgithubprojects

📦 cdcailf/cjkfold

CJKFold: The Silent Glitch Fixing Chinese PDF Search

Text extracted from Chinese PDFs often contains invisible characters that look exactly like normal words but are actually different code points, causing search engines to fail silently. CJKFold is a lightweight Python tool that fixes this by identifying and replacing these look-alike glyphs with their standard equivalents, ensuring your text matches correctly. It also includes a command-line scanner that diagnoses these hidden issues before they break your data pipeline. If you work with Chinese documents, this tool solves a frustrating problem that standard normalization misses, making your text processing reliable and accurate again.

🆔 @hackernewsgithubprojects

📦 albertoarena/laravel-truss

Laravel Truss: Your Database's Live Map

Laravel Truss turns your live database into a scrollable, zoomable map of how everything connects, without ever touching your actual data. It scans your table structure and renders it as an interactive diagram right in your browser, so you can see relationships at a glance instead of guessing. The coolest part is the focus mode, which lets you spotlight any single table and instantly see exactly which other tables link to it through foreign keys. It is perfect for understanding complex schemas without opening a heavy database client. It reads only the structure, keeping your data private while giving you a clear, visual picture of your database health and connections.

🆔 @hackernewsgithubprojects

📦 yoursjeyo/claude-switcher

Claude Switcher: The IP Leak Shield

Stop letting your real IP address slip through the cracks. Claude Switcher is a fail-closed environment manager that physically prevents your identity from leaking. It runs an eleven-point fingerprint audit and includes a hand-written WebRTC probe to catch hidden connections before they expose you. The standout feature is its kill-switch, which instantly cuts your network if something looks wrong. It’s like a bouncer who checks every ID and kicks you out if the vibe is off. This tool ensures your digital footprint stays exactly where you want it, no more accidental identity slips. If you care about privacy, this is your new safety net. Simple, strict, and effective.

🆔 @hackernewsgithubprojects

📦 openmoss/moss-image

MOSS-Image: The Text-to-Image Model That Actually Spells

Draw images where the text is actually spelled correctly. Most AI generators mangle words into gibberish, but MOSS-Image fixes that. It is a compact open-source model that renders crisp, legible characters directly into your pictures. You can even paste a simple layout file to tell it exactly where to place every title and object. It handles long prompts and looks great in anime or realistic styles without needing a supercomputer. If you need clear text in your visuals, this is the tool to use.

🆔 @hackernewsgithubprojects