Building headless automation with Claude Code | Code w/ Claude

Good afternoon everybody. My name is Sedara. Um  I am an engineer on the cloud code team. Um and   today we're going to be talking a little bit  about uh the cloud code SDK and uh the cloud   GitHub action that was just announced today.   Um cool. So a little bit about the agenda.    uh we do a little quick start for the SDK  uh just to give some examples of how to get   started and how to use the SDK. Uh we will  then dive into uh a live demo of the GitHub   action which should be fun. Uh the GitHub action  was built on top of the SDK. So it's meant to be   um a source of inspiration for the kind of things  that you can do using the cloud code SDK. Uh we'll   then dive into some uh more advanced features of  the SDK. Um, and then we'll end with having all   of you set up uh the cloud GitHub uh GitHub  action on your repo so you guys can start using   it and build on top of it. Um, cool. Actually,  before we get started, uh, can I get a show of   hands here? How many people have used cloud code?   Okay, it's a lot of people. And, uh, of the people   who have used cloud code, uh, how many have used  cla or know what that is? Okay. far fewer   people. Uh it's good to know. Um if you guys  don't have cloud code installed in your laptop,   uh that's how you get it. Uh I'd encourage you  to install it in your laptops and follow   along. Uh there will be parts of this uh this  talk that will be beneficial to follow along   and then if you don't want to, you don't have to.   It's all good. Uh cool. So what is the cloud code   SDK? It is a way to programmatically access the  power of the cloud code agent in headless mode.    Uh this is powerful because uh it's a new kind  of primitive and a new kind of building block   that allows you to build applications that just  weren't possible before. Um things that you can   do with the with the SDK are like super  simple things to get started. Like for example,   you can use it like a Unix tool. Uh the Unix the  Unixish tool philosophy is what really makes cloud   code powerful because you can plug it in anywhere  where you can run bash or a terminal. Uh so you   can like use it in your Unix pipelines. Uh  you can pipe stuff into it, pipe stuff out of it,   have like make like complex chains out of it and  stuff like that. You can then build CI automation   on it. So you can have Claude review your code.   Some people actually get Claude to write new   llinters for them too. So like Claude can lint  your code. If there are specific things that you   can't define programmatically, you can get Claude  code to do it. Uh and then we get into fancier   applications as well. So if you want to build  your own chatbot that's powered by cloud code,   that's certainly possible. Um if you want  cloud code to write you code in like a um a new   environment or a separate remote environment, uh  you can build those kinds of applications as well.    Um and finally, the these are a few features.   We'll talk more about the features in the coming   slides. Uh and we have Python and Typescript SDKs  or like bindings for the cloud code SDK coming up   soon. Uh so that should make it much easier for  you guys to consume it and build on top of it.    Uh so let's jump into some basic examples. Uh  calling claude the cloud code SDK is as simple   as just doing claude-b and following it up with  the string that you want to ask claude. So in   this example uh I'm telling claude to write me  a Fibonacci sequence generator. Um and if you   notice I also give it a d-allow tools write which  is uh a way for me to proactively give it access   to the right tool so it can write files to my  file system. Um, and then this is something I   like doing too. Uh, piping logs to claude. So you  can do cat log cat app. log and then pipe that into   cloud-b looking at logs manually. So this is  something I do quite often. Um, and as you can   see, it does a pretty decent job of summarizing  what the log failures were. Uh, similarly, uh, if   you're anything like me, I just can't get myself  to understand the output of if config. I still   don't know what it means, but Claude does and  Claude does it for me over here. Um, and finally,   this is uh this is kind of what makes the SDK  tick, right? It is this is a we have an output   format. If you do d-output format JSON, uh, cloud  code will actually output things or its response   in JSON as opposed to plain text. And, uh, you  can parse this JSON and build on top of it. Um, so   we we'll talk more about details for how this is  what else you can do with this JSON, but uh   I wanted to throw that example out there. Uh let's  get into a significantly more complex example now   uh which is the uh the cloud GitHub action. Uh so  cloud GitHub action was built on top of the SDK   um and it can be used to uh review code. It can  be used to create new features. It can be used to   u triage bugs and so on. Um and uh this is also  open source. So I'll include a link at the very

