Claude Code Accidentally Open Sourced Itself

Through some mishap, the source code that creates Claude code, which has famously been closed source for eternity, someone pushed that to the NPM registry, the source map, which is a file that allows you to decompress all of the all of the jumbled up code that a build process creates, and actually define all of the files that make up Claude code, revealing loads of juicy secrets. Command AI Surprised the person in the chat is like, "What Claude code leak?" Cuz this was everywhere. Jesus Christ. Uh the leak could help give software developers and Anthropic's competitors insight into how it built viral coding tool. And nonsense, that's important, nonsense of customer credentials were involved in the exposed. Thing is, I have to be honest, you know, I got sent this by people. I actually do not care. To be honest, it's juicy gossip and we'll go through everything, but I don't care. Like, I'm not one of those people who is like in a mad thing about open source and Claude code needs to be open source. Um but you know, it's there. And someone actually has rebuilt it in Python, which is a which makes it which is completely legal in that you've you've created something you've created something a derivative piece of work from something, which means it's it's no longer like copyright claimable. So, that was the first thing that was done. So, go and get that. Obviously, they're DMCA'ing everyone who picks up Claude uploads Claude code open source to GitHub. They can DMCA request those, and they will and have been doing. Uh but the Python library, and I think there's a Rust version as well, I'm not too sure. Those are all, you know, go nuts with those. They're they're all there. But I think it's what we've learned from it that's the the most interesting thing. I I find it interesting how it is for OpenAI and other competitors potentially. Are they like sitting down and looking into the code and trying to understand what they can learn from it potentially to make their products better? Absolutely, for sure. And and and Claude code does have a magic to it. Um the not only do are they are they got a great starting point with the actual model itself with Claude. Like, it's like the Apple Apple ecosystem, you know? The software runs great on the hardware because they have that control of the entire pipeline. So, there's just a a real nice magic the two of it like synchronicity between the two that may or may not be able to be translated to other tools, you know? It's just it's just how it is, how it runs with with Claude. But they'll definitely be looking into some of the stuff. Interestingly though, I saw a video by the Primergen. And one of the things that someone found was [music] if it's trying to understand if it's a negative like if you're frustrated or you're pissed off with with Claude and whatever. And instead [clears throat] of instead of using their AI to actually query whether it's a negative sentiment, they literally just use search and replace like plain old JavaScript to look through whether it's like piss, [ __ ] [ __ ] off, rubbish, crap, and all whatever. Uh you've got it in front see them. Yes. So, [laughter] exactly. And so, you'd think that that is some of the stuff would be actually really really complex or they'll be using the most sophisticated thing, but actually, they're using basic rejects find and replaces on on But it is cheaper, right? And what they do with this to be specific, they are just going to flag the conversation. They are just going to From what I understand, they do use this for analytics. Now, they might be able to go and look at the conversation. I'm not sure about that. Is it uh is it something that they can do, look at your conversation, or is that encrypted? I'm not sure. But it is I I know for sure that they are for the least, they are using this for analytics. To see how satisfied or unsatisfied you are with the uh with the session. Exactly, yeah. And also, there is some leaks that happened. One of them has actually already been within like a day, they already released it, which is this buddy feature. Tamagotchi virtual pet. And I hatched that as soon as I saw it, I was like, "What was this buddy thing?" And you type {slash} buddy, and it hatches this creature. And there's different rarities. It was like this Tamagotchi {slash} Pokémon hybrid thing. If we have a look at some of the other stuff here, we've got Kyros, autonomous demon, always-on agent with GitHub webhooks, 5-minute cron cycles, {slash} dream. It's got auto dream. This is kind of creepy, actually. Yeah, memory. So, obviously, things running in the background, just understanding what's going on and constantly observing your GitHub and and certain things like that. Ultra plan. So, we have ultra think, which expands the memory uh thinking budget that you have. You can you can just type {slash} ultra think, it goes all rainbow, and then it just can think for longer. This is ultra plan. Offloads planning to a remote 30-minute Opus session, 3-second polling, teleport sentinel returns results. So, that's going to be quite interesting. I am a fan. Do you know what 3-second polling means? Whilst it's running for 30 seconds, you're going to be checking in every 3 seconds to see what's going on. Cuz seemingly, you can change that polling number. So, Interesting. [clears throat] I I did use this in T I did use this in the T3 code. Yeah, in T3 code, I use ultra plan. I didn't I don't think it went for 30 minutes, but it did go for I think I saw 18 minutes. And is it remote? Is it done on your machine, or is it remote? I'm not sure. I I I didn't check. I I I used it, and I didn't check too much. Cool. Yeah, so this one's obviously sending off to you know, remote remote server doing its thing. Probably going to have to pay extra for that, to be honest. Tungsten tool. This is an internal thing, direct keystroke and screen capture of virtual environment. Is that like a you know, a bit for a data collection thing, I'm not too sure, but seems like a internal only thing. Voice mode. We already had voice mode. So, maybe it's using Deep Grounded over 3 as its model now instead of something else, I don't know. Um and coordinator mode, multi-agent swarm with workers and scratchpads. Again, something I thought that we already knew about or already have even. Bunch of other stuff, undercover mode, anti-distillation, interesting. So, they've already started to patch that. [snorts] Session memory, auto memory, auto dream, and native client. So, yeah, few things have leaked, particularly that that buddy thing. I'm not too I don't really understand it, to be honest. It just keeps, you know, winging and whining in the background. Um And saying things about my code, I don't really know what. I'm not sure. You know the spinner verbs? So, when when you send off Claude to do something, it says "Thinking," and it also [snorts] uses other words instead of thinking, these verbs. And the list also exists. They have 187 uh of them, and it's just a cute thing to, you know, This is, by the way, Vess found it. And yeah, it is just fun to look at this. I didn't know most of these words, uh but I'm going to look them up. It's just kind of fun. Zesting. Yeah, sketching and puzzling and I don't know, gusting and frosting and [snorts] I don't know, it's just forging and form forming grouping and imagening. Yeah, it's just a bunch of beep booping. It's just kind of [laughter] It's kind of cute, so uh I think they also asked who came up with this, and they they did mention uh I I forgot the name, but one of the employees, and then the rest of the employees from Anthropic uh came together and just added more verbs. So, yeah, it is quite funny. Yeah, and a few other notes that I saw from other creators um covering this were like uh there's a tons of comments, like more than usual, and the comments sound like they were not written for human. They were written for the next agent looking over the code to understand what it is. So, that kind of makes sense, and it is the new era that we are in. Uh and also uh I I saw comments on how it does look like a code that was written all by just prompting. That meaning that it's not like the highest quality code. Mm. Mm. Yeah. But I I'm not qualified to judge that. I mean, you know, it is it's it's the right things or not, ain't it, really? It's it's quite crazy how um how we've just good is good enough. You know? Yeah, I mean I I'm I'm paying 200 bucks because it it is the best. Yeah, I don't know. That's the Claude code league. I don't think there's much um much much more to say about it to be honest. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI