WEBVTT 00:00:00.160 --> 00:00:40.130 Hundreds of new open source AI projects hit GitHub every single day, yet only the smallest fraction of a percent are actually worth your time. But today, I'm gonna be highlighting 10 that are. Almost every single tool we're gonna cover today has just come out within the last month. So unless you are as obsessed with this stuff as I am, I promise you're gonna get exposed to at least a few new tools. Now the first tool on the list is my favorite because it's one I use literally every single day at this point, and that is the caveman skill. This repo's gotten over 50,000 stars within its first month of release, and the whole idea is this just a skill that we can use inside of Claude code or codex that makes our agent talk like a caveman. 00:00:40.130 --> 00:00:50.115 AKA it's not gonna be so damn verbose. So you have some examples here where you have your normal Claude code response where it kinda just goes on and on and on. But if I use caveman, 00:00:50.275 --> 00:01:08.850 well, it's kinda just gonna get to the point. This repo has taken the idea of why say many word when few do trick and just codified it. Now the cool thing about caveman is that there's levels to it. Like, we don't have to go full Neanderthal. We can do caveman light, which is what I sit at. We can also do full or we can do ultra. Now I will say this repo 00:01:09.185 --> 00:01:16.625 gets a little excited when it comes about how much how many tokens you're saving. You'll say, like, hey. We're saving, like, 75% of output tokens. Understand 00:01:17.025 --> 00:01:26.250 that the way Caveman works is it's just going to be changing how many words you see. It doesn't change how it does its thinking. It doesn't change the amount of stuff it's ingesting. So overall, 00:01:26.410 --> 00:01:29.450 if we take it all together, you're looking at about a 5% 00:01:29.450 --> 00:01:46.055 or so savings when it comes to tokens. And I've done a full video on this, and I'll link that above if you wanna do a deep dive. Now I think the secret win when it comes to caveman, and I think they kinda bury the lead here, is the idea that our large language models might actually do better if they're forced to give more concise responses. 00:01:46.215 --> 00:01:49.975 And this comes from a March 2026 paper. It's called Brevity Constraints, 00:01:50.135 --> 00:02:44.315 Reverse Performance Hierarchies and Language Models. And, basically, the long and short of it is when we have powerful models and we force them to be concise, they're more likely to give us correct answers because they're essentially not going to talk their way into the wrong answer. And it's actually a really interesting study, and I highly suggest you take a look at it. So we take those things together where we're going to be saving tokens and I'm potentially getting an actual quality increase. What's not to love about this thing? And it's just a simple skill. Installing this is super easy. You can just run the commands here inside the repo, or you can just copy the repo URL, put it inside of Cloud Code, and say, hey. Let's start running Caveman. And if you wanna do Caveman Lite, just say Caveman Lite. If you wanna do Ultra, do Ultra. It's very easy to execute. I'm always a huge fan of these lightweight tools that give us the wins on the margins without any real downside. So if you check out nothing else here, check out caveman. But before we move on to tool number two, a quick word from everybody's favorite sponsor, 00:02:44.555 --> 00:02:58.360 me. So I recently came out with my Claude code masterclass, and it is the number one way to go from zero to AI dev, especially if you don't come from a technical background. I update this every single week, and we really focus on real life use cases 00:02:58.520 --> 00:03:03.800 and building upon the foundation of Cloud Code with things like an entire agentic OS system. 00:03:04.205 --> 00:03:15.645 So if that's something that you would be interested in, you can find it inside of Chase AI plus. There's a link to that in the pinned comment. Now tool number two is all about memory and knowledge graphs, that is Graphify. 00:03:15.645 --> 00:03:28.940 Now what Graphify is able to do is it reads our files to build a knowledge graph. And because we now give Cloud Code a clear structure to understand what we're working with, we're able to execute our task while using way less tokens per query. 00:03:29.180 --> 00:03:30.860 They quote 71.5 00:03:30.860 --> 00:03:58.490 times fewer tokens per query versus reading raw files. Now when we talk about knowledge graphs and memory, a lot of us first start to think about things like Obsidian. But Obsidian, while this does give us a knowledge graph in theory and that's what we're looking at now, this isn't a true knowledge graph in the sense of, like, a graph rag system, something like LightRag or rag everything. Graphify is much closer to that true rag structure than something like Obsidian is. Remember, Obsidian, for all intents and purposes, 00:03:58.730 --> 00:04:04.305 is just a nice interface for us to be able to deal with markdown files and markdown files exclusively. 