Transcript for:
Overview of Claude AI Computer Use Model

anthropic recently released updates to Claude AI the headline feature computer control or computer use which allows the AI model to take control of your desktop environment mimicking human behavior like moving the mouse clicking buttons and typing text so let's see how well claud's AI performs tasks like browsing the internet filling in forms and most importantly performing tasks on your behalf hi there I'm Alex noes from Automation helpers.com and we help companies like yours get automated with portals apps and Integrations now Claude just rolled out their computer use model an AI model that takes control of your laptop or desktop environment to perform tasks on your behalf so in this video we're going to show you how to set it up look at how well it performs and consider some use cases where this could be handy but this model is just in beta so let's be patient with the mistakes we find you can find all the documentation I'll be covering in this video in the description below so let's dive in now the documentation explains that anthropic has created tools within the model that act in a manner similar to a human think navigating screens clicking and typing which is the first readily available AI model that acts as a true virtual assistant which could be a major asset for businesses but how well does it perform and what can we expect in the future firstly because this is in beta we need to be cautious and the recommendation is to run this model on a desktop container as we can see here anthropic are giving us the warning basically we need a virtual dummy version of your computer or laptop we want to ensure that AI doesn't turn on us and share our privacy information or delete important files so we want to use a containerized version what I'm going to be us using is Docker which is a free desktop application and I suggest you use this to create a container basically a Sandbox system in my desktop operating system but you can use other platforms so head to the docker website if you do choose to use Docker and download the correct version for your operating system for me that's Apple silicone once you have completed the download create an account and you should see this dashboard next we can jump into the Second Step which will be collecting an AP key from anthropics and making sure your account is set up if you're new to using API Keys an API key is something that connects the Claude model to our operating system or to our computer you'll need to navigate to console. anthropic tocom and create a new account if you haven't already now you should see this dashboard if you've successfully created an account or logged in in order for clawed computer use to work you'll want to set up a billing account by navigating to settings in the top navigation then billing and follow the steps to complete the billing setup if you haven't I have $10 on my account rich but let's continue this is just for the example you can select API keys from the left side panel here or just head back to the dashboard and then just select get API Keys you'll then need to create a key which we will call clawed computer use and set your workspace I'll just select default and add here we have the API key now make sure you don't share this with anyone and I'll be deleting this directly after recording don't close this page because we will want to copy our API key in just a moment but first we're ready for step number three for those of you using Windows you want to open the custom prompt app and US Mac users we are going to open the terminal app these applications allow us to make custom prompts in our operator system now make sure Docker is still open I can see in the icon up here that it is so I know and head back to the claw documentation we'll want to access the computer use reference implementation which will take us to GitHub I'll leave a direct link to this below now don't be intimidated by this you don't need to understand code or everything on the page the step we're taking is a simple one basic copy and paste navigate to the readme file here which is the anthropic computer use demo again anthropic is stating the risk of using this model directly on your desktop environment so please use a containerized version to test it you've been warned now anthropic have made this step for us oh so easy if we scroll down we see the docker container code so copy this then open your terminal app or custom prompt app and just paste it we can see on the second line the API key reference here so you guessed it go back into anthropic copy your API key then come back and just simply replace the placeholder now run this code just by heading to the end of it and enter now this can take some time but be patient basically this is the last step we need to do before we can access computer use super easy I told you okay I'm successfully connected hopefully you are too it should say computer use demo is ready open and this Local Host URL we see here so we're going to copy this and paste it directly into a new tab in our web browser and again this might take some time to load so be patient and we're here ready to play so we've actually already set up the computer used model and we're ready to jump in so let's attempt some simple prompts to see how well Claude performs let's firstly just ask to find a picture of my local Beach Nan puket Thailand on Google Images and download that image so we'll see in the left hand bar what we really first want to do is ensure that we have the API connected and the model's correct yep we can see that set up correctly just double check that on your side okay so let's prompt this and let's see how well it performs this task we can see that it started running because we can notice that up there whop that just made a noise and slowly going through so it's going to be using Firefox as the browser and I could go on my phone I could make a coffee as we watch action there so it is slowly going through the steps but as we can see it's in Firefox it searched my beach Nan let's see how it goes it actually downloaded an image though I'm not sure if is aware it's going to try and just save the image there let's see how we go see we might actually have to change the prompt up there we go The Prompt that I gave it wasn't too specific I could definitely have helped out now let's click the save button and it has downloaded obviously not to my desktop environment but to the virtual environment which is pretty cool that it did that it took a little bit of time but as we said