Transcription
Hi everyone. I would like to showcase you the application that Palin Turos and myself have been working in the last one and a half months. Uh, this is called Craft agents, and we open source it so that everybody can use it. But what is really important is that this is an application that we developed to enhance our internal workflows. So, this is our agentic approach, how to improve our workflows internally in the company. But we found it so nice that we wanted to share it with you. That's why we open sourced. And in today's video, I would like to showcase you what are the capabilities of this application.
So, first of all, let's imagine, uh, we start from fresh. If you open your browser and just type agents.craft.2, then you can open our website, and here you can find all the necessary information. Uh, here are the download links for Mac OS with Apple silicon, Mac OS with Intel, Windows, and Linux machines. And also, there is a link to our open-source GitHub repository. So, you can even, uh, download yourself if you are technical enough. You can, uh, create your own version of Craft Agent.
So, when you first download the Craft Agents, this is the window that you will see in the application. It doesn't matter if it's an Apple MacBook or a Windows laptop or a Windows machine or a Linux machine. Our solution is multiplatform. The first thing first, when you get started, you have to select how do you want to connect to the AI. By default, we usually recommend to use Cloud Pro or Cloud Max, uh, subscriptions because that gives you the best experience. However, if you are using third-party applications, third-party providers, you can also select from, uh, the direct entropic using API key, Open Router, Vercel AI Gateway, Ollama, or any custom third-party options that you want to use. In my scenario, I'm going to use my code subscription. It goes through the browser. From the browser, it will authenticate me and give me a code that I need to paste. And now I'm connected. I can click on new conversation here or just click new chat here, and it will open a new chat for me. In a new chat, I can just say hi, and this will automatically start, uh, the discussions and the connection with the AI endpoints. We are using Claude Agent SDK. It doesn't really matter for your use case. If you are interested, this is how it works under the hood.
What I would like to showcase you is what is the whole UI, how it is set up. The UI consists of multiple areas. The first area, which you can see on the left side, is basically where you have all the organizational tools, uh, that you want to need during your work. You have an old chat view, but you can also have flag view, and you can also have statuses. You can also have labels. You can have sources, which I will describe in a moment, and also skills. If you are not familiar with skills, you can also just click on the learn more, and it will guide you, uh, to our, um, documentation, which I will show you eventually as well. When you are here, uh, on the left sidebar, you can also go to the settings. In the settings, there are a lot of options to parameterize and customize the application. I will show you what kind of solutions we have, but like, there are some basic, uh, stuff. You can also select different, uh, themes as you wish. So, it already supports a lot of customization out of the box. We also try to support all the functionality that is available around the globe. So, you can even select, uh, how do you want to send messages? Do you want to use Enter directly or use Command Enter, for example, on a MacBook?
Then, of course, we are thinking about workspaces. So, when you are in Craft Agents, we mainly suggest you to be in a workspace. So, you can add, create, and reuse workspaces across your PC. Within your workspaces, you can define, uh, specific parameters like what model you would like to use, uh, what permissions you want to use. Permissions are super important. Right now, for example, if I open, uh, a new chat window and I say, okay, so now I'm in, let's say, my downloads folder. So, I go to my downloads folder and I just tell the agent, okay, please reorganize my downloads folder and highlight which folders and files could be deleted. As you can see, it works similar way how code is working. So, basically, it has access to your machine. What is happening? There's a nice UI which can tell you what is happening under the hood. You can see the different steps that the AI is executing, and you can even jump into it and see how it did. So, you can see what commands it was running and what was the output. In my case, you can see it was running a command and then it received bash command blocked. So, there was a reason that this command was blocked, and then my agent will realize that the main reason for being blocked is that you are in explore mode. So, to do the changes, it will ask me to switch from explore mode to a different mode. So, now in explore mode, it has a clear picture. It saw how many files I have high priority deletions. So, I can save 23 GB, and then it can see there are some, uh, installed app installers that also can be deleted, and there are some sensitive files that it should keep.
Now, what you can see, I will say, it, please delete the unnecessary files. It will fail to delete it. What's the reason for it? You will see. So, working with AI nowadays has a really important section, and that's called planning. So, planning phase is super important nowadays with working with AI. In explore mode, right now, it cannot execute the commands. So, it created a plan. So, this way, in our solution, you can enforce the agent to go into the plan mode and create a plan for you. Once you read the plan, so you can see what it would like to do, then you can do accept, accept and compact. By just accepting, it means you stay in context, and whatever you were discussing before, uh, will be preserved. When you say accept and compact, you basically ask, uh, the AI to compact the context window and summarize what has been already discussed. And in our solution, it is going to inject the plan again so that the AI knows, uh, what it needs to do. When I click accept, it will automatically switch to execute mode. So, then it can execute all the tasks that I want to do. So, now, for example, if I say accept, you can see it changed to execute mode, and now it is able to execute all the different steps, run through all the different commands, delete the files, move the files, create folders, etc. So, you can see it already started. It created the to-do list, what it would like to do. At any moment and any time, I can just stop it and say, please do not continue, and then I can just interrupt the flow of the agent. Right now, I'm fine with the plan. I accept it to delete these files. So, then it's going to go through and execute it. At the end, it gives me the summary of what it did. And then now I'm nice. We cleaned up quite a bit of, uh, storage on my system.
Of course, I can also do ask to edit. So, if I'm a bit, uh, having trust issues with the AI with the ask to edit, I can do the following. Okay. Now, create a new folder called archive under downloads and move some files there. So, now I'm in ask to edit mode. So, it won't execute right away. Uh, it should ask me, uh, for permission to do the changes on my file system. As you can see, it listed the files and now it wants to create an archive folder. It shows what's the command that it wants to do, and then I can select hello, always hello, or deny, similar to how Claude does, but here for us in a UI, in a much nicer view. Again, it gives me what's the command, the move command. So, these are all the different, uh, files that it wants to move under the archive. So, I will just say, yeah, please move them. And then if I trust, so now I say, okay, it's getting a bit tedious. So, I don't want to ask so many questions. So, I just switch back to execute. I allow. And now, from this moment, it again is going to execute all the commands that it will do without asking questions.
For Craft Agents, connections are super important. The whole application was built to make configuration of connections as easy as possible. And for that, we have the sources menu. In the sources menu, you can see there are APIs, MCPs, and even local folders. But if you don't want to spend time and realize what's an API, what's an MCP, you can just come here, click on the sources, click on the learn more, which will guide you to the documentation that we have online. Let's just do a really simple add source. When you do this, it will open up a popover window with a little chat, uh, inbox. And here you can type really simple with human language, connect to Linear. Linear is a tool that we are using to store our issues and tickets. So, what you can do here is basically ask the agent to connect to Linear. What it will do is going to look up the documentation on how to set up Linear and then it will guide you through the setup of the connection. Perfect. So, it went through our documentation and also did a web search to look up how it can connect to the Linear official MCP server. It is also asking some questions like how do we want to use, uh, Linear, what we want to use it for, any specific teams or projects we want to connect. I just want full access. You can always check what the agent is doing. So, in execute mode, you can see it automatically executes all the steps that it requires. So, again, it reads the config, then it actually creates the configuration file, and then if you are interested what's inside the configuration file, you can also check it, but you can also see inline how many lines of changes it was doing. So, basically, in this scenario, it was adding 18 lines to the config JSON, and then it also generates a guide with 100 lines, and then also generates a custom permission for the Linear MCP server. If you want to see what's inside, you can also have this really nice edit view. You can have multiple options like highlighting. It also shows how many lines removed, how many lines added, and it helps you to better understand what the AI is doing in the background. Once it finished, it gives me the option to log in to Linear using OAuth. It pops up a browser window where I need to sign in and approve. Once I approved it, it can continue, uh, the setup. Now you can see, excellent, Linear is now fully set up. Then I can just enable the Linear, uh, connection in this chat, even, and then ask a question. As you can see, it will start using the Linear tool. You can also see here, sorry, here. So, you can go into details and you can see like what it's actually sending. It might be too much information for some people, but if you really want to dig deep and audit what it does, then you can just go dig deeper. So, you can see there are some tickets which are in progress, and then it returned them. So, this is how you can set up, uh, a source, a connection, and you name it. So, it can be an MCP server, it can be an API, or it can be, um, local folders. Even just to demonstrate how you can connect to APIs, for example, I can also ask it to connect to GitHub API.
