Claude in the Box - Use Anthropic Agent SDK in a Sandbox

Hey everyone, I want to show you something super  cool that I've been playing around with. Now,   I know we all love our coding agents. It just kind  of goes off and does stuff, right? It asks when   it needs permission and it has all the access to  those lower level things like reading and writing   files and all sorts of shell commands. Did you  know that Anthropic lets you run that yourself?   Well, guess what? You can run it as an agent with  code. And the best part, you can host it yourself.   Here, let me pull the docs up real quick. So,  I'll link these in the description as well. Uh,   this is how you host the agent SDK. And what you  need is you need to have a sandbox, right? And so,   if in case you missed it, uh, we have one. We  have a Cloudflare sandbox. It's awesome. Uh,   also links in the notes here. Uh, and that  means that you can run your agent autonomously,   no sweat. And I'm sure like all of you, writing  code isn't the only thing that you do, right?   I make tutorials and I make videos like this  one and I would love to kind of have that same   sort of agent, right? Help me out. There's always  stuff that slips through the cracks. Maybe it's   missing links or it's unclear steps, whatever you  name it. So, I had this shower thought. What if I   could feed an agent a tutorial and have it review  everything for me? And this is that thought. Now,   this is totally work in progress, but  aren't we all? I'm running locally. Uh,   which is awesome because I can iterate on this  quickly conceptually. So, what I'm going to do   is I'm going to toss it a URL. Let's see. Uh,  here's the AI search tutorial and I'm going to   click check tutorial. And you'll see that it says  starting the claw in the box. It already has gone   and got a sandbox. I'm again running this locally.  So, it's using Docker and it's going to go get a   repo and it's checking that out. And I made the  pattern pretty extensible because I am pretty   sure that I'm going to be using this a lot. So,  while this is cooking, let me show you some code.   So in here we will see that we go and just get  the sandbox, right? So we have this get sandbox.   It's coming from this Cloudflare sandbox. I  get the sandbox. I'm giving it a unique name.   I don't I don't need to uh I can name it name it  something but the name is unique to it. So I can   get a hold of this sandbox later uh if I need  it. I'm going to store that sandbox uh locally   because I'm going to use it. And uh I with that  sandbox I pass across environment variables,   right? So I need to have my anthropic API key.  I need to have a Cloudflare account ID and API   token. I'll show you why that is here in a second.  Uh, and so I'm using Hono. I'm using the awesome   Hono library and I'm doing this stream text. So  what I'm doing is I'm streaming out. You saw those   lines were coming out one by one. Uh, I am saying  what it is. Uh, each one of these sandboxes have   a a work a workspace and I'm just making sure  if it exists. I don't want to overwrite it in   case something was there. Uh, and then I just do  this get checkout. Right? So I'm checking out this   repo URL. Now, I want to show you what that repo  URL looks like. And I'm telling it to go to that   cloud agent directory. So, uh, and then actually,  let's just go a little bit further. Uh, I I go I   get in there, I do this npm install, right? I get  all the dependencies and then I go ahead and call   this npm start and then I pass across that URL  that came through because what I wanted to do was   I wanted to go and run this agent repo. So, let's  take a look at that agent repo. So, there's a main   function and it comes in and basically this is the  prompt here. So, so what I've done is I said first   use the skill right now because this is its own  repo. I can actually run it locally on my machine   and make sure that it's doing what I want it to do  before I run it in the container approach. And and   the way that this is working is I'm I'm running  this prompt here, right? I'm running this prompt   and you pass it to this uh thing called query and  it sends the messages. Uh you tell it what it's   allowed to do and I'm using skills. Now, if you  haven't used skills yet, Anthropic just released   this pretty recently, and I'm really digging how  I can just put things where they belong, and it   works. So, uh, right. So, I'm saying that it can  use this allowed tool of skill, and in my folder,   I have thisclaw skills. And I have a couple of  skills in here. So, this, uh, first skill that's   in here is this markdown fetch. And I have a file  called skill.md. And one of the cool things that I   was able to do is sometimes you could just do a  straight web fetch of a page, right? So it gets   what happens when you do that though is it gets a  lot of surrounding garbage. And the default tool   actually if we look at the default tool called  web fetch, uh it doesn't currently support   websites dynamically rendered with JavaScript.  And I want that, right? I want anybody to be   able to drop their tutorial in there. Uh and so  what I wanted to do is I wanted to convert that   page to markdown. So what I did was I made that  skill, right? I I I discussed that skill and I was   just kind of playing around here. This is really  powerful. You can kind of just say what you want   to have here and it will check that there's these  these variables and I gave it the endpoint, right?   There's an endpoint for it and I kind of explained  what happens. And I straight up just told it that   I want to save things in a specific location. And  then the other skill, this tutorial checker, comes   in here. And check this out. You just kind of  describe what it is that you want to have happen.   I love living in the future. All I just told  it like this is how I want you to evaluate it.   Super super cool. So, uh, so if I flip back and I  look at it, I say I want to use the skill markdown   fetch to fetch it from the URL, right? I pass  the URL through and then I want to use the skill   tutorial checker to analyze the markdown. So, so  it's using those skills and if I come down here,   we'll see that I am just going ahead and I'm  writing out to the console to standard out.   I'm writing out the message, right? because on the  other side I'm going to grab it. And I wanted to   make sure that I redacted some things. So I put a  little safety in there. Let's see how things are   going. So I'm going to pop back. Uh cool. It's  it's it's outputting some stuff. Uh and it's   it's seeing that it has the Cloudflare account  ID. It's making sure it did a curl. It's got and   gone and fetched it. And so now it's stored that  file. And now it's about ready to start writing   that review. Right. So let's go ahead and uh let's  look at what it looks like on the other side. Uh,   so where I'm using that, let's get back into  my my other repo here, the claw in the box. So   I've I've called npm start and I've passed the URL  and I said stream equals true. And so I have this   function that will go ahead and grab this. Uh,  it'll pull it out. And we can see that uh we have   the standard out and we're doing this JSON parse  of that data. Uh, the user thing is pretty noisy,   so I turn that off. So you'll see that the type if  the type system, it returns back the session ID.   And I could actually write an interface that sends  data back in, but I don't need to in this case.   So, uh what happens then is we just go through and  I get the text out um and I make sure that I show   the tool and I show what's happening uh there.  And then I just handle very at the end, right?   This is specific to my app. I uh wrap things up  here and I say I want you to go read that file,   right? Because a sandbox is able to read files.  So, I'm going to read the file and then I stored   it up in KV. So, so, uh, I stored it so that let's  go take a look really quick at what came back. So,   uh, it finished. And so, I've got here's the  content that was fetched, right? There's the   that's the markdown. Uh, and then here's the  review of the tutorial. Uh, really nice. Uh,   I love this. It It's awesome. I can sit  here and I can read and see what's missing,   right? I can come here and kind of think about,  oh, that's an interesting take on it. And I can,   you know, improve this as it goes, but this is  awesome. This is totally awesome, and I need this,   right? I need this. I need some other eyes to to  take a and of course the best part is you can take   this idea and run with it. So one of the things  that I was thinking about as I was literally   talking to you about this there's nothing that  is stopping us from spinning up a sandbox and   having it try to run the tutorial steps and  see if they work. Isn't that cool? Imagine   the QA possibilities. Now keep me posted if you  build that. I think I'm going to go give that a   go right now and I'll share too when I get it.  Big thanks to Anthropic for making all of this   possible. And now you've got your own claw in the  box to try out. Have fun. Thanks for hanging out.