AI Repo: microsoft/SkillOpt

Skyler: Hey everyone, welcome back to Git News A I — I'm Skyler, and this is the show where we dig into the hottest A I repositories blowing up on GitHub so you don't have to scroll for hours. Big thanks to today's sponsors: A-I-E-D-A, helping teams integrate A I into their workflows; Ago Consulting — pronounced "Ahh-Go" — your go-to for silicon development from A I all the way to SoC; and Zen Semiconductor, an A I and silicon venture group whose Vect appliance runs local L L Ms and vector search on dedicated hardware. Great lineup of sponsors, and an even better lineup of repos today. We've got a Microsoft project that teaches frozen L L M agents reusable skills, a blazing-fast local A I engine built specifically for Apple Silicon, a grammar checker built in Rust that never phones home, a compute platform that turns scattered cloud resources into one giant A I supercomputer, and an M C P toolkit that basically gives your coding agent an I D E. Let's get into it. Skyler: Alright, first up we have microsoft/SkillOpt, sitting at nearly fifteen thousand stars and climbing fast. So what is this thing? SkillOpt is a text-space optimizer — and I know that sounds a little abstract, so let me break it down. Imagine you have a frozen L L M agent, meaning you can't touch the model weights, you can't fine-tune it, nothing like that. The model is locked. But you still want that agent to get better at doing specific tasks over time. That's the problem SkillOpt is solving. Instead of touching the model, it trains reusable natural-language skills — basically human-readable instruction sets that the agent can draw on. The way it works is really clever. It uses trajectory-driven edits, so it watches how the agent performs on a task, analyzes the full trajectory of that run, and then figures out how to improve the skill text based on what went wrong or what could be better. Then before any update actually gets applied, it goes through validation-gated updates — so there's a quality check in the loop to make sure a new skill version is actually better before it gets promoted. And when a skill is ready for prime time, it gets saved as a best_skill.md artifact — a plain markdown file you can inspect, version-control, and deploy. That's a huge deal because it means your skills are auditable and portable. The recent activity shows they're actively patching things too — they just merged a fix that makes a session end installation step non-destructive, which is the kind of operational detail that matters when you're running agents in production. If you're building L L M agent systems and you want a systematic way to improve agent behavior without retraining, SkillOpt is absolutely worth your time. Go check the show notes for the link. Skyler: Next up, let's talk about raullenchai/Rapid-MLX — three thousand three hundred and sixty-five stars and the description alone got me excited. This is billed as the fastest local A I engine for Apple Silicon, and the numbers they're throwing around are wild. They claim four point two times faster than Ollama, and a cached time-to-first-token of just zero point zero eight seconds. If you've ever sat there waiting for a local model to start spitting out tokens, you know how painful that latency can feel, so zero point zero eight seconds cached is genuinely impressive. But speed isn't the only story here. Rapid-M L X also ships with one hundred percent tool calling support and seventeen — seventeen! — tool parsers out of the box. Tool calling is one of those features that separates a neat demo from a production-capable agent, so having that fully baked in matters a lot. They've also got prompt caching, reasoning separation — so you can cleanly split chain-of-thought reasoning from final output — and cloud routing if you need to offload to a remote model. And here's the part that really stands out to me: it's a drop-in OpenAI replacement. That means if you've got code already talking to the OpenAI A P I, you can point it at Rapid-M L X instead with minimal changes. It works with Claude Code, Cursor, and Aider right out of the box. Recent commits show they're fixing memory management on long prompts and cleaning up M C P server output, which tells me the team is very much in active polish mode. If you're on a Mac with Apple Silicon and you want the fastest local inference experience available right now, this one deserves a serious look. Show notes have the link. Skyler: Okay, shifting gears a little — let's talk about Automattic/harper. Twelve thousand three hundred stars, and yes, this is Automattic, the same folks behind WordPress. Harper is an offline, privacy-first grammar checker built in Rust, and I want to talk about why this lands in A I territory even though it's not an L L M. A lot of us are building A I-powered writing tools, coding assistants, or document processors, and one of the recurring headaches is grammar checking. Most grammar APIs are