AI Repo: earendil-works/pi

Skyler: Hey everyone, welcome back to Git News AI — the show where we dig through the hottest trending A I repositories on GitHub so you don't have to. I'm Skyler, and before we dive in, huge shoutout to today's sponsors: Ada, helping teams integrate A I into their real-world workflows; Ago Consulting — that's "Ahh-Go" — your go-to for silicon development consulting from A I all the way down to SoC design; and Zen Semiconductor, an umbrella company with a whole portfolio of silicon-industry companies, plus they're behind Ada for deploying A I at scale in any business — and their Vect appliance? Local L L Ms and vector search running on dedicated hardware. Really cool stuff. Alright, today we've got a packed lineup — a slick A I agent toolkit, a massive collection of L L M apps, a self-driving product platform, a GPU kernel D S L, and a Codex plugin marketplace. Let's get into it. Skyler: First up, we've got a repo called pi from earendil-works, sitting at over a hundred and eleven thousand stars on GitHub — and yeah, that number is not a typo. So what is pi? It's an A I agent toolkit built in TypeScript, and it's doing a lot of heavy lifting in one package. You get a unified L L M A P I, an agent loop, a terminal U I — or T U I — and a coding agent C L I all baked in together. So instead of gluing together five different libraries to build an agent that can actually write code, pi gives you that whole stack in one coherent toolkit. Skyler: The unified L L M A P I piece is really the backbone here — it means you're not rewriting your integration every time you want to swap out models. Whether you're pointing at OpenAI, Anthropic, or something else, pi abstracts that away. And then the agent loop handles the back-and-forth reasoning cycle that makes A I agents actually useful — the observe, think, act pattern that's at the heart of most agentic systems. Skyler: The coding agent C L I is where things get really fun for developers. Imagine having an agent running in your terminal that can help you write, debug, and reason about code — kind of like having a very fast pair programmer living in your shell. And the T U I makes the whole experience feel polished rather than janky, which honestly matters more than people admit. Skyler: Recent activity on the repo is a little eclectic — there's a commit about playing a 3D animation in fullscreen mode on a specific route, which feels like the devs having a bit of fun, and there's a new unreleased section in the changelog which means more features are cooking. TypeScript developers and anyone building agentic tooling should absolutely check this one out. Links are in the show notes. Skyler: Next up — and this one is a beast — we have awesome-llm-apps from Shubhamsaboo, and it is sitting at over a hundred and forty thousand stars. That makes it one of the most starred A I repositories we've ever covered on this show. The pitch is simple: over a hundred A I agents, agent skills, and R A G apps, all free and open source. That's R A G as in retrieval-augmented generation, where you're feeding an L L M relevant context from your own data at query time. Skyler: What makes this repo special isn't just the quantity — it's the breadth. You've got agents for productivity, research, coding, data analysis, and more. And because everything is open source and written in Python, you can actually crack open any of these apps and learn how they're built. It's basically a living textbook for L L M application development. Skyler: If you're an M L engineer or data scientist trying to level up your agentic app skills, this is the resource. Instead of starting from a blank file wondering how to wire up a R A G pipeline or build a multi-step agent, you can just clone one of the examples and start experimenting. The learning curve on L L M app development is real, and repos like this are what flatten it. Skyler: There's no recent commit activity to report, which usually means a repo is either dormant or just very stable — given the star count and the breadth of content here, I'd guess it's the latter. This is a bookmarks-worthy repository. If you're building anything with L L Ms, check the show notes for the link. Skyler: Alright, number three — PostHog. Now, some of you might already know PostHog as a product analytics platform, but it has evolved significantly and now has a strong A I angle that makes it very relevant to this show. The repo has over forty thousand stars and it's written in Python. The tagline they're using now is "the leading platform for building self-driving products," and that framing tells you a lot. Skyler: What does self-driving products mean? Essentially, PostHog is building toward a world where your product can observe itself, diagnose its own problems, and even surface fixes — with A I doing a lot of that heavy lifting. They've added A I