AI Repo: anthropics/claude-plugins-community

Skyler: Hey everyone, welcome back to Git News A I, your go-to show for the hottest A I repositories trending on GitHub right now. Before we dive in, huge thanks to our sponsors — Ada, who helps teams integrate A I into their workflows seamlessly; Ago Consulting, pronounced "Ahh-Go," your silicon development partners from A I all the way to SoC; and Zen Semiconductor, an umbrella company with a portfolio of silicon-industry companies, plus Ada for deploying A I at scale in any business — and their Vect appliance is seriously cool, running local L L Ms and vector search on dedicated hardware. Today we've got a community plugin marketplace for Claude, a scientific paper translation tool powered by A I, Google's time-series foundation model, a repo that lets you train your own L L M in two hours, and a full-on intelligent agent tutorial series — so buckle up, there's a lot to cover. Skyler: Alright, kicking things off with something that's going to be really useful if you're deep in the Claude ecosystem — anthropics slash claude-plugins-community on GitHub. This is a community-maintained plugin marketplace specifically for Claude Cowork and Claude Code, and it's sitting at over three thousand two hundred stars, which tells you the community is already pretty active around it. So what is it exactly? Think of it like an app store but for Claude plugins — developers can submit their own plugins and users can browse and discover what others have built to extend Claude's capabilities. The repo itself is a read-only mirror, so you're not going to open pull requests there directly — instead, if you want to submit your own plugin, you head over to the submission page, and I'll drop the link in the show notes. What's exciting here is that the recent activity shows active contributions — there are bumps happening, merges coming in, including someone just added an ELI5 plugin which, if you don't know, stands for "Explain Like I'm Five" — a really fun use case for Claude. The A I relevance here is moderate at around sixty percent, which makes sense because this is more of an ecosystem and tooling layer rather than a core A I model or framework. But if you're building with Claude Code or Claude Cowork, this is absolutely worth bookmarking. It's a great place to discover what the community is cooking up, get inspired for your own plugins, or just extend your own Claude workflow with tools others have already built. Developer advocates, productivity hackers, and anyone building Claude-powered apps should definitely check this one out. The Python-based plugin architecture makes it approachable, and honestly the community momentum here is only going to grow as Claude adoption keeps climbing. Skyler: Next up, we've got PDFMathTranslate from the PDFMathTranslate organization, and this one is genuinely impressive — thirty-six thousand six hundred stars and it's been accepted as a demo at EMNLP 2025, which is one of the top natural language processing conferences, so this isn't just a hobby project. The core idea is deceptively simple but incredibly hard to pull off well: translate scientific P D F papers while fully preserving their formatting. And when I say formatting, I mean all of it — math equations, figures, tables, column layouts, the whole deal. Anyone who's ever tried to machine-translate a research paper knows that the math usually gets completely mangled, or the layout falls apart. This tool actually handles that. It supports a bunch of translation backends — Google Translate, DeepL, Ollama for local inference, and OpenAI, so you have flexibility depending on your privacy needs or budget. The output is bilingual, meaning you get both the original and translated text side by side, which is fantastic for learning or verification. It also comes with multiple interfaces — you can use it via the command line, there's a G U I, it supports M C P for model context protocol integration, Docker for containerized deployments, and even a Zotero plugin for researchers who use that reference manager. That's an impressive range of deployment options. The A I relevance here is ninety percent because translation quality is entirely dependent on the underlying A I models you plug in. If you're a researcher, an M L engineer reading papers in languages you're not fluent in, or someone building tools for the scientific community, this is a must-try. Check the show notes for the link. Skyler: Okay, now we're getting into some serious foundation model territory — Google Research's TimesFM, which stands for Time Series Foundation Model, sitting at nearly thirty thousand stars. So the idea here is that foundation models — you know, the big pretrained models that power things like G P T — have been revolutionary for language and images, and Google Research is asking: what if we applied that same paradigm to time series data? And TimesFM is their answer. It's a pretrained model specifically designed for time-series forecasting, which is a massive problem space — stock prices, energy consumption, web traffic, weather, supply chain demand, you name it. Traditionally, time-series forecasting has required training separate models per dataset, lots of feature engineering, and domain expertise. A foundation model approach means