Claude Auto Gives Itself Permission to Do What It Wants

I'm surprised that nobody else came up with this before. It took so much. Okay, you go on and I'll explain. Command AI all familiar with an actual designer, I don't know. With when you're coding, when you're vibing out on Claude code, it's asking, "Do you want to perform this action? Do you want to perform that action?" And it could be quite, you know, annoying. You can set permissions and things like that in place, but it could still be quite annoying and you can you know, work through them. But progressive disclosure is is a an important part and safety is an important part of AI, right? So, to bypass this, dangerously skip permissions is one of those things. But auto mode is an alternative to that. Today, we're introducing auto mode, a new permissions mode in Claude code where Claude makes permission decisions on your behalf with safeguards monitoring actions before they run. And it's a research preview only on the team mode. So, to me, this is basically an AI judging whether it should or shouldn't be allowed. You know the Obama giving Obama uh medal. You tell [laughter] your AI You can You can give your your own some permission, so [laughter] It will be really fun when it starts to hallucinate [laughter] giving itself more and more permission. That would be really fun like getting in a loop. But I'm surprised that nobody else came up with this before because it kind of like jokes aside, it kind of makes sense. Like having a different agent being like very careful, maybe it's like the security agent that is like, "Yeah, sure you can you can do this. Yeah, sure you can do that." And it it's talking usage. Depends what you're doing cuz there's extra token usage at the end of the day, right? Um cuz you are right, but then do people really want to be paying extra tokens for a decision that they're they People are scared of AI, right? It took It took a while to get to where we are now because people are like, you know, "No, I'm not going to let AI do this for me, that for me." But now we're slowly coming around to AI being quite actually all right. So, yeah, whilst I agree, I think it needed to it needed to get to this stage. It's not like people didn't know or didn't want to. We We've wanted to for quite a while, but Claude code's default are purposely conservative. Every far right bash command asks for approval. Some developers choose to bypass permission checks with dangerously skip permissions. Auto mode is a middle path that lets you run longer tasks with fewer interruptions while introducing less risk than skipping all permissions. Before each tool runs, a classifier reviews it to check for potentially destructive actions like mass deleting files and sensitive data exfiltration or malicious code execution. So, I don't need to labor on this point any off any more really. I think you know what it does. It's just an auto approval thing. So, yeah, only available for teams. I'm not I'm not part of a team. I'm alone I'm a solo dev. I'm a lone lone wolf they call me. Uh so, I can't show you how to do this, but it's a simple flag for admins to disable auto mode. Um kind of a double negative there, so it's quite hard to understand. Or developers run Claude enable auto mode. So, it's actually a flag when you run Claude. Um I saw my friend on GitHub today created a uh process that auto does this all the time. So, if you forget to run it, whatever, it's not quite nice. You could just set an alias, I guess. You could just set reconfigure Claude to run Claude enable auto, but yeah, that's kind of it. I want to understand permissions a bit more, if I'm honest. I want to cuz you can set permissions up per project. And but they're this crazy, you know, it's not It's not like natural language. There's a format you've got to follow. Maybe I can even show you. Yeah, so it's this it's this crazy thing. It's like bash and then npm and and you know, web fetch, all of this stuff. But like you, it just feels like like there must be a way. I mean, technically, you could probably do this. And that will allow you to run any bash command, very very dangerous. But I think it's just so specific. Like if something changes, let's say for example, a write happens, uh I don't even know how to explain it, but Can Can you select all select all and just do a star for a for a moment? Like just everything in that list. That Is that what you mean? Allow everything. I did that. that's dangerously skip permissions, right? Yeah. I It still asked for permission and I don't know what I got wrong. Well, Maybe I'm missing something wrong. exit Claude and then enter back in? Could it have been something like that? 100% yes, multiple times. All this to be said, I want to understand permissions better and I want to make a video on it. I want to add it as part of my my masterclass thing that because I've I'm still like you, just a bit like, "Huh?" You know, like I'm setting things and I'm still getting the whatever. Or you just go full auto mode and let the good times roll, baby. 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