📦 frontier-harness-eval/eval

The 17x Cost Gap in Coding Agents

The same AI model can solve a coding task for one dollar or nearly twenty, depending entirely on the software wrapper used to run it. This benchmark, called Eval, tested a single AI model across nine different coding tools on thirty identical software problems. While the pass rates stayed surprisingly close, the cost difference was massive. The cheapest tool spent just over a dollar per solved task, while the most expensive spent nearly eighteen dollars for the same result. The repository publishes these public results and provides a workflow for anyone to test their own tool against these exact tasks.

🆔 @hackernewsgithubprojects

📦 qdrant/skills

Qdrant Skills for AI Agents

Qdrant released a toolkit that teaches AI coding agents how to act like senior database engineers. Instead of just reading code, these agents now possess specialized knowledge on complex topics like scaling, performance tuning, and migrating data between models. The most surprising part is that the system doesn't rely on static tutorials. It uses a live advisor that fetches the most current guidance directly from the official documentation every time you ask a question. This means your coding assistant always gives advice based on the latest best practices, rather than outdated training data.

🆔 @hackernewsgithubprojects

📦 zefan-cai/open-jev

Open-Jev: AI Decisions Without the Guesswork

This project skips the usual AI guessing game. Instead of watching a model invent an answer and then parsing that text, Open-Jev gives you actual numbers. You feed the model your context and your specific questions, and it spits out typed probabilities directly. It is a clean way to make decisions without the messy business of reading generated sentences. This repository is an open implementation of that idea, built on Qwen models. It handles everything from routing customer service tickets to playing simple games. The best part is you get the raw confidence scores right away. No more hoping the AI wrote the right words.

🆔 @hackernewsgithubprojects

📦 fstandhartinger/chat-seek-vscode

Chat Seek: Find Any AI Chat in Seconds

Chat Seek is the VS Code extension that finally lets you find any past AI conversation using plain English. You know that one time you asked Codex to fix a tricky bug, or when Claude helped you draft a complex email? Finding it again is usually a nightmare. Chat Seek solves this by indexing your local chat histories from Claude Code, Codex, and OpenCode right inside your editor. The coolest part? It uses a local AI model to rerank your results so the most relevant chat jumps to the top immediately. You can even resume the session with one click, dropping you right back into that conversation.

🆔 @hackernewsgithubprojects

📦 pyd021226/fly-paradise

Fly Paradise

Fly Paradise puts a tiny fruit fly colony directly on your Windows desktop, letting them land on your icons and scatter when you move your mouse. The project offers two modes, one where you raise a limited group of flies to a specific goal, and another where they multiply without limit, designed specifically to start up automatically on your friend’s computer. It uses a transparent layer to track your screen and make the flies interact with your real desktop environment. This digital pet turns your screen into a living habitat where you can swat, wash, or care for the insects.

📰 https://news.ycombinator.com/item?id=49762770

🆔 @hackernewsgithubprojects

📦 pymc-labs/daimon

Daimon: Your Chat-Based Data Scientist

Drop a CSV into Slack, ask Daimon if your new pricing strategy actually helped, and watch it write the code, run the analysis, and post a clear chart right back into the thread. It is an open-source data science agent that lives inside your team's chat apps, handling everything from messy data cleaning to forecasting future trends. The coolest part is that it does not just give you a number; it provides a fully runnable notebook so you can verify every step yourself. No more copying and pasting between spreadsheet tabs and coding editors.

🆔 @hackernewsgithubprojects

📦 tarun-032/mellow

Mellow: The Pixel Dog That Actually Helps

Mellow is a pixel-art dog that lives on your Windows desktop and acts as a hands-free assistant. You can hold a key combination and talk to him out loud, and he will answer, read your screen, and even point to buttons with a little bone pointer. He also handles reminders, focus timers, and meeting notes, all without needing an internet connection if you choose local models. Whether you want a helpful tool or just a cute pet to keep you company, you decide how much intelligence he gets. It is a charming, open-source way to make your computer feel a bit more alive and responsive to your voice.

🆔 @hackernewsgithubprojects

📦 volotat/mini-agi

Mini AGI: The Laptop Brain That Never Forgets

A tiny neural network is training itself from absolute zero on a single eight gigabyte laptop graphics card. This project called Mini AGI solves a massive problem: standard AI models freeze forever after training, meaning you can never teach them new things without erasing old memories. Instead, this model works like a living brain. It stores its knowledge as ordinary files on your hard drive, loading only what it needs into memory. This means it can grow bigger by simply writing more files to disk, and it learns continuously from every new document you feed it.

📰 https://news.ycombinator.com/item?id=49783133

🆔 @hackernewsgithubprojects

📦 monteduro/killmyidea

Kill My Idea: The Brutal Truth Test

Type your startup idea and hit submit to get a brutally honest verdict from a digital judge named Jev. This tool cuts through the fluff by asking ten rapid-fire questions about your market, your money, and your actual execution plan. It skips the cheerleading and instead calculates a hard score to tell you if you should kill the idea, fix it, or finally ship it. It’s the reality check every founder secretly needs but rarely gets. If your idea survives the gauntlet, you’ve got something worth building. If not, you just saved yourself a year of wasted effort. Stop guessing and start building with confidence.

🆔 @hackernewsgithubprojects

📦 irfndi/prism-liquidity-agent

Prism Liquidity Agent

A bot on Solana now watches your trading pool every ten minutes to stop you from losing money while you sleep. Most people who put crypto into automated trading forget that the market moves, so their positions quietly lose value without them noticing. This project, called Prism, fixes that by constantly checking if your money is in the right spot. If the market drifts, it automatically shifts your position to keep earning fees instead of bleeding out. It even remembers past mistakes so it learns not to repeat them. It is basically a night watchman for your digital assets. No more waking up to check your phone.

