What Claude Stats Tell Us About AI Coding Tools

If you have ever found yourself wondering what the scale of agent-assisted development is, or you've wondered what types of problems Claude is helping engineers solve, then I've got good news. I'm going to answer a couple of these questions for you. Who am I? My name is Joran Olberts. I'm the head of AI engineering at Advanced.io, a company in South Africa. We're a part of a larger group called Group Elephant. I have about a decade's worth of experience in software engineering, and the last couple of years that has been primarily focused on AI engineering and the development of agents.

A couple of months ago, as a side project, I built a website called Claude's Code, and Claude's Code is a dashboard that displays a number of metrics that I track off of public GitHub. So how that started was I started seeing the Claude logo in the contributors list for a large number of repositories that I visit. This sort of piqued my interest, so I thought I would see if there's a way to find out how many repos did have Claude as a contributor. Initially, what I did is I looked at whether there were commits authored by the Claude bot.

Now, in this case, the Claude bot would be mostly people using the GitHub Action to action some sort of code change on their repository. This means typically code reviews, small bug fixes on PRs so you can integrate the action on GitHub. I only found 8,000 commits, which I thought was too low, but that was mostly because I was only looking at commits for the bot. Now, after that, I thought, okay, well, I will expand my search space. So I've expanded my search space to look at commit messages that were signed by Claude.

So what that means is at the end of the commit message, typically, if Claude code was used to author the commit, there will be a signature saying that Claude is a co-author. This yielded many, many more results. So after defining the pipeline, I now have a continuously running pipeline that reads those commit messages, and then I have a couple of interesting stats on my dashboard that I display. Today, I want to talk a little bit about the stats that I found there, but I'm going to talk about why I think that viewing agentic coding as a phase or a fad, or as a non-serious contributor to the software engineering world, is a mistake.

Briefly, maybe an interesting story about the dashboard. When I originally launched it, I posted it on Hacker News. It was on show HN. It sort of got nowhere. I didn't get any traction. Nobody found it too interesting, which is fine. That happens, I think, frequently with these projects. A couple of days after posting it, though, someone else posted referencing my dashboard, and this took off. Within a couple of hours, I was on the home page of Hacker News, and suddenly had thousands of visitors. If anybody has experience to go live, I'm sure you know things don't usually go as planned.

So that night, I didn't sleep much, but the little server I have running, the little potato server, did survive. And over the course of a couple of days, I got around 20,000 visitors. And more interestingly, some attention from, I think, some, let's say, giants in the space. For instance, Boris Churney bought me a coffee. So thank you, Boris, if you ever see this. Which helped keep the server alive for a little bit longer. So what did I see initially? I think the big thing, and the question I wanted to answer initially was, what is the scope or scale of Cloud Code deployments or development?

Now, over the last few months, we found an enormous number of commits. Cloud Code was released to the public in February of 2025. And that means we're 14 months or so down the line at this stage, 15 months down the line. In that time span, I've picked up 45, a little bit more than 45 million Cloud-signed commits. There could be more. Of course, I expect there will be more, some that I didn't pick up, but this is already an enormous scale.

On top of that, there's around 2.4 million unique repositories. So this is quite widespread. We're not necessarily seeing a very small batch of repositories generating an enormous amount of code, although there are some exceptions to that rule. So definitely the first thing you can see is that there's an enormous volume of Cloud Code-generated code. Specifically, we're looking at around 50 billion lines of code that have been contributed to public GitHub. And that's the net code. So we have more, it is 85 billion or so, but there have been several deletions, so sort of refactoring, you can imagine.

We can also see something interesting in the distribution of languages. Now, you'll see there's a large number of unknown languages. So that means on that repository, the language has not been sent or couldn't easily be detected. I'll circle back to that. But what I want to touch on is the TypeScript and Python popularity. Now, I think anybody who works in the AI space is not going to be surprised by this. Specifically, Python is quite a popular, let's say, AI-focused language these days. It's driven largely by machine learning and data science development. But also, I think more recently, with the popularity of vibe coding picking up, many sort of pet projects have been spun up using Python because it's pretty easy to get started.

It's driven largely by machine learning and data science development. But also, I think more recently, with the popularity of vibe coding picking up, many sort of pet projects have been spun up using Python because it's pretty easy to get started. And it seems that the agents are very good at coding it. I think it's a similar story for TypeScript. But with TypeScript, we have other drivers, I think, contributing here. Many people are developing web applications, or some version of cross-platform applications. And I also think the spike in popularity for things like Next.js and React have contributed to this.