end of the talk. Uh so you guys can go have a look  at the source for inspiration for how to use it.    Uh but for now let's jump into a live demo on my  laptop. So I have cloned a popular uh small like   uh open- source quiz app for this for the purposes  of this demo. Um and we are going to fire it up   just to see how that works and then we will tell  cloud to build something on top of it for us. So,   I just did an npm start, which opened up my my  shiny new quiz app. It's actually pretty nifty. It   allows you to like choose a bunch of categories.   Uh how many questions you want, difficulty,   definitely easy for me. I suck at trivia. Um  type of questions. And then there's like a   countdown timer. So, we're not going to actually  answer these unless someone feels very strongly,   please shout out the answer. But I'm just going  to just fly through these uh just to show you   guys how this like little quiz app works.   There we go. Uh not surprising. We got a great F,   but that's okay. Um we So this was the little like  demo quiz app that that's open sourced. And if we   look at the issues for this uh for this um uh  repo, we see a couple very interesting ones. Um,   there's one issue that says, uh, we should add  powerups for 50/50 elimination of options and skip   questions for free. Um, because I suck at trivia,  I really like that feature and I want to build   it. Uh, and I before this presentation, I  already installed cloud the cloud GitHub action   on my repo. So, it it's already available. Um, but  we'll go over like how to set that up uh later   too. Um, okay. So, here's the issue. Um it has  pretty sparse details on how to implement this.    It's just literally a feature a wish list really  like a wish feature. It's saying uh add a power   up option in the config 50/50 elimination for the  skip question. It should avoid user points even   even though the user the question was skipped. Uh  and user should be able to configure this from the   config page. So there's a lot of like creative  room for claw to do whatever it wants to do   uh in this case and I'm excited to see what it  actually ends up building. So what I'm going to   do is say at cloud please implement this feature  and comment on it. So it usually does take a four   or five seconds for it to respond. And while it's  doing that for good measure we'll just also take   this other GitHub issue. Uh this is talking about  uh a per question timer. So we saw there was like   a global timer on the quiz app but there was no  per question timer. Uh so that's what this one's   uh talking about. So let's go and say cloud please  build this. And now we have two things building.    Um cool. So now when I get back to this tab you  see that claude responded with a comment on this   GitHub issue. Uh saying that it's working. Uh  it also has a link to the job run which is the   GitHub action run. If I click into it and if  I actually like click on the logs, I'll see   that it's doing a bunch of stuff. You can see all  this JSON being output. This is from the SDK. Um,   so we won't look at the JSON too much because it's  not much fun to parse it manually. But over here   we can see that it also created a to-do list for  us. So claw is now going to actually go through   this to-do list and try to implement uh implement  the uh powerup feature. Uh and similarly for the   question timer uh it's going to do something  something uh similar. Uh one more thing that   we should do here is um there are already a  couple pull requests that have been opened   uh for this repo. Um and let's get claw to  review it or change some of these pull requests   uh just for fun. Uh there's this one which  is change background color to blue. All right.    I actually think I like green better. So,  I'm just going to be like, "All right, Claude,   please change this to green. " Uh, this one is  fairly easy, and I'm pretty sure Claude's going   to do this, but I just wanted to show you guys  that it can also add commits for a pull request   that's already open. U Okay, so this is going to  take a few minutes to run. And while this runs,   uh, let's go back to the presentation,  and then we'll check up on how this is doing,   um, towards the end. Um, okay cool. So, let's do  a little bit of a deep dive on the features of the   SDK. Uh, when you call cloud-ba it has uh  it has no edit or destructive permission access.    Uh, which is great for safety, but it's not great  for actually getting things done. uh which is why   there is a d-allowed tools option which allows you  to preconfigure cloud with uh any permissions   that you think it might need in the future for for  for your given task. Uh so in this case the first

