AI Repo: kyutai-labs/pocket-tts

Skyler: Hey everyone, welcome back to Git News A I, the show where we dig into the hottest A I repositories blowing up on GitHub right now. Today's episode is brought to you by LimitLess A I — more on them in a bit. We've got a packed show today covering text-to-speech that runs right on your CPU, autonomous penetration testing agents, a supercharged K V cache layer for L L Ms, a classic OSINT tool, and a slick menu bar app for tracking your A I coding usage. Let's get into it. Skyler: Alright, first up we have pocket-tts from kyutai-labs, sitting at just under seven thousand stars and climbing fast. The tagline is literally "a TTS that fits in your CPU and your pocket" — and honestly, that alone had me clicking immediately. Text-to-speech has traditionally been one of those things that either sounds robotic, requires a GPU, or costs you money through a cloud API. Pocket TTS is trying to blow all of that up. The whole idea is that you get high-quality speech synthesis running entirely on CPU hardware, no GPU required, no internet connection, no subscription fees. That is huge for developers who want to ship voice features in apps without depending on external services. Think about offline assistants, embedded applications, edge devices, local productivity tools — this opens up a ton of doors. The recent activity in the repo is also telling — they're actively adding community projects to the README, including something called LocalVocal and seshat-tts, which means there's already a growing ecosystem forming around this library. If you're building anything that needs a voice layer and you want to keep it private, fast, and cost-zero at inference time, pocket-tts absolutely deserves your attention. Check the show notes for the link. Skyler: Next up, we've got pentagi from vxcontrol, and this one is wild — nearly twenty thousand stars on a repo that describes itself as a fully autonomous A I agents system capable of performing complex penetration testing tasks. Let that sink in. Penetration testing, or pen testing, is essentially ethical hacking — you're simulating attacks on a system to find vulnerabilities before the bad guys do. It's skilled, time-consuming work, and this project is building A I agents to do it autonomously. The fact that it's written in Go is also interesting — Go is known for being fast, concurrent, and great for networked systems, which makes a lot of sense in a security context. Now, the recent activity is focused on installer fixes and error handling improvements, which suggests the team is actively hardening the deployment experience, making it more reliable for real-world use. This isn't just a demo or a proof of concept — it looks like people are actually running this thing. If you're in the security space, a red team operator, a developer building security tooling, or just someone deeply curious about where A I agents are headed in high-stakes domains, pentagi is one to watch. Just use it responsibly, obviously. Link is in the show notes. Skyler: Okay, now we're getting into some serious L L M infrastructure territory. LMCache from the LMCache organization has crossed ten thousand stars, and it's targeting one of the most critical bottlenecks in deploying large language models at scale — the K V cache. So quick background: when an L L M processes a prompt, it generates what's called a key-value cache, which stores intermediate computations. If you're serving lots of users with similar or repeated context — think system prompts, shared documents, or RAG pipelines — recomputing that cache every single time is massively wasteful. LMCache positions itself as the fastest K V cache layer you can bolt onto your L L M serving stack. It essentially lets you store, reuse, and share that cache across requests, which can dramatically cut latency and compute costs. The recent activity shows the team doing real engineering work — refactoring legacy allocator paths and cleaning up logging style for code quality, which tells me this isn't vaporware, it's a maturing, production-minded project. If you're running L L M inference at any kind of scale, whether that's with vLLM or another serving framework, the efficiency gains here could be massive. M L engineers and infrastructure teams, this one's for you. Seriously, check the show notes for the link. Skyler: Alright, switching gears a little — sherlock from the sherlock-project is an absolute legend in the open source OSINT world, sitting at an insane eighty-six thousand stars. OSINT stands for Open Source Intelligence, and sherlock does one very specific thing extremely well: you give it a username, and it hunts that username across hundreds of social networks simultaneously. It's basically a reconnaissance tool for finding where someone has a presence online. Now it's not strictly an A I project, but the applications in the A I space are real — think training data gathering, research pipelines, security analysis, and increasingly A I-powered OSINT workflows where sherlock is one piece of a larger automated system. The recent activity shows they added an OSINT Industries banner, which suggests some interesting partnerships forming in the broader intelligence community. With eighty-six thousand stars this is clearly battle-tested and beloved by the community. Security researchers, journalists, red teamers, and developers building any kind of identity or verification tooling — if you haven't looked at sherlock, you really should. The link is waiting for you in the show notes. Skyler: Last but definitely not least, CodexBar from steipete, coming in hot at over seventeen thousand stars. This is a Swift app, so it's macOS-native, and it does something delightfully practical — it puts your OpenAI Codex and Claude Code usage stats right in your menu bar, and it does it without requiring you to log in through any web interface. If you've been using A I coding assistants heavily, you know the pain of trying to figure out how much you've actually consumed, what your costs look like, or whether you're about to hit a limit. CodexBar just surfaces all of that in a clean, always-visible way. The recent activity is great to see too — they added a Cinnamon applet link, which hints at community ports being built for other platforms, and there are translation updates happening including Catalan, which means this thing has genuinely global contributors. It's a small tool but it scratches a very real itch for developers who are deep in the A I coding assistant workflow. If you're a heavy Codex or Claude Code user on a Mac, honestly just go get this. Link's in the show notes as always. Skyler: And that wraps up today's episode of Git News A I! A huge thank you to LimitLess A I for sponsoring the show — if you haven't checked them out yet, seriously do, they're building some fascinating stuff in the personal A I space. If you found any of these repos useful or interesting, make sure you hit subscribe so you never miss an episode, and all the links we talked about today are sitting right there in the show notes waiting for you. Until next time, keep building, keep exploring, and keep pushing commits. See you in the next one.