observability tooling which is huge right now. If you're shipping an A I-powered product and you want to understand how your L L M features are actually performing in production, that's the kind of visibility PostHog is going after. Skyler: Beyond that, you still get the full analytics suite — session replay, feature flags, A/B experiments, error tracking, and logs. So it's not just A I observability, it's a full developer intelligence platform. And the fact that it integrates with Slack, web, desktop, and M C P — that's the Model Context Protocol — means your A I agents can actually interact with your product data programmatically. Skyler: Recent commits show work on bulk feed row interactions and feature flag evaluation syncing behind a setting — these are the kinds of incremental reliability improvements that matter a lot in a platform at this scale. If you're building A I-powered products and you need to understand what's happening in production, PostHog is worth a serious look. Show notes have the link. Skyler: Okay, number four is one for the low-level performance nerds — and I mean that as the highest compliment. TileLang from tile-ai is a domain-specific language designed to streamline the development of high-performance kernels for G P Us, C P Us, and accelerators. Eight thousand stars and climbing, written in Python, and it's got some really interesting recent activity. Skyler: So what's a kernel in this context? When you're running A I workloads — matrix multiplications, attention mechanisms, convolutions — those operations need to run incredibly efficiently on hardware. Writing these kernels by hand in C U D A or other low-level languages is notoriously painful and error-prone. TileLang gives you a higher-level D S L — domain-specific language — that lets you express these computations more naturally while still squeezing out serious hardware performance. Skyler: What's exciting about the recent commits is the addition of a P T O backend for the Ascend 950. That's Huawei's A I accelerator chip, and seeing TileLang expand beyond Nvidia G P Us is a big deal. The A I hardware landscape is diversifying fast — you've got Nvidia, A M D, custom A S I Cs, and now Ascend — and tools that can target multiple backends without you rewriting everything are incredibly valuable. Skyler: They also just bumped to version 0.1.15, so this is an actively developed project with real momentum. If you're working on M L systems, inference optimization, or custom A I hardware, TileLang is the kind of tool that could genuinely change how you approach kernel development. This one's for the infrastructure engineers and M L systems folks — check the show notes for the link. Skyler: And our fifth repo today is awesome-codex-plugins from hashgraph-online — describing itself as the number one Codex marketplace and a curated list of OpenAI Codex and ChatGPT plugins, skills, and resources. About eleven hundred stars, written in Python, and it's essentially a community-maintained directory of plugins and integrations for Codex and ChatGPT. Skyler: If you're building on top of OpenAI's ecosystem — whether that's creating Codex plugins, ChatGPT integrations, or just trying to discover what's already out there — this repo gives you a structured starting point. The curation aspect is what I appreciate here. Anyone who's tried to discover useful A I plugins by just browsing around knows it can be chaotic. Having a maintained list with submission guidelines adds real signal to the noise. Skyler: The recent commits are all documentation updates — they've been refreshing the plugin scanner submission guide across a couple of patch versions. That tells me the maintainers are actively keeping the contribution process clean and accessible, which is a good sign for a community-driven resource like this. Skyler: It's a smaller star count than the others we covered today, but don't let that fool you — niche, well-curated resources like this can save you hours of discovery time. If you're in the Codex or ChatGPT plugin ecosystem, it's worth bookmarking. Link is in the show notes as always. Skyler: And that's a wrap on today's episode of Git News AI! Big thank you to our sponsors — Ada for helping businesses integrate A I into their workflows at scale, Ago Consulting — "Ahh-Go" — for world-class silicon development consulting from A I to SoC, and Zen Semiconductor for their incredible portfolio of silicon-industry companies and the Vect appliance bringing local L L Ms and vector search to dedicated hardware. If you want more great tech content, go check out AI Special Report — another fantastic show in the Ada podcast network that dives deep into the A I landscape. And hey, if you're getting value from this show, hit subscribe wherever you're listening, and check the show notes for links to every repo we covered today. We'll see you next time on Git News AI.