example you see that I've given it permissions  bash permissions to uh npm run build npm test   and the right tool uh which is a good set of  uh permissions because this allows cloud to uh   self-verify what's what it's writing uh and build  uh build your project and test and then   continue writing. Um similarly for MCP if you have  MCP servers configured um you can allow list those   MCP tools as well. So it's a very similar uh  very similar process. Uh then structured output   uh we already saw an example of structured output  both from the GitHub actions logs and uh also the   the little screenshot I showed you earlier. But  there's two modes here. There's stream   JSON and JSON. Um it it's it does exactly what  it sounds like. If you select stream JSON,   it'll actually stream messages to you as and  when they're available versus JSON will just   give you one giant blob of JSON at the end. U  and parsing this JSON and building on top of it   is really how you can make use of the cloud  code SDK um and create features for for   your users. Um and then you can also configure  the system prompt. So you can do d-system prompt   talk like a pirate and you can get cloud code  to talk like a pirate for the rest of your day.    uh which is actually quite fun. If you haven't  done it, I'd encourage you to try it out. Um we also have uh a few user interact interaction  features built into the SDK. Um and what that   means is like the first one is uh is resuming  session state. So um when you uh when you call   cloud-p in structured output or JSON mode, it's  going to return a session ID. Uh and this session   ID is useful because you can then reference the  session ID to go back to the same context state   that claude had when it finished that  process. Uh so by preserving these session   IDs and keeping track of them, you can enable or  like build user interactive features where like   you the user says something, you pass that on to  claude, uh claude returns a response and now you   want the user to give feedback on that response  and that's how this kind of enables you to to   build those types of interactions in your apps.   Um, and then the last one, and this one's actually   pretty interesting, and it's fairly recent,  too. Um, it's d-permission prompt tool. Um,   we talked a little bit about how to give cloud  permissions using the allowed tools flag. And that   requires you to preconfigure them in advance.   Uh, but what if you didn't want to do them because   you don't know what tools cloud wants to what tool  tools cloud would want to use in the future? Um,   in that case, you can use the d-p permission  prompt tool and offload the permission management   to an MCP server. U so you can ask users  in real time for whether they want to accept a   tool or reject a tool and you can have an  MCP server kind of handle that for you as   opposed to trying to predict which tools are  okay and which tools are not. Um, so this is   uh this is fairly recent and we'd love to get  feedback on this if you guys end up trying it out. Uh, okay. Let's go back to our demo uh and  see what Claude's done. Um, all right. So, this is   the power up issue. Uh, we can see that Claude  has actually gone through his to-do list. Um, okay. I'm going to open a PS. There's a  there's a link over here to create a PR.    And I'm going to click that uh and see  what that gives us. I'll actually create   the pull request too so it's easier  for us to review. Um I don't really   know how this codebase works but we'll  still eyeball it just to see if you know   it's doing the right thing. U so you see  some set power up stuff seems all right. Okay. There's like some configuration in the  main component. Um all right. I think   what we should do and what would make this fun  is we should just get this branch locally and   see what Claude did because there's no way  that we can actually figure out what it did   in the short amount of time that we have. So, I'm  going to go back to my terminal, do a good fetch, check out the branch that Claude just  created, and restart our process. Okay.    Uh, awesome. Uh, it looks like we have a powerup  section now, uh, at the bottom of our config page.    And it's a little check box. I like that touch.   Uh, we'll keep both of them on. And, uh, no, let's   select general knowledge. Uh, let's start playing  this game. Let's see what it did. Oh, sweet. So,   we You see it has like this little 50/50 button  on the bottom left and a skip questions button on