🆔 @hackernewsgithubprojects

📦 0xnatoshi/jev-codex-router

Jev Codex Router

The Jev Codex Router is the routing tool that dynamically picks the right AI model for every single step of your coding session. Instead of forcing you to manually swap between expensive and cheap AI engines, this system automatically decides which model and thinking depth handles each specific task. It runs locally on your computer, quietly analyzing your code requests to ensure you get enough power to finish the job without wasting your monthly quota. It even includes a safety feature that keeps your code running if the main AI service ever has a hiccup.

🆔 @hackernewsgithubprojects

📦 jkudish/jev-mcp

Jev MCP: The Bouncer for Your AI Agent

Give your AI agent a built-in bouncer that actually reads the room. Jev MCP plugs into your coding assistant to run ten quick, cheap judgment calls. It screens fetched pages for sneaky prompt injections before they touch your context, verifies that every claim in a report matches the source, and even scores a code change to see if it is actually safe to merge. Instead of letting a massive, slow model guess, it uses a specialized model to return a typed verdict in under half a second for less than a cent. It is the difference between trusting a random intern and having a rigorous editor checking the work.

🆔 @hackernewsgithubprojects

📦 ghalebdweikat/winnow

Winnow: The Context Sieve for Claude Code

Winnow stops your coding assistant from wasting time reading irrelevant information. It acts as a smart filter that checks every large result from your code tools before the model sees it. A fast, specialized judge decides which parts of the file or command output actually matter for your current task. Anything deemed unnecessary gets swapped for a tiny summary, saving hundreds of tokens and keeping the AI focused. The full text is safely stored locally, so you can bring it back instantly if you change your mind.

🆔 @hackernewsgithubprojects

📦 ellipsis-dev/blink

Blink: Find Code by Asking Questions

Ask a plain English question to locate specific code files within a massive project. Blink uses an AI model to read directory names and assign likelihood scores to every path. It then releases hundreds of digital walkers that roam the file tree, with the most likely routes getting the biggest crowds. When a walker hits a file, it stops, and the final table shows which files captured the most traffic. This turns a chaotic search through nested folders into a simple probability report. It is a clever way to navigate complex codebases without manually guessing where to look.

🆔 @hackernewsgithubprojects

📦 skhanzad/selective-amnesia

Selective Amnesia: The Memory System That Forgets On Purpose

Selective Amnesia is the memory framework that teaches AI agents to forget on purpose. It builds a knowledge graph to track what a chatbot learns about you, but adds a smart pruning module that deletes less important details over time. This mimics how human brains work, keeping recent or critical facts while letting minor trivia fade away. By testing this against standard benchmarks, researchers show that this controlled loss of data actually helps the AI answer questions more accurately. It proves that for long-term digital memories, knowing when to drop information is just as vital as knowing what to keep.

🆔 @hackernewsgithubprojects

📦 modelscope/diffsynth-webui

DiffSynth-WebUI: Train AI Models in Your Browser

DiffSynth-WebUI runs the entire process of training custom AI models for images, video, and audio directly inside your web browser. Usually, fine-tuning these complex systems requires memorizing obscure command-line flags and manually editing text-based configuration files, which is a common source of errors and frustration for developers. This tool removes that technical barrier by replacing the terminal with a simple visual interface where you can upload your data, let AI automatically write descriptions for your images, and select training settings using clear menus. Because it tracks every choice you make, you can always reproduce your results later.

🆔 @hackernewsgithubprojects

📦 blackwell-systems/gcf

GCF: The Format That Speaks AI

GCF is the data format that finally lets AI read your info without the bloat. Instead of dumping giant blocks of JSON, it squeezes the same facts into a tiny, plain-text package that models understand perfectly. The real magic happens in long conversations. As your chat gets longer, GCF remembers what you already shared and only sends the tiny difference, dropping the cost by nearly one hundred percent. It is lossless, so nothing gets lost in translation, and it works with tools you already use. If you are building agents that talk a lot, this cuts your bills and speeds up every response.

🆔 @hackernewsgithubprojects

📦 quchangle1/turnsight

TurnSight: How AI Learns From Its Own Mistakes

TurnSight teaches AI models to learn from their own past mistakes without needing a human teacher. It works by looking back at a conversation where the AI already called a tool and saw the result. The system then compares what a smarter version of itself would have predicted with what the model actually did. By finding the gap between these two views, the model learns exactly where it went wrong in each specific step. This makes the AI much better at multi-step tasks involving external tools. It is a clever way to improve reasoning using only the data the model already generated.

🆔 @hackernewsgithubprojects

📦 tpyun/justdepth

JustDepth: Real-Time Radar Depth

JustDepth is the real-time radar depth tool that finally makes autonomous driving see clearly without expensive sensors. It takes a standard camera photo and a cheap automotive radar scan to instantly predict exactly how far away every object is. The magic happens because it skips the slow, messy steps usually needed to combine these signals. Instead, it compresses the radar data into a tiny, fixed-size format so the math stays lightning fast no matter how cluttered the road gets. This means your computer gets a complete depth map in under fifteen milliseconds.

🆔 @hackernewsgithubprojects

📦 snitro/pointer-cad-v2

Pointer CAD v2

Pointer CAD v2 is a new model from the recent computer vision conference that changes how computers draw technical designs. Instead of guessing shapes, it plans the construction steps first, using specific measurements to build precise parametric objects. The team says the actual code is still coming soon, so right now this is just the official announcement and paper details. It is interesting because it aims to make AI generated engineering drawings much more accurate and structured than before. This is a placeholder for a future tool, not a ready to use app.

🆔 @hackernewsgithubprojects