I can also select models. So, by default, because I'm using my code subscription, I can select from the code models. So, for example, for the previous setup, I used Sonnet 4.5. Now, let's say I use Opus. And then I can also set the max thinking. So, this way, you can even manage the token usage a bit better as you wish. So, now I'm asking, connect to GitHub API. Again, it is going to ask me some questions. So, some key information and how it works. I will just tell it that I want to list my issues, check on my pull requests, and organize. It's an AI, so it will eventually understand even if I have a typo. So, now it knows what I would like to do and can continue setting up my GitHub. But don't forget, I'm in explore mode. So, it is not going to be able to create the configurations. So, first, it will probably either ask me to switch from explore mode to execute mode or will give me a plan on how to set up the GitHub, as you can see. So, it actually proposed the plan. So, here it says what it's going to do: create the source folder and the configuration, the guide, the permissions, run the test, and then request, uh, the access token from me. So, now I will say, perfect, I like this plan. As you can see, it gets to the point that it requests me to provide an API key. So, I'm going to just provide the API key. So, now I'm providing the API key. It gives me a custom, uh, user form, a custom card to get the API key. Then I click save, and then it can connect, uh, to GitHub. And now you can see that if I add GitHub, then I can ask and do list my issues in the Craft Agent OSS repository. Perfect. As you can see, that's how it works. If I switch to, uh, my main workspace, you can see I have much more, uh, sources. For example, here. And then I can open down the sources can be automatically, uh, categorized into APIs, MCPs, and local folders. I can set up all different, uh, APIs and MCP servers. So, you can just ask the agent, and it will set it up for you. Of course, it may run into some issues, but it usually fixes it itself, and then if it requires authentication from you, it will ask for it in a secure way. It's really important that all the credentials, we usually, uh, set up an authentication card for them. So, it will not be stored in your sessions, and they will be stored in a secure storage, uh, a credential storage on your local machine.
So, now we discussed how sources are working. Let's jump into statuses. You can see I have a lot of statuses: to-dos, ideas to work on, in progress, needs preview, getting pizza. I don't know why, it makes no sense. So, let's change the name of getting pizza. Switch getting pizza status into important things to do. So, now I'm instructing an inline agent to change my config setup. As you can see, it's really similar to the standard, uh, chat window. It will see, uh, what it wants to do. It will explain me what it is going to execute. Rename it. As you can see, now it changed to important things to do, as I say. Okay, just change it to important stuff. Now you see it actually updated my UI to important stuff, and then if I click out, it automatically disappears. So, this way, you can easily, uh, customize your UI to the needs of yourself. However, you can see I left the pizza slice there. So, let's delete, change the pizza slice icon for the important stuff and exclamation mark. So, it's not just able to change the wording, but it is also able to change, uh, the icons for the statuses. So, this is how you can create, um, new statuses. But how can you move, uh, conversations across statuses? Really simple. When you open a conversation, automatically it goes into to-do state. But then with one click, you can select, uh, where do you want to move it. So, with one click, I just move to my important stuff. And now you can see it shows up in important stuff. But I can also use a shortcut using the hashtag. So, now, for example, if I want to move it to in progress, it will jump back to in progress right away. So, you can see now I'm in progress. But let's say, move it back to important stuff. And this is how statuses work.
But we also have labels. So, if you want to organize your projects not just based on their current state, but also maybe the topic that you want to do for that, it's really simple that you can create, uh, labels. For example, let's say I want to add a new label. It works the same way as it were. But what happens if I want it to be automatically assigned the different sessions under the labels? I can go to settings, select labels, and in labels, if you scroll down, there are auto apply rules. So, let's do the following. I would like to assign all sessions that contain bill invoice or what maybe what no bill will invoice in them to the invoices label. Again, it is going to start a new chat because auto apply rules may be more complicated. It actually gives you a full chat, but as you can see, it's using Haiku in the background. So, this is also to make sure that we don't waste too much time on configuration options. Also, these small agents that are spawning, they are Haiku agents, which is the cheapest model and fastest model that Claude, uh, provides. And now you can see it created the auto rule. So, I can also double-check, uh, the auto rule creation. So, you can see it created the auto rule. So, let's say, how does that work? If I go back to my status and let's say here's my imports and stuff. I need to collect my invoices. Do not do anything yet. I am just stating my intent. So, now I will say invoices, and you can see automatically assigned it to the invoices label. So, you can see it's in the important stuff, uh, status, and also assigned to the label invoices. You can also remove the label if you want. But you can see with the auto apply rule, you can automatically get it there. It helps a lot to organize your different sessions into different projects, uh, activities as you like.
What's also really interesting, for example, if you have sessions, uh, with a large number, like myself, and you want to search for something, you can also have filters. But in this scenario, I don't want to filter on anything. So, let's see all the conversations that I have under, for example, to-do. And what I want to do is search. Works the following way. So, I'm looking at Polaton, which is a lake in Hungary. So, I can see where I was talking about Lake Balaton. In this scenario, you can see I was looking for, uh, different weather forecasts, and you can see it highlights in my session list, but also highlights inside, uh, the body of the discussion. And you can see these, uh, titles are not so great. So, I can also do and ask to regenerate the title. You can see it regenerates the title using AI. So, it will look into your content and see what is the discussion about, and then it will update the title accordingly. The search is really fast. We tested it with thousands of sessions also open, and it helps a lot to go through actually, uh, your history and find where you were discussing already some topics. So, it also helps that you can actually work and live inside the Craft Agents because all the information that you want to reach should be reachable, uh, within a few keystrokes. The, the search should be really fast.
And now let me show you how connections work. So, I get actually, I have quite a few connections set up. So, I have my Crop Dog space, Slack, Outlook, Linear, and for fun, Catfacts hooked up into my session. So, let's do some examples. I can ask Outlook to list my meetings for the week. You can see it again goes through, uh, my connected connections, look, goes through the tools that it can actually reach, and you can see it found Outlook, and then it listed all the different, uh, activities that I have. Okay. Now, please check if my weekend is free. My weekend is free. Perfect. So, please create a blocker for Saturday, a whole day, and look for activities in at Balaton, but based on the weather forecast using, and then with the at sign, I can open a quick menu for my sources and also my skills, and I have a skill called Hungary weather. So, I will just paste it. What you can do in a skill, you basically can specify workflows and execution steps that you want the AI to do. So, in this scenario, for example, for Hungary weather, I just specify the locations that I'm usually going within Hungary, and then it should check the weather forecast for those, uh, locations. Now it is going to execute the task. Now, as you can see, it generated for me a plan, but I also want these also add some funny cat facts. It makes no sense, but just for the demonstration that it can actually use multiple sources within the same session without any issues. I just showcase it. Get facts is an API access, uh, added as a source. Now it created a plan. I said it is perfect, but next time do the weekend planning in a better way using a skill. So, please create a skill based on the steps we did. So, now I'm basically just asking the AI that all the steps that we were discussing. So, first, look up my calendar. Do I have a free weekend? If yes, then come up with a nice, uh, plan for the activities, and then, of course, check the weather and give me some funny, uh, cat facts for the weekend. So, these steps I did it in a few minutes, but next time I can just trigger it through a skill. So, now it created me a plan to create a skill. It's called Balaton weekend. So, how to plan a weekend at Lake Balaton. So, I can just, uh, accept it. And now it switched to execute mode. It is going to create the skill. And next time if I again want to have some nice fun at the lake in Hungary, it is going to basically use that skill, which I will show you in a moment. How does it look like? Perfect. Now it created the skills. So, if I go to my skills, I can see actually there should be a Balaton weekend planner. You can see it created the skill. It gave it a really nice icon. It has a summary of what it does. And then here are the details, the instructions. So, what is the workflow? First, check weather forecast. Second, recommend activities based on weather. Add fun cat facts. Fourth, create calendar blocker. And then here you can find the response format. And then there are also some guidelines. Always check the weather first before suggesting activities. Prioritize indoor warm activities in winter months. Include thermal bus options regardless of weather. It's always nice. You can always change it. So, for example, you can ask the AI to modify it as you wish. And for that, it will just use, uh, a cheaper Haiku model.
Let me create a new chat and I show you another really nice thing. When you want to understand things, it's really nice to have a visual representation. So, for example, if I want to understand how does my Balaton weekend planner skill work. Please create a flowchart using Mermaid diagram. In Craft Agents, we also use the beautiful Mermaid diagram. As you can see, it creates really nice visual flowchart, visual representations of whatever you want to do. In this scenario, it created a visual representation of the Balaton weekend planner skill that I just created. You can see it starts the weekend planner here. Then it does a web search to get the weather forecast. Then based on the weather conditions, it will either give me winter activities, autumn, spring activities, or summer activities. And then you can see it also has some ideas on what activities for the different, uh, climate I would do. Then at the end, fetch my cat facts. It also highlights what, uh, command it will use, what URL it would use through the API. Then create an Outlook calendar event named all-day event Balaton trip, and at the end, present the plan to me, actually the user. So, what you can see, uh, we have a really nice visual representation as well, which can help a lot to better understand what is happening. Of course, in this scenario, for the skill, it can be also really useful, but also for more complex skills or if you want to understand processes, then you can just ask the agent to create a nice diagram out of it, and you can view it in full screen. You can zoom in, you can zoom out, and you can even save it.
So, I hope I could give you a really nice, uh, overview of what Craft Agent can do. And let me just quickly also show, uh, our website. Here you can see a summary of what I was basically showing you, how to use Craft Agents, what you can do with it. So, if you really liked it, give it a shot. Download, try to use it. Just, I would like to highlight for the best experience, I would recommend to have at least a Claude Pro subscription, but you can also hook it up with other solutions. But so far, we have been, uh, working with it with Claude, Claude's subscriptions, and it was working really nice, and we are having not just, uh, a lot of work with it, but also a lot of fun with it. So, thank you very much. Thank you very much for your attention, and have a nice.