the right. Um, I'm just going to with 50/50  because I have no idea what the answer this   is. Does anybody know what that is? D. Okay,  there we go. That makes sense. Cadbury. Yeah. Uh, all right. I'm going to  skip this one and then let's   just breeze through the other ones  for sake of time. Um, all right. Still got a still got an F, but  we got one correct answer. Uh,   which is better than zero correct  answers. Um, and yeah, uh, I guess uh, yeah, it tricked us. Uh, that  was a good one. Um, but yeah, I mean,   it seems like it worked. I think there's  definitely more we could do here. We could   like show how the power like which questions  we use the power upon over here. And there's   like definitely more we can do. But at the most  basic level, I think uh Claude was able to do   u do the task that we assigned it to do. Um  which is exciting. Like this is kind of the   power of the GitHub action because you didn't  really have to run this on your own infra. You   can just literally comment on a thread saying  please build this for me. It uses your your   uh GitHub action runners and just like does the  thing. Um we let's also look at the PR that we   told it to change from blue to green. Uh it's  all hex code. So let's just see what it did in   the commits. So you see there's two commits and uh  Claude has added this last one to switch it from   blue to green. And it did it for all three of the  places where uh we uh where the color was defined,   which is awesome. Um okay. Uh I'm not  going to go over the last one, the question timer,   because we might run out of time. Um, but  this hopefully gives you uh insight into what   the cloud GitHub action can do for you. Um,  let's go back to the presentation now. Okay. So, just as a recap, uh, the cloud  GitHub action um, as it's implemented today,   uh, is able to read your code. It's able to  create PRs for you from GitHub issues like we   just saw. It's able to create commits for you.   So if you already have a PR and you commit or   you comment on it, it can add a commit to an  existing branch or an existing PR. Um it can   answer questions. It doesn't have to do something.   It can just literally answer questions for you.    If you don't understand something, you can be  like, "Hey Claude, how does this work? " And you   can get it to answer questions and it can of  course review your code. Uh the best part of   all of this is that you don't have to take care  of the infra. It runs on existing GitHub runners   which almost everyone has configured if you're  using GitHub actions. Um, so that that's kind   of the really nice thing about this is you  don't have to worry about any of the infra. Okay. Um, so how were actions built, right? Um,  I think I may have mentioned that these actions   were built on top of the SDK. Um, we so the SDK  does form the foundation of how these actions were   built and then we have two other actions on top.   We have the cloud code base action. Uh this is a   thin layer that just uh it just implements  the piece which talks to cloud code and returns   the response from cloud code. Um and then we have  another action on top of this which is called the   PR action. And this action is responsible for  all the fancy things that you saw um on the   PR. So it's responsible for making comments, for  the to-do list, for rendering it the right way,   uh for adding the PR links and things like that.   So um it it's kind of three layers uh in which   it's built. Both the base action and the PR action  are open sourced. Uh so I would encourage you guys   to go have a look uh you know take inspiration  from how that works and maybe that inspires more   ideas. Um yeah. Um yeah and then uh finally uh we  also um uh you guys can install the cloud GitHub   actions today. Uh the easiest way to do this is  to open up cloud code in a terminal uh in the repo   that you want to install it in. Uh and once you  open up cloud code, just do slashinstall github   action. Um and that is going to present you with a  nice flow which uh guides you through configuring   your GitHub action as well as merging it. So the  end the end result of this would be a PR which   would be a YAML file for your GitHub action. Um,  and once you merge that in and you configure your   API keys and things like that, uh, you're off  to the races and you can go ahead and   start tagging Cloud and using Claude, uh, like  we just did right now. Um, uh, small caveat,   if you're a Bedrock or Vertex user, uh,  the instructions are a little bit different and   a tiny bit more manual. So, uh, please have  a look at the docs. the docs uh are pretty   comprehensive in uh in in helping you set up  uh the GitHub action for both bedrock and Vortex.

Uh cool. Finally, uh resources. Uh these are  resources for things that we've talked about   today. If you want to snap a picture, uh go  ahead. Um the the open source repos for   both the base action and the cloud code action  are here. Um and we absolutely love your   feedback as well. So if you guys have any feedback  on the SDK, on the GitHub action or on cloud code,   uh please go to our um public cloud code  GitHub repo and file an issue there and   someone will have a look um and get  back to you. Um cool. That's all I have   for today. Uh thanks for joining me and I  hope you guys have a good rest of the day.