294 | 3 Surprisingly Simple Ways To Connect Claude Cowork To External Applications, To Automate Everything In Your Business with Isar Meitis

Can Claude really run major parts of your business… if it can’t even be triggered from the outside world?

Turns out: yes—but only if you know the workarounds. In this episode, we tackle one of Claude’s biggest limitations and break down practical ways to connect it with the rest of your business ecosystem. If you're building AI-powered workflows and feel like you've hit a wall, this episode may save you weeks of experimentation.

The takeaway: don't wait for perfect tools. The smartest operators build systems around constraints. You'll learn three practical approaches to make Claude work with external platforms—and how to choose the right one for your use case.

In this session, you'll discover:

• Why Claude's inability to accept external triggers creates operational bottlenecks
• How a LinkedIn content machine generated a 250% increase in engagement
• How ClickUp, N8N, Claude, and Gemini can work together in one system
• The pros and cons of using API-driven workflows instead of Claude Cowork
• How Scheduled Tasks can act as a substitute for real triggers
• Why Claude Code Routines may be the closest thing to a complete solution today
• How GitHub repositories unlock cloud-based AI workflows and reusable skills
• Which option makes sense depending on your business needs and technical complexity

About Leveraging AI

• The Ultimate AI Course for Business People: https://multiplai.ai/ai-course/
• YouTube Full Episodes: https://www.youtube.com/@Multiplai_AI/
• Connect with Isar Meitis: https://www.linkedin.com/in/isarmeitis/ 
• Join our Live Sessions, AI Hangouts and newsletter: https://services.multiplai.ai/events

If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!

Transcript

GMT20260518-193442_Recording_avo_1280x720 0:00

Hello and welcome to the Leveraging AI Podcast, the podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business and advance your career. This is Isar Meitis, your host, and today's episode was triggered by a very interesting conversation that happened in the recent Friday AI Hangouts, which we have every Friday at 1:00 PM Eastern. By the way, it's open to the public. It's an amazing community of people who just share what they learn and different things about AI implementation in business. If you wanna join us, you can find a link to that in the show notes and just join or send me a message on LinkedIn and I will send you a link on how to join. But anyways, we were talking about Claude Cowork and how to connect it to other applications, and this came out twice actually in the same call, and I had to explain it twice, which made it very obvious to me that people are interested in learning how to do that. So what we're going to cover today is how to use Claude Cowork and trigger it from external applications, or how to use external applications in conjunction with the knowledge and skills and other capabilities that you have inside of Claude Cowork. So let's get started. And what I'm going to show you today is three ways to trigger Claude Cowork or to use Claude in conjunction with other things that are happening in other applications. I will be sharing my screen, but for those of you who are either driving or doing something where you cannot watch the screen, then don't worry about it. I'm going to explain everything that we are seeing. So Before we start, let's talk about what is the problem and what is the challenge. And the truth is Claude is amazing. It is literally running two and a half businesses for me right now with a growing amount of functionality every single day. So roughly every day, I build a whole new process that AI agents run on my behalf across accounting and finance and operations and customer service and the training that I deliver. And practically every aspect of the business has more and more fully autonomous or almost fully autonomous capabilities, and I decide where I draw the line in the sand. So Claude is amazing, but the biggest limitation that Claude has that literally drives me crazy, and probably many of you as well, is that there is no way to trigger Claude from the outside world. Meaning whatever amazing tools and functions and skills and plug-ins and automations and agents and all of that that you build inside of Claude, you can trigger from Claude, and Claude itself can trigger other tools like N8N and ClickUp and all the tools that you have connectors for. But you cannot trigger processes inside of Claude from external applications. So that's the limitation. The solutions are three different solutions, actually three and a half, because the last solution you'll see is actually two different versions, exist. None of them is perfect. Each and every one of them has pros and cons, and the last one is going to be the closest thing to a full solution. I assume I'm not the only person that's complaining about this, so I assume Anthropic will come up with a solution. But for now, I'm going to show you three different options on how you can do that. But before we get started, I want to give you a specific example of what I mean by triggering Claude from the outside world. So one of the capabilities that I developed is what I call the LinkedIn content machine. The LinkedIn content machine, and those of you who are looking can see that has a very complex process with multiple components, with multiple agents and multiple N8N processes. But what it does as a, on a nutshell, is it's learning from the best people on LinkedIn on what they're doing, how they're doing, how long they're writing, when they're posting, and so on. And then it tailors my content, what I talk about on the podcast, what we talk about on the Friday AI Hangouts, what I share on other channels, and it takes my content and then applies what it learns from the best performers on LinkedIn in the AI space and creates the output for me. So I call that the LinkedIn content machine. Again, those of you who are watching can see all the different steps that it's doing in order to deliver this amazing outcome. The outcome, by the way, is since I launched this, I saw a 250% increase, so three and a half x improvement in engagement across what I'm doing on LinkedIn. And I did post daily before that and engaged daily before that, and I'm seeing significantly better results right now, so it's working. By the way, if you are watching this, what happens when I put my mouse over any of these components, it actually shows me all the inputs and the outputs. So as an example, now we're looking at the strategy optimizer, and when I put my mouse over it, it is showing me that it's getting information from the LinkedIn content analyzer, from the content performance analyzer, from the content strategy generator, and that it pushes information into the LinkedIn performance tracker. So you can kind of see how that works, and it also tells me exactly what this component does. So if you wanna learn how to build things like this, by the way, any kind of automation for anything in your business, this is exactly what we're teaching in the multi-agent orchestration course that we have launched recently. We've already sold the cohorts for May, June, and July, and we're now selling the cohort for August. So if you want to learn how to be able to automate literally any digital work in your business and do it at scale without increasing the amount of people that work for you, even if it's just you, then come and join the next cohort because they're selling out fast. If you join between now and mid-June, we have $100 off. We had $500 off for the first session, uh, $200 off after that, and now we have only 100. And once we get to the middle of June, there's going to be no discount. And probably after that, we're gonna increase the price because other people teaching this are charging between $3,500 and $5,000. And I can guarantee you our course is incredible from the practical perspective that it provides, because I've taught this to companies as workshops, and I see the results in my meetings following these workshops. By the way, if you're looking to do this as a workshop for your employees, reach out to me either through my email or through my calendar or through LinkedIn. There's a lot of easy ways to find me, and I will gladly talk to you about this as well. But back to the episode, uh, this is the LinkedIn content machine. So what does that have to do with everything we talked about before? So this is the process of the LinkedIn content machine. Before I show you how this connects to what we talked about before, how does this connect to Claude? Let me show you the outputs. So this is an example for a post that I posted on LinkedIn. You can see it has pretty cool graphics. It's all aligned with my brand guidelines. This is the post itself. So it writes the post, it creates the images, and you can see that this post, uh, received over two thousand impressions, which is not a huge number, but it's definitely not a bad number either. Here's another example of a post that it created on its own, and again, it's using actually my language and my content. It's not making stuff up. It is using what I post on other channels, and then it repurposes it based on best practices that it's learning on its own on any channel that I want. This one, again, has a detailed post and the graphics, and it generated eighty-seven hundred impressions on LinkedIn. So how does that connect to how I started this session? Well, the way I manage this is through a ClickUp board. Those of you who don't know ClickUp, ClickUp is like monday.com or like Asana or Jira or Notion. It's basically a Kanban board or a list in where I can see all the different tasks, and then those of you who are watching can see the board. The board has multiple stages from plan to prompt review to, graphic in progress, graphic ready, graphic for review, approved, and posted as different stages. And the cool thing about this, all I have to do in this board is verify, review, and comment, and when I review and comment or move it to specific statuses on the board, then the AI takes over and does whatever it needs to do. So it generates ideas for the post. I review them, I comment on them. It generates ideas for the graphics. I review them, I comment on them. It generates the graphics. I review, I comment, and then in the end, I have an approved post. I'm working in a regular board as if I had an actual full LinkedIn marketing team or any other social media, because I can repli-replicate this easily to other, socials as well, because it's learning on its own what the best practices are. But I work with it through this board. I don't actually engage with Claude anywhere other than writing comments in the chat of the specific tasks. So the problem to connect it how do we start it, there is no way to trigger a Claude Cowork skill or any other thing inside of Claude Cowork when things change inside of ClickUp or Notion or Jira or Monday or emails or whatever. It knows how to read it, but you have to initiate the process. And so how did I solve the problem? How does this thing work? Well, there are three different solutions, and I'm gonna give you all three very, very quickly in here. The first one is I don't actually use Cowork, which is what I've used in this particular case that we're looking at right now. I use different API calls to different AI tools that are triggered through different statuses of my board. Again, you can connect this to anything else. Why do I use N8N specifically versus other automation tools like, Make.com or Zapier? Well, the reason are twofold. One is you can self-host N8N, which is what I'm doing, and then all my automations, and there are hundreds of them running every single day, cost me seven to ten dollars a month, so I'm not paying per automation. And the other reason is that Claude is very good at creating these automations on its own, and so I don't need to create them at all. So if you look at the specific example that relates to this particular process that we just saw, what I'm going to open is the automation that is one of the automations that runs behind this board. So what we're seeing right now is that it gets triggered by a ClickUp status. So every time something hits, it triggers this process. Then there's a filter that finds specific things. Then it gets the task details, it gets the comments details if I wrote any comments. It extracts the information from the comment, and then it builds a Gemini prompt. So you can see here, if I'll open this thing, that it has a bunch of instructions. Some of them are static, such as, the Multiply brand identity and the visual styles- categories and the mandatory rules for every prompt and so on, and some of them are dynamic. The dynamic parts connect to other things that I have in different places in order to make it relevant for the specific prompt and in order to continuously learn over time. So it learns between one prompt and the other. But this is a call for one AI. Then there's a call for Gemini. So then it takes the prompt that was created using other AI tools, and it sends it to Gemini to create the output. Then it extract the generated prompt, so it can share it to me on the board. That also happens with another AI call, and then it actually creates the response inside the task itself. So it's either posts the prompts in the comments and it sets the status for review, or if there's an error, it tells me that there, there's an error and what the error is. That's it. Now, to run the entire board we've seen previously, there are several of these that are running, and it's not actually connected to Claude Cowork, but Claude Cowork created everything in here, including the workflow itself, including the prompts to extract the context, including the prompts to Gemini, including the prompts generator that generates the prompts to Gemini and so on and so forth, and it's just deployed through this methodology. So let's talk about the pros and cons of this option. So the pros is that it can literally access any AI at any time, right? It is, connecting through the API. The other benefit is that it can connect to more than just Claude. As you saw in this particular example, it calls Claude to write the prompts, and it calls Gemini to generate the images. I have another one that calls ChatGPT to generate images, and I can call Grok, Grok or Chinese models or anything that I want. So these are the benefits. The disadvantages is that, A, it's more complex to set up, even though Claude is the one that's actually creating it for me. It's a longer process than some of the other options we'll see in a minute. Uh, you are paying for every token, right? So this consumes tokens rather than using your subscription to whatever tools that you already have. It does not have access to skills and plugins from Claude, which is a big disadvantage, right? So if you build an entire universe of skills and plugins and capabilities inside of Claude, then it cannot work in this because you're literally just calling the Claude API or any other API, and it does not have access to your files that exist on your computer that also might be relevant context to what it is that you're doing the way you're running it in Claude. So these are the pros and cons of this one. Option two is scheduled tasks inside of Claude Cowork. So what are scheduled tasks? When you are in Claude Cowork, all the way at the top, you have the selector between Chat, Cowork, and Claude, and then you have New Task, Project, and Scheduled. It also shows as the very first list section as the scheduled task that I'm running. You can see I have many, many, many of them. Some of them are running on a very high frequency. So what do these things do? When you go to Scheduled and you create on a new scheduled task, it allows you to either create it with Claude by providing a prompt or set it up manually. I actually usually don't do this or the other. I just go to Claude and have a regular chat with Claude and ask Claude to just create the scheduled tasks for me. But what do they have? They have a name, they have a description, both of these are just for you, and then they have a prompt that tells it what to do. So then it will do the thing that you tell it to do. You can choose a specific model, you can leave it to the default, you can work within a specific project, so the folders that you have on your desktop, just like it works regularly. And then the frequency can either be manual, hourly, daily, weekdays, or weekly. If you set weekdays, it will allow you to pick whichever, uh, hour you want, and the same thing with every other selection you make. So if you say Daily, then you can define whichever hour of the day you want it to run. It does have a warning. If you're running it on peak hours, it's going to tell you that you're consuming more tokens than you should if you run this off-peak hours. So just, uh, food for thought and something to remember. So what is this actually doing, and how does it actually connect to what we did before? The reality is this is not really a trigger, meaning it's not getting triggered by any external system, but you can poll whatever other tool you want because now you're triggering it from Claude on whatever frequency you want. As an example, and I'll show you something, I have a proposal auto-trigger, and you can see that it ran 195 times. It ran more than anything else on my different scheduled tasks. Why? Because it runs every single hour. Why does it run every single hour? Because I have lots of calls during the day. At the end of every call, it goes and check if somebody has requested a proposal or if I suggested one, and then it runs an entire sequence with multiple different components that queries about this company, it runs a whole research, understand what they were requesting, connects it to different offerings that I have, and writes an amazing proposal. Then it applies my brand guidelines, then it writes a draft email. It up- it updates my CRM, and so on. It does all these things, all the stuff that a salesperson would have done, it does. But to do this and get the proposal in the fastest turnaround possible, it actually checks every single hour whether there is a new request for a proposal. If it doesn't find anything, it just stops, and if it does, then it will trigger the entire process. So while this is not really a trigger, you can use it in lieu of a trigger and still get the benefit of being able to use whatever is in Claude connected to what's happening in other platforms. And so what are the pros of the scheduled tasks? One, it takes two minutes to set it up. You literally ask Claude to set up a scheduled task for you. You tell it what you needs to do. You can bounce ideas back and forth until you finalize what it needs to do, and then it will run regularly. The other huge benefit is that it has access to all your Claude universe. So every skill, every plugin, and also every file that Claude has access to in the folders you connected to it, it can use and it can see, and it can manipulate for the needs of that particular task. And this is a huge benefit compared to the previous thing we looked at. The disadvantages or the cons are, first of all, it's not a real trigger. It needs to poll whatever you tell it to look at on whatever cadence. This could be, again, every hour, every day, every week, whatever you need. In many cases, that's enough. So I have several different things who run just once a day to do a cleanup for me. As an example, go through all my new files that I created and everything