So not too surprising seeing TypeScript and Python rank highly. What I did find interesting is that we see languages like Rust, Dart, and Zig also rank relatively high. Now, what I think is happening there, and these are still much smaller than the rest, but what I think is happening there is agents are making it much easier to code in languages that might traditionally have been considered slightly more technically difficult. So what that allows developers to do, let's say someone like myself, where I might be interested in developing an application using Rust, or I've developed an application already, and I'd like to improve the performance.

In the past, it would have been a monumental undertaking to rewrite that code base. But now what I can do is I can use an agent like Cloud Code. And then I can have that run through the rewrite with me. In some cases, it could just do the rewrite outright. And I think that's why we're seeing an uptick in languages like Rust and Zig, rewrites of libraries, but also just pet projects that would have been written in other languages like Python, which are a little bit easier to understand or wrap your head around, are now being written in more difficult languages like Rust or Zig.

Initially, the dashboard got a lot of attention because someone mentioned that more than 90% of the commits are two repos with fewer than two stars. So why does it matter how many stars the repo has? Many people will say it doesn't necessarily mean it's a good or bad repo. I agree with that. But if you have more than two stars, it at least means one other person has found some sort of value on the repo. If there's one or fewer stars, it could be that you've self-starred the repo, or it's just a personal project. So this assists a little bit in removing some of that personal project or noise from the stats.

Now, this, again, is not unexpected. It makes sense that most Cloud Code projects are personal projects or a new project, it's a relatively new technology. And I think you'll notice that the original stats were that more than 90% had fewer than two stars. And that, it seems to be dropping. So these projects, in some cases, are actually picking up users and picking up stars, which I think is an interesting occurrence. Just in numbers, that means only around eight million of the 45 odd million commits are in stars and repos that have a few stars.

The other interesting thing is the growth. Some of this just makes sense. So again, the launch, the actual first, I think, beta was in February. So you see a small uptick in commits there, and then public launch, I believe, was in May, open to the public, and then you see an uptick here. Now, these vertical lines that you see, these are product releases by Anthropic. Some of them are models. And then, of course, in this case here, we have Cloud Code's own release there. So sometimes they correspond to upticks. What I think is telling here is that we start to see almost a plateau. And then right after the Christmas season, there is an exponential growth in usage.

And if you look at this breakdown on a day-to-day granularity, you can actually see that uptick happen directly after Christmas. I can only speculate why that happened, but my feeling is that tooling matured and models matured. And also, at that time, people had a little bit of free time. So they could play with perhaps agents they hadn't tested before and found that they were quite capable. So after December, we see this exponential growth in the number of commits per day.

Currently, it seems to have started to even out. So we're not seeing as much growth. The doubling time initially in that sort of exponential period was a matter of days. You saw the number of commits double in days or weeks. Now, we are looking at a doubling time of around four months. Still impressive. If you take into account that, would mean we have upwards of 19 million commits then. So current pace is around 400,000 commits and new commits to public GitHub per day signed by Cloud Code.

So I wanted to zoom in a little bit on that data. And to do that, I selected what I consider to say the top 100 most popular repos and then picked a time frame that includes some of this time that we had spike and some of the time prior to that. So August to March. The repos were selected mostly based on star count again. But I think you'll agree, if you look at the top five here, that these are popular and widely known repositories. I then split them into AI-focused projects and non-AI-focused projects. Now, I know the non-AI projects have AI components in some cases, but I do think that the AI projects are sort of expected to make use of agentic development.

So some of the signals or interesting information, I think, is maybe highlighted better by projects that are not specifically AI-focused, making use of agentic coding. Some of those immediately jump out, right? Dano, Home Assistant, Bunn, really all of the non-AI ones are quite popular. And then on the AI side, you'll see just familiar terms, even if you don't know the project, right? MCP being quite a big one there. So then I tried looking at how are these projects using it. How I did this is I took my sample of 2,500 commits from those repos, then I decomposed them into more atomic changes. If there were multiple changes, then I ran a sort of open labeling on that to summarize what the changes were, which categories they fall in. And then I ran a clustering algorithm to see which clusters emerge.