00:04:04.465 --> 00:04:06.145 Graphify is multimodal. 00:04:06.145 --> 00:04:15.330 Now it's not multimodal in the sense that it's gonna be ingesting pure video, something like, you know, Google's embedding too, but it is able to look at things like PDFs, 00:04:15.330 --> 00:04:15.890 screenshots, 00:04:16.130 --> 00:04:17.010 diagrams, 00:04:17.010 --> 00:04:29.490 and it's able to take videos and then use Whisper to pull what it needs out of there. Furthermore, Graphify doesn't use embedding. So when we're talking about sort of that spectrum between something like this, Obsidian, and a true rag system, something like LightRag, 00:04:29.915 --> 00:05:04.235 I would say Graphify sits somewhere in the middle. And it's something that we can essentially layer Obsidian on top of. So if you're someone who loves Obsidian, wants a little extra power in terms of what's going on with your memory and your files under the hood, yet you don't wanna take the step into some sort of true rag system with embeddings and everything like that, well, Graphify might be perfect for you. And definitely stay tuned for a deeper dive on this particular topic for on a video that might be coming out in the next few days. Now tool number three is one you probably haven't heard of before. It's Claude Video. Just came out last week. We're at 400 stars. 00:05:04.395 --> 00:05:15.080 And what it does is it gives Claude the ability to watch video. Now what do I mean by that? Because we know Sona and Opus can't ingest video. Well, this tool has a pretty clever approach. 00:05:15.240 --> 00:05:18.440 Once it's given a video, it uses FFmpeg 00:05:18.440 --> 00:05:20.440 to extract the frames 00:05:20.440 --> 00:05:25.080 at a particular rate depending on the length of the video. Obviously, if it was 60 FPS 00:05:24.875 --> 00:05:26.395 and it's a ten minute video, 00:05:26.635 --> 00:05:38.795 that would cost an insane amount of tokens. So it gives it a default frame budget based on the duration of the video. So a thirty second video would be 30 frames. If it's ten plus minutes, it would only be a 100 frames, so it gets kind of sparse. 00:05:39.115 --> 00:05:44.910 But it essentially feeds screenshots to Cloud Code, it grabs the audio via Whisper, 00:05:45.150 --> 00:05:49.150 and it uses those two things in combination to essentially watch videos. 00:05:49.390 --> 00:06:01.315 Now I think this is a really useful tool because when it comes to handling videos, there's really only two other pathways right now when it comes to something like Cloud Code or Codex. And that's, alright. Let's just send it off to something like Notebook LM 00:06:01.555 --> 00:06:06.595 and have it figured out. Or in kind of in that same category is, let's invoke 00:06:06.940 --> 00:06:07.820 Gemini 00:06:08.060 --> 00:06:12.700 via an API call and just send it that way. This gives us sort of a, 00:06:13.100 --> 00:06:26.775 you know, different approach where we aren't beholden to Gemini to deal with these videos for us because we're breaking it down via screenshots. Obviously, when we talk about longer videos, three minutes plus, ten minutes plus, you're gonna run into issues just like what are you actually trying to do with these videos. 00:06:27.095 --> 00:06:37.760 But I think anything that gets us closer to having Cloud Code being able to handle video is a great tool for us to play with moving forward. Now tool number four is one I did a video on recently and that is OpenDesign, 00:06:37.760 --> 00:06:38.880 which is essentially 00:06:39.040 --> 00:06:49.040 an open source clone of Claude Design. You can now use Claude Design or something pretty close to it with any sort of coding agent. So you could do this completely locally 00:06:49.445 --> 00:06:56.165 for free. You don't even have to be on Claude Code. They've copied the exact layout of Claude design in terms of being able to create prototypes, 00:06:56.245 --> 00:07:14.920 slide decks, and added some additional functionality like also being able to call APIs for image creation and for video creation. And OpenDesign itself is really built upon four other open source tools. The first one being Hua Shu design, which is basically another Clot of Clot design, but it's purely inside the terminal. 