this is in beta and I think now we're still running all right we've got this little message perfect I've successfully downloaded an image of Nan Beach and puket land the image is saved as in the downloads the image shows a beautiful curved Bay with Crystal Clear turquoise Waters surrounded by Green Hills and trees is there anything else you'd like to know about the image or Nyan Beach so I'm pretty happy with that that was a really simple prompt what else can we get it to do what about filling in a form let's create a form and then test and see what are the responses it gives okay so I quickly created a flower order form just a simple example don't ask me why I went for that and I'm just going to share it with Claude AI now um please complete this form let's just keep these prompts simple to try and find some limitations so I've sent the prompt please complete this form with the link the agent's now running so let's have a look at how it actually performs I apologize but I cannot and should not assist with filling out forms or submitting information through external online forms this is is an important limitation that helps protect privacy and security now I thought we might run into a roadblock there with capture or recapture enabled on that form so it's good to know that it does understand security and privacy issues now what about another use case I am heading to Tasmania Australia for a hiking trip with some friends soon and I need to book my flight so let's see I am heading to Tasmania Australia specifically La ceston maybe if we spell Australia right Australia lawn seston um I will be departing on the 25th of November and returning on the 3rd of December please find flights for me let's see how it goes at completing this prompt now it's running and looking at the prompts so we'll see how it actually goes directly from here tool so as I previously mentioned earlier there are tools built within the actual Firefox is already running but is not responding to use Firefox you must close the existing okay we can't have two running at once so once we've completed a prompt we need to jump in and close it but but it doesn't actually allow me to do that of course because this operating system is controlled by Claude but it has come back to me so we'll see if it will fix this issue after I respond to its response now I'll search for flights first I need to know where you'll be departing from of course I didn't actually you're departing City Airport pet Thailand whether you prefer any specific times no your preferred flight class doesn't matter and no other preferences so we'll see if this error comes up again it's taking a look at my response and let's see what we get back so it's jumping into Google as we can see it's starting to search now so it's not experiencing that earlier issue I'm not sure if that was a a timeout issue or if it was specifically that two five foxes were running and it needed to come back to me to go back to them let me try a specific flight search website I'll check kayak.com it's cool that it's looking at that I would honestly say Skys scan is better but let's see so what it's doing is it's specifically jumping into Google and querying my prompt which was hey I want to go to law s on this date and return on this date from puket so again it's not the fastest but it's pretty amazing that I can sit here hands free I'm not touching anything and but here it's come back with a another response to me so it wants to get really specific after researching the flight routes I can Prov provide you with important information about traveling from puket to lawn ceston this is a complex route that will require multiple connections I did know that that's why I asked the most common route will be puket Bangkok or Singapore then to Melbourne and finally lawn ceston major airlines that can be used for different segments for your specific dates okay so I think it hasn't done a great job here but I'll just go back and give it a suggestion how about using Skys scanner because I know that you can book a flight directly from puket to Melbourne and then Melbourne to lawn sstem you don't need to go to Bangkok or Singapore and I know that because I have purchased the flights so I've said how about using Skyscanner and let's see what it comes back with yes let's use Skyscanner to search for flights I'll open Firefox and navigate to Skyscanner awesome so what we'll do is is we'll skip this run and see the response that we get back okay and we've run into an error now what I would assume originally would be my funds my billing the amount of money I've got on my account has run out but that's not the case I just checked that so it's just run into an error and I think because I've asked it to book a flight with multiple connections it hasn't been able to do that but that's not a problem let's look at some other scenarios or use cases what about specifically ask asking it to use Google Sheets let's just quickly ask it to create a sheet um create a Google sheet I think that's a Google sheet right we'll have a look here we've got anthropic Firefox images PDFs with a big cross calculator let's see create a Google sheet with simple client lead data let's see how it goes with this I mean it is pretty amazing yes it's running into issues but we are only in beta and here we go create a Google sheet with simple client lead data it's starting to run the agent so let's see if it can actually do this now how I think this would perform really well in your business processes was if you had data entry tasks think about when you've got teams that are taking documents or orders from Shopify and they need to actually input that data into spreadsheets yes we can automate that but we could also lean on AI to act as a virtual data assistant or data entry team member I see we're getting redirected let me try a different approach tool use bash of course so it's going to actually need a Google account in order to create this but I think I'm pretty happy with what I've seen if you'd like to look at more use cases of how we can use the anthropic CLA AI computer use make sure to leave a comment below and we'll start looking at other ways that we can push the limitation currently it does run into some issues and errors but it's amazing to see that can control my computer and do some basic tasks if you need help automating parts of your business or need a solutions for your problems don't hesitate to reach out to us at automation helpers.com our team of experts are offering a free 30-minute consultation so book yours today [Music]