Here we have the six biggest clusters that did emerge, and these make sense. Bug fixes, feature changes, documentation, refactoring, tests, and then UI or visual changes. A couple of things here I find interesting. Firstly, there's a little bit of contrast between the AI and non-AI repos. We see that the AI repositories are contributing more features as opposed to non-AI repos focusing mostly on bug fixes. I think that makes sense. On the one hand, you do expect AI repositories to be newer, so they may be generating more features, but these are typically quite mature repos. So it's not that they're implementing the system still, right? They have to have reached some sort of stability to be as popular as they are. And then you expect the non-AI repos to be implementing mostly bug fixes, but interesting nonetheless.

Then the other thing I think is interesting is that we are seeing UI and visual changes still being quite a big portion of this, but the least of the six most popular. And this makes sense to me. I think in UI and UX space, we are seeing that designers and developers are still more involved in that space, but this seems to be changing with more modern multimodal models. Other than that, the rest of them track documentation, refactoring, and tests. Those are things we would want AI to do because they're not always the most fun. I picked up a couple more interesting, let's say, contrasts between the AI and non-AI ones. For one, AI repos seem to have Cloud Code commits more frequently.

So in a window of 11.8 hours, every 11.8 hours, you typically see a commit by Cloud Code on one of these repos. That's longer on the non-AI repos, 16.5 hours. Again, I can only speculate as to why that is, but I think that the AI forward repos, the AI repos are using agentic coding in a slightly more autonomous way, maybe less involved in the process, might be trusting, so to speak, the AI a little bit more. You're also seeing that AI repos have, due to that, right, more commits. And you're seeing that the size of changes by AI repos are typically a little bit smaller. So it seems they're committing more atomically. It's not an extremely large gap, 40 to 39 lines compared to the 64 lines in non-AI repos.

So the cadence really is what you would pick up here, that on AI repos, we're seeing more frequent, smaller commits by agents as opposed to the non-AI repos. In both cases, we do pick up some very large changes now and then. To look at what those might be, I picked two repositories and sort of zoomed in again, looking at specifically what they've done using Cloud Code on their repository. So firstly, Apache Superset, which is a BI platform, a business analysis platform. And Superset is not overtly sort of an agentic product or anything like that. So it was categorized in the non-AI repos, in my case.

What I think stands out here is that they also used Cloud Code to do, and I quote, comprehensive TypeScript quality improvements. These are front end changes and they're deferring to the agent to improve the quality of human written code, which I think is significant, right? Means SuperSig definitely is putting quite a lot of responsibility in the agent's hands. The second use case I found interesting, and this is not a repo in the top 100, actually, but I noticed this myself on the dashboard, is Clickhouse. So, Clickhouse is a database usually used for data analytics or data science. And Clickhouse is extremely popular. It's picked up quite a lot of popularity now as well. And they have been using extensive use, but they're making extensive use of Cloud Code.

But what I think stands out here is that they also used Cloud Code to do, and I quote, comprehensive TypeScript quality improvements. These are front end changes and they're deferring to the agent to improve the quality of human written code, which I think is significant, right? Means SuperSig definitely is putting quite a lot of responsibility in the agent's hands. The second use case I found interesting, and this is not a repo in the top 100, actually, but I noticed this myself on the dashboard, is Clickhouse. So, Clickhouse is a database usually used for data analytics or data science. And Clickhouse is extremely popular. It's picked up quite a lot of popularity now as well. And they have been using extensive use, but they're making extensive use of Cloud Code. So, in the Clickhouse case, we're seeing a different type of use, I think, many more commits. They have, you know, 10 times the commits that I picked up on the SuperSig side over the same time period. Not many feature additions, in their cases, largely skews CI testing and infrastructure. And I think this is a good example of a different type of use of Cloud Code, where the agent is picking up your sort of chore-type work and making sure that engineers are spending their time on what I would consider the more interesting and more exciting work, which is feature development and maybe some bug fixes. So, that's not saying the agent isn't doing bug fixes for them, but the largest portion of this is sort of optimization and continuous integration and infrastructure. Specifically, a couple of commits I picked there. I think most of this is not the most interesting thing. So, we have that sort of test and infrastructure-type thing in known failures. But then what we do see here is that they are also writing Cloud Skills in their repo. Now, this is expected in some cases, but we're seeing more and more of this, where repos, existing projects now have Cloud Skill directories in their repos. So, in this case, we have Cloud Writing Skills for itself to help with query profiling and PR workflows, which I found really interesting. All right. So, I think a lot of information to gain there. And I'm still learning a lot from what I've mined and will be learning for a while, I think. Something that I've been picking up recently is that we have an uptick in repos that explicitly are Cloud Skill or Cloud Adjacent, Cloud Tool-type repositories. Specifically, there's around 45,000, 44,000 repos that mention Cloud Skills or Cloud Tools in their naming or in their descriptions. So, these are repos, not commits, which is quite a large number of them that exist just to be Cloud Skill repos. And that's growing. That number is growing. If we look broader to look at repos that mention Cloud, generally, we mentioned Cloud Tools, MCPs, or Skills. Again, specifically focused around Cloud, we're seeing around 170,000 of these repositories. So, that's 6.6% of all of the Cloud repos or repos that have Cloud commits that I've read, which I think is a significant number, exist purely to cater to Cloud. And again, that's a growing number.