00:07:15.425 --> 00:07:21.025 The Goozang PowerPoint skills, so allowing us to create these PowerPoints and then actually extract them properly, 00:07:21.505 --> 00:07:23.105 as well as OpenCodeDesign 00:07:23.105 --> 00:07:49.905 and then MultiCAD. So it's taken all four of those, added a package of 31 skills, and voila, we essentially have local Claude design. So if you're someone who really likes claw design, especially the graphical user interface portion of it, I highly suggest checking this out if you've already hit your usage limits for the week. Now if you're someone who cares about where your tokens are going and how much money you're throwing away every single month on these coding tools, then you are gonna like tool number five, which is CodeBurn. 00:07:50.065 --> 00:07:59.585 CodeBurn tracks token usage cost and performance across 16 AI coding tools and allows us to get a much better look at where our tokens aka our money is going 00:08:00.080 --> 00:08:07.360 well beyond what, you know, forward slash usage is gonna give you inside of Cloud Code. You can see in this dashboard, it breaks it down by activity, 00:08:07.520 --> 00:08:08.240 project, 00:08:08.480 --> 00:08:53.675 model, core tool, shell commands, MCP servers, and shows us not only how many tokens we're using, but, like, the actual dollar amount, which is really important, especially if you're on the API. Now more importantly than just telling us where our tokens are going and how we're losing money, it gives us ways to fix the problems. It tells us how to optimize our systems so we stop burning so many tokens. So just like Caveman, think CodeBurn is one of those lightweight tools that is almost pure upside. So definitely take a look at this one. Tool number six is Impeccable. Now Impeccable came out a couple months ago, but they recently came out with their three point o version just last week, which is why I kinda wanted to include it because their updates to Impeccable include the ability to actually edit front end designs in a browser. And if you didn't understand by now, Impeccable 00:08:53.675 --> 00:09:09.930 is a tool for front end design. Impeccable ships with a single skill yet that single skill includes 23 different commands that are all about making sure your web pages don't suck. What I like about Impeccable is it includes this website where I can actually see what each and every command does. So it shows a before 00:09:10.205 --> 00:09:21.245 and an after. And you can see, okay, like, what will actually happen if I use this skill? Furthermore, it now has a live mode where you can actually bring up your web page, click on different components, 00:09:21.405 --> 00:09:37.080 and then go through different variations on the browser itself. I actually did a whole deep dive on this and I will link that video above if you wanna see this in action. But I think the best part arguably might just be the website and the ability to see all these before and afters and just kinda give you inspiration for like, alright. Here's what 00:09:37.485 --> 00:10:00.530 my AI slop looks like versus what it should look like and seeing the different ways you can make minor adjustments on individual components, but in totality, that can really change the way your website looks and feels. And again, this live mode just got released. So if you've used impeccable in the past without it, highly suggest you take a look at it again. So sticking with the front end design theme, tool number seven is design extract. 00:10:00.770 --> 00:10:06.850 Now a big repo that came out a little while ago, I've talked about in the past is awesomedesign.md. 00:10:07.115 --> 00:10:08.795 Now awesomedesign.md 00:10:08.795 --> 00:10:24.230 has taken off since it first came out about two months ago. It's up to 70,000 stars. And the idea is is they give us this repository of all these popular websites, say, for example, Eleven Labs. I click on it and I can see essentially an entire breakdown of what 00:10:24.790 --> 00:10:39.965 their website looks like from an aesthetic point of view. You know, what are the cards? What are the colors? What's the spacing? What's the font? Etcetera etcetera. The problem with Awesome Design MD is I can only choose from these. I mean, there's a lot to choose from, but I'm limited as to what I can do. DesignExtract 00:10:39.965 --> 00:10:46.765 takes it a little bit further because it's essentially allowing us to get the same thing I showed you here inside of DesignMD, 00:10:47.500 --> 00:11:03.535 but for any website we want. So we point this design tool at any website we wanna use as inspiration as a foundation for what we are building, and it's gonna grab the layout system, responsiveness, interaction states, motion language, component anatomy, brand voice, on and on and on and on. So we have a comprehensive 00:11:03.535 --> 00:11:15.055 thing we can then bring into Cloud Code and build upon with our brand. And it does all this by using a headless browser to actually grab all this information. So it's a bit more than just taking a couple screenshots 00:11:15.215 --> 00:11:24.230 and saying, hey. Copy this. If So you're someone who loves this awesome design repo but wishes there were some more websites on here that you could essentially use, well, 00:11:24.870 --> 00:11:26.470 definitely check out DesignExtract. 