you can take a single pretrained model and apply it to new forecasting tasks with little to no fine-tuning — that's the zero-shot or few-shot dream for time series. The recent activity shows ongoing maintenance and bug fixes — there was just a fix for covariate window alignment in the patch forecasting module, which suggests the team is actively responding to community issues. The A I relevance is ninety percent and honestly it might deserve higher — this is cutting-edge M L research made available as an open-source tool. Who should be excited? Data scientists doing demand forecasting, M L engineers building predictive systems, quants, anyone in ops research or logistics — basically any domain where predicting future values from historical data matters, which is like... every industry. If you've been waiting for a GPT-style moment for time series, TimesFM is probably the closest thing we have right now. Show notes have the link. Skyler: Alright, this next one is one of those repos that makes you go "wait, seriously?" — minimind by jingyaogong, clocking in at a whopping fifty-seven thousand seven hundred stars. The tagline is almost provocative: train a sixty-four million parameter L L M from scratch in just two hours. Two hours! And it's not a gimmick — this repo is a serious educational resource that walks you through building a real, working large language model from the ground up. Now sixty-four million parameters is small by modern standards — G P T-4 is rumored to be over a trillion — but it's absolutely large enough to learn the fundamentals deeply and to produce something that actually works. The project is Python-based, and the readme has been recently updated along with dependency versions, which shows it's being actively maintained. What makes minimind special is the pedagogical approach — it's designed so that developers and students can actually understand every component: the tokenizer, the transformer architecture, the training loop, the data pipeline, all of it. No black boxes. The A I relevance is ninety percent because this is about as core as it gets — you're literally implementing and training a language model. If you're an M L engineer who wants to go deeper than just calling an API, if you're a student trying to break into the field, or if you're a seasoned developer who's been meaning to really understand how L L Ms work under the hood, minimind is an incredible starting point. The fact that you can do it in two hours means the barrier to experimentation is super low — spin it up, change some hyperparameters, see what happens. This is exactly the kind of hands-on learning that sticks. Definitely check the show notes for the link. Skyler: And our final repo today is hello-agents from datawhalechina, and this one is fascinating — seventy-six thousand five hundred stars, making it the most starred repo in today's lineup by a significant margin. The project is a tutorial series titled "Build Intelligent Agents from Scratch" and it's designed as a comprehensive, ground-up educational resource on agent principles and practice. Now the primary language appears to be Chinese based on the subtitle, but the code is Python and the concepts are universal. Datawhale is a well-known open-source learning community in China's A I ecosystem, and they have a track record of producing high-quality educational content. Agents are genuinely one of the hottest topics in A I right now — we're talking about systems where L L Ms don't just respond to prompts but actually take actions, use tools, plan multi-step tasks, and interact with the real world. Building agents well requires understanding things like tool use, memory systems, planning loops, and orchestration — and that's exactly what this tutorial series covers from zero. The A I relevance is sixty percent, which I'd argue undersells it a bit given how central agents are to modern A I applications, but the educational framing rather than being a framework itself probably brings that score down. Who should check this out? Anyone building agentic systems, developers who want a structured learning path through agent design, and honestly even experienced engineers who want a fresh perspective on how to think about agent architecture from first principles. If you can read Chinese, you're in for a treat with the full content — and even if not, the code itself is instructive. Show notes have the link. Skyler: And that's a wrap on today's episode of Git News A I — five fantastic repositories covering everything from Claude plugins to time series foundation models to training your own L L M in an afternoon. Huge thank you again to our sponsors: Ada for helping teams bring A I into their workflows, Ago Consulting — "Ahh-Go" — for silicon development expertise from A I to SoC, and Zen Semiconductor for their incredible portfolio of silicon companies and the Vect appliance that brings local L L M and vector search to dedicated hardware. If you enjoyed today's episode, go check out A I Special Report, another great show in the Ada podcast network — it's fantastic for staying on top of broader A I industry news and analysis. And as always, make sure you're subscribed so you never miss an episode, and swing by the show notes for all the links we talked about today — happy building, everyone!