All of these seem to be quite active. These are not mostly stale projects, 99.9% of them have recent activity. So, this is definitely a currently growing trend.

All this to say, I think I've answered the question that I had, right? Is AI doing real work on real projects at a significant scale, a noteworthy scale, and is that happening frequently? Is this now a day-to-day thing? I think the answer to the question is a resounding yes. Definitely, at least from the perspective of Cloud Code and on public GitHub, Cloud is contributing a significant amount of code to code bases that are seeing daily use, significant, important projects, repos that are not necessarily AI repos, that are not necessarily trendy, just real core infrastructure critical software is seeing a very real, seemingly valuable contribution by Cloud Code and definitely by other agentic coding tools.

So, I hope I've answered a couple of the questions you may have had. And it's been a pleasure taking you through my descent into madness. If anybody does have any other questions, we just want to send a hello, something interesting, feel free to reach out. Thank you so much, Elon, for your massive research and for presenting us all the outcomes and insights. While we wait for more questions from our virtual attendees, I have a couple of ones for you.

So, this was about Cloud Code, that clearly marked itself as a model contributor. Are there any other models or maybe coding agents who self-identify themselves the same way like Cloud does? Yeah, it's a good question. Firstly, thank you to everybody who tuned in. Yeah, Maxim, so I've actually found a couple of the other agents do self-identify. These Frontier labs are sort of motivated to have the agents identify because they would like to avoid training on code generated by AI in many cases. Specifically, interestingly, I'm seeing more and more of Mistral's models identify, which I didn't exactly expect. But every now and then, I do a little manual search to check if there's something else I should be paying attention to. The reason I picked Cloud Code is because I think it was the first very popular one that was fully agentic.

We had Cursor initially, but that was AI-assisted coding. But Cloud was the first one that really went, okay, I'm generating all of the code, writing the commits, and it seems like Cloud is definitely the one that does it most aggressively, even though you can disable it. I've seen in my own work that sometimes even disabling it doesn't always make it not sign. By the way, I also noticed, I do a lot of building with AI myself, and I noticed that when I use GitHub Copilot, it also explicitly adds Copilot as... First, it was just contributor. Now it's co-contributor. Now all this code is marked as two developers, Copilot and actual developer. Did you notice that as well? And also, did you run some research on GitHub Copilot specifically? Yes, I have noticed that. Actually, Cloud also does both ways in some cases. So some commits you will see Cloud listed as a co-author, and in others, Cloud will be listed as an author. It doesn't seem like there's necessarily a pattern with that, although I have noticed that Opus, for instance, is more aggressive for listing itself as a co-author. So I'm not sure if that is something that Anthropic handled in training. I have noticed the same thing with Copilot.js. I haven't been pulling the stats for that explicitly, but I do plan on doing a little bit of monitoring for Copilot and some of the other agents. I think Codex's popularity I'm seeing tick up as well. So at some point, hopefully, I'll have something there. Please keep us posted about your next research.

Have you found any correlation between security issues and intensive use of Cloud? Let me generalize that a bit. Do you see that the more actively people use any AI systems, maybe more or less, I don't know, security issues they end up with? Yeah, that's a good question. And it's actually one not necessarily related to security, but more generally, I am trying currently to track whether there's a correlation between the number of issues generally opened on a project and the use of an AI agent, not specifically security, but I think there will likely be security issues there as well. Thus far, what I've seen is projects who adopt AI development do see an increase of, say, the number of issues opened on GitHub.