00:11:26.470 --> 00:11:49.390 If you've ever thought about using Claude code to help you apply to jobs or get your resume in order, well, you will like this tool, and that is CareerOps because that's exactly what it does. As they state here, CareerOps turns any AI coding CLI into a full job search command center. It evaluates the offers of the jobs out there. It generates tailored PDFs. It scans portals. It processes in batch and essentially tracks 00:11:49.630 --> 00:11:56.110 everything related to the job search process, which is brutal. And importantly, this isn't a tool that's just like a mass 00:11:56.685 --> 00:12:04.285 application tool. This isn't like, oh, go on LinkedIn and now apply to every single job under the sun. Like, this is much more of a scalpel that's going to tune 00:12:04.605 --> 00:12:29.475 your resume to the job and make sure the jobs you're actually looking at makes sense for you. This isn't just like, alright. Go out there and just like throw up all over the job application process. Under the hood, it's using Playwright to actually navigate the pages. It evaluates the fit based on your CV and then adapts it per each listing. And here's how the general flow works. You paste in a job URL or description. It then classifies it. It then figures out, are you a match? 00:12:29.795 --> 00:12:52.620 Before then generating a report, the PDF, and then updating the tracker. So definitely a useful tool if you or anyone you know is trying to leverage something like Cloud Code to help them in their job search. Now Now tool number nine is one I think you're gonna hear a lot more about, and that is browser harness. So think of Playwright if Playwright was self improving after every single run. So the way it kinda works is if I used browser harness to say do something 00:12:53.285 --> 00:12:54.245 on Amazon, 00:12:54.485 --> 00:13:15.830 every time it went to complete a task on Amazon as this Autonomous Browser agent, it would update its own agent skill file saying, okay. This is what we did for Amazon. Here's what worked. Here it didn't. And almost in a sense, almost like a mini Ralph loop where we've given it a task. It's gonna always update its files to see, hey. Did it work? Did it not work? What did we already try? And then try again 00:13:16.310 --> 00:13:25.015 based on the information it wrote about itself and sort of the, like, self healing thing. So it's still pretty new. It's only been out for a couple weeks. It's just under 10 k stars. 00:13:25.335 --> 00:13:34.695 But I think this sort of agentic approach to these browser agents is something you're gonna see a lot more. Now I cheated on the last tool on the list because it isn't technically 00:13:35.040 --> 00:14:13.990 open source and even any then itself isn't technically open source. It's fair use, but, you know, you can use it locally so it gets a little confusing. And that is the any then MCP server. Now I think the death of any then has been greatly exaggerated, but let's be honest. It isn't in the same place it was even six months ago. Yet they've begun to realize and pivot into being a tool that Claude code can use very, very easily, especially with this brand new MCP server. So this MCP server is a little different than any other any in an MCP server that has come out because there's been a few out there, and they were open source. The difference is this one uses TypeScript instead of just trying to generate a JSON file automatically. 00:14:14.310 --> 00:14:16.550 So I give the N8N MCP 00:14:16.710 --> 00:14:19.590 some sort of command like build me whatever automation. 00:14:19.910 --> 00:14:20.390 It then 00:14:20.825 --> 00:14:22.505 builds it in TypeScript, 00:14:22.505 --> 00:15:05.463 which allows it to actually validate the automation to see, hey. Do these notes make sense? Will this actually work? From there, as a last step, it gets changed to JSON, and then it populates inside your instance. So if you're someone like me who still really likes n eight n and there are use cases for it, although it can be kind of niche, this is an awesome tool. It just came out a few days ago, and I did a full video on that as well, and I'll link that above. So those are my 10 favorite open source tools for Cloud Code that have come out within the last month or so. Like I said, this space is literally always changing. It is impossible to keep up. So I hope by watching this, you were able to see at least a few of them that you might wanna check out. As always, let me know what you thought. Make sure to check out Chase AI plus if you wanna get your hands on that master class. 00:15:05.703 --> 00:15:07.463 And besides that, I'll see you around.