Ever been told your 4GB GPU can’t run anything serious? Yeah, me too. But every local-LLM guide says “you need at least 16GB VRAM for a 7B model,” and my RTX 3050 laptop is sitting here calling that a lie. Turns out the model card might be wrong — not the hardware. So AirLLM (25.2k stars, Apache-2.0) runs a full 70B Llama on a single 4GB card, no quantization tricks, and it crossed the Trending list again this week. ...
Pointing a model at a whole repository and saying “find the bugs” rarely works. I’ve tried. The findings come back vague, duplicated, and unranked — a pile of “potential XSS here” with no way to know what’s real. open·kritt is the open-source answer to that, and the team behind it has $1.5M in bug bounty payouts to prove the approach works. The project just hit 628 stars in under two weeks (created July 20, sitting at v1.2.0 when I pulled it). Yet it’s not another “AI code scanner” toy. It’s a self-hosted security research platform that turns a single vague instruction into dozens of small, well-defined agent tasks running in parallel — then de-duplicates, ranks, and validates what comes back. And because it ships as Docker Compose, you run it on infrastructure you own. That last part is why it earned a full review here. ...
Every agent tool I’ve reviewed on this site is single-player. You, your terminal, one agent. QM is the first one that made me stop and think: what if the whole team had an agent that actually works like a teammate instead of a glorified autocomplete on one person’s laptop? That framing alone is why it earned a full review here. Four days ago the YC software team open-sourced yc-software/qm, and it’s already sitting at 4,954 stars — roughly 1,200 a day. That’s not hype, that’s a release moving. So I pulled it, deployed it to my own VPS, plugged a Slack channel in, and let it triage my inbox. Here’s the honest rundown. And if you’re wondering whether it’s production-ready yet, read the 63 open issues before you answer that. ...
Every serious coding agent has the same dirty secret: idle it in the background and it chews through RAM like it’s free. I’ve had Claude Code, Cursor, and Codex all parked at once on my Ryzen workstation, and by the end of a long session my memory graph looked like a ski slope. jcode — 14,795 stars and climbing fast at +812 today — is a Rust harness that treats RAM like a budget instead of an unlimited line of credit. Still, the numbers it posts aren’t marketing fluff; I watched them happen live. ...
Three hours into a refactor session, Claude Code had rewritten half my project. It felt productive. And it also felt like watching money leave my account through a window I couldn’t see through. But the terminal kept churning out files, running tests, calling tools — ask me what that session actually cost, or which of those changes were even verified, and I’d have shrugged. That’s the exact gap agentacct (537 stars in its first week, MIT, Python) is built to close: a local-first dashboard that joins your agent’s token bill to its work log, on one screen, on your machine. ...
Every AI-agent ecosystem has the same bottleneck: your skills, MCP servers, and connected services live locked inside whichever agent you happened to open. Switch from Codex to Cursor and you start over. OpenWork wants to be the layer that travels with you. It’s an open-source desktop app — an alternative to Claude Cowork — but the part that got me is the MCP server, because that works in the agents you already have. ...
Ever watched your AI agent write 50 lines for something the browser already does natively? Yeah, me too. Three files, a utility module, a type definition, one dependency — all for an HTML date picker. So that moment when you realize you’re paying tokens for code your users’ browsers shipped five years ago? That’s the problem Ponytail solves. I stumbled across ponytail-improved three days ago on GitHub Trending. 564 stars. 188 stars per day. And that kind of velocity doesn’t come from another “yet another Claude Code skill.” It comes from scratching a real itch. Honestly, it delivers. ...
Disclosure: Some links below are affiliate links. If you sign up through them, I may earn a commission at no extra cost to you. Vultr — high-memory cloud GPU instances from $6/mo DigitalOcean — $200 credit for new users, great for GPU droplets So Kimi K3 dropped two days ago and instantly became the most exciting open-weight model release this month — 7,300+ stars, 2.78 trillion parameters in a Mixture-of-Experts architecture, multimodal. Problem: nobody has a cluster sitting around to serve it. Enter Deltafin, a research project that does exactly what its README says: “runs a Mixture-of-Experts model far larger than the machine it sits on.” I spent the evening digging through its codebase, and honestly? The ambition matches the execution. ...
Ever had an AI coding agent go rogue — read a file it shouldn’t, fire off an API call you didn’t approve, or quietly delete something important? Yeah, me too. That moment when you realize your Claude Code session silently touched /etc/passwd or your Codex agent issued a curl to an IP you’ve never seen before — you can’t prove it, you can’t stop it, and you definitely can’t go back and figure out what happened. At least, you couldn’t until last week. Perplexity AI just open-sourced numbat, and it changes the game. ...
Your AI agent works in silence. That’s the problem. I’ve been running coding agents for months — Claude Code, Codex, custom toolchains. But every interaction goes through a terminal: type a command, wait, read the response, type again. Yet it works, but it’s glued to a desk. Last week I had a stretch of manual work in my workshop — soldering a new sensor board — and I thought: why can’t my agent just talk to me while I work? ...