Thus far, what I've seen is projects who adopt AI development do see an increase of, say, the number of issues opened on GitHub. But the problem is that's not necessarily causation. I think what's happening is they're simply producing more code. And because you're producing more code, inherently, you do have more issues. These issues are not necessarily bugs. Many of them are sort of functional issues. But it's a good question. I don't have a simple answer to that. I think it's a question many of us would probably want to know the answer to.

Yeah, I'd say this is a multi-dimensional question. Of course, the final result depends not only on the model itself or not only on the harness, but also really much depends on the context you attach while you use this model. And I don't remember who made this research, but I believe it's available somewhere that clearly shows the worse condition your code base is in, the better chance that you'll get smelly code as an output from the model. Because, I mean, it's a pattern game, right? A pattern that's learned from code during training and the same patterns learned during inference. You just provide bad context.

Yeah, sorry, Martin, one more point on that. There is one thing worth noting here is that it seems AI is more prolific in encoding TypeScript. I mentioned this in my talk as well. That TypeScript and JavaScript are very popular frameworks or languages, and the frameworks associated with them are very popular. And at least at the moment, if you look at the supply chain research, supply chain attacks are quite prevalent in npm. So that also contributes. The engineer is still responsible for setting the scene, as you said. If it's a bad code base to start off with, you likely will end up with some bad code. If it's a bad tech stack, you likely will end up with some problems.

100%. And I like that the next question is kind of opposite mood. It's about the Clickhouse popularity. So it's growing and someone is asking us over you, do you think there is correlation between the growing popularity and the fact that they actively use Claude for building this product? I would have to speculate to answer this question. I think that I don't think that it's necessarily the use of Claude that's driving it, but I do think Clickhouse is using this actively to accelerate their development. And at the moment, that's allowing them to sort of cash in on popularity that they've picked up and sort of exploit that. So I would say yes, I think it's contributing. I don't think it's necessarily that Claude is making it more popular, but it's allowing them to develop and fix things faster. And that's helping the project I would guess.

Yeah, of course. I mean, I feel this one more funny question. And I mean, that would be too good to be true, right? If you just use any AI assistant, your product becomes more and more popular. Definitely that contributes, right? But if you develop this fast and in the right direction, what's even more important, maybe, than it is. I love that, you know, this kind of differences between speed and velocity, how we explain that. So speed, this is just speed of like movement, right? But velocity is also including direction, right? Because you can move fast, but slightly different direction, and then it's totally wrong.

Next question is about... So you represented some metrics, right? I mean, also different ones that might be interesting to consider. What made you choose specifically the set you presented? Yeah, so firstly, the reason I did that, started doing the tracking in the first place, was just to satisfy my own curiosity. I've been following the development of large language models. I come from an AI background, so for me, it was fascinating in that sense. And when we started using agentic coding tools, AI coding tools, generally, I think, sometimes you wonder if it's just your bubble that's making use of it, or whether it's seeing widespread use.

So I had that same question, and that's why I initially started tracking the metrics. But the specific metrics that I discussed, is because I was quite surprised, I think, to see how quickly, especially in this year, from January till now, short codes use spread. I mean, it was like wildfire, absolutely exponential growth.

But the immediate next question is, is it contributing in a valuable way? So that's why I tried pulling those metrics. I was trying to think, how would you know if this is useful contributions? And I think how it's useful is if it's working on critical code, and seeing continued use. Hopefully, projects like Kickhouse, for instance, if they're not seeing valuable contributions, would stop the use. But they haven't. They're seeing an uptick in useful code.

Absolutely. I mean, in my sessions about AI and native development, I also highlight the importance of selecting right metrics, because you can measure so many things, but not obligatory something that gives you some insights. Does it bring you return on investment into these tools or not? Yeah, I think that we actually addressed a question from Chris about other tooling besides cloud, and you even promised us to share all the next research on an alternative model providers, alternative vendors.

And I believe we actually addressed the one about correlation of the models to commits, right? This is actually the way how you make this research. Yes, yeah, absolutely. Many of the models do identify the specific model. I didn't track all of that on the site, but Opus and Haiku are quite happy to volunteer whether it was them who contributed. Good, thank you so much, Jörgen. It was an insightful session and discussion, and of course, let's stay connected. Bye-bye. Yeah, thank you, Maxim. Thank you, everybody. Bye-bye.