cloud-based, which means your text is leaving your system, which is a huge problem for enterprise users, sensitive documents, or anyone who just cares about privacy. Harper flips that on its head entirely — it runs completely offline, nothing leaves the machine, and because it's written in Rust it is fast. Like, really fast. There's no spinning up a model, no API call, no latency tax. It's rules-based with smart linguistic awareness, not a neural network, which means it's deterministic and lightweight. For developers building applications where you want grammar checking baked in without a cloud dependency, this is a fantastic option. It's open source, which means you can audit it, modify it, and ship it in your own products. The recent commits are kind of fun too — they just added a rule to correct "one's self" to "oneself" as one word, which is exactly the kind of careful linguistic detail work that makes a grammar checker trustworthy. If you're building anything involving text — and honestly, when are we not — Harper is worth throwing into your toolkit. Check the show notes for the link. Skyler: Alright, let's talk infrastructure for a minute because skypilot-org/skypilot is doing something I think is really important for the A I ecosystem. Ten thousand three hundred and seventy stars, and the pitch is straightforward but powerful: SkyPilot turns fragmented A I compute into one A I supercomputer. Here's the problem it's solving. If you're a serious A I team, you're probably juggling compute across multiple clouds — AWS here, GCP there, maybe some on-prem GPUs, maybe spot instances — and managing all of that is a nightmare. You've got different APIs, different pricing, different availability, and no unified view of what's running where. SkyPilot abstracts all of that away. You define your job, and SkyPilot figures out where to run it, optimizes for cost and availability, and handles the orchestration. It's positioned as a platform for frontier teams, and I think that framing is right — this is for teams doing serious model training, large-scale inference, or complex multi-job pipelines where manual compute management would eat your engineering hours alive. The recent activity is interesting too — they're actively working on a dashboard, and the latest commits are about things like dark mode favicons and making settings cards plugin-replaceable. That tells me they're investing in the operator experience, not just the core engine. If your team is burning time managing compute across clouds instead of building models, SkyPilot is exactly the kind of tool that can give you those hours back. Show notes have the link. Skyler: And last but absolutely not least — oraios/serena, sitting at a whopping twenty-six thousand eight hundred stars and it is the top performer in our lineup today. Serena is a powerful M C P toolkit for coding, and the tagline is "the I D E for your agent." I love that framing. So M C P stands for Model Context Protocol, which is the standard that's emerging for giving A I agents structured access to tools and context. What Serena does is provide semantic retrieval and editing capabilities on top of M C P, so instead of your coding agent just fumbling around with raw file system access, it gets intelligent, structured ways to understand and modify code. Think symbol-aware navigation, semantic search across your codebase, the kind of context a real developer has when they sit down to edit code — except now your agent has it. The recent commits show some really thoughtful feature work: they just added support for nixd language server paths and config, and they made TypeScript server progress start timing configurable, which shows they're sweating the details of real-world language server integration. This isn't a toy — it's being built to handle the messy reality of actual codebases with multiple languages and toolchains. If you're building coding agents, or if you're using agents like Claude Code or Cursor and you want to level up the context and editing capabilities, Serena is a must-explore. Twenty-six thousand stars don't lie. Check the show notes for the link. Skyler: And that's a wrap on today's episode of Git News A I! Huge thank you to our sponsors: A-I-E-D-A for helping teams bring A I into their workflows the right way, Ago Consulting — "Ahh-Go" — for world-class silicon development expertise from A I all the way down to SoC, and Zen Semiconductor for pushing the frontier on local A I hardware with their Vect appliance. If you're enjoying this show, make sure to check out the A I Special Report over in the A-I-E-D-A podcast network — it's a fantastic companion show for staying sharp on what's happening across the broader A I landscape. And of course, subscribe wherever you get your podcasts, drop us a review if you're feeling generous, and all the repo links from today's episode are waiting for you in the show notes. I'm Skyler — go build something cool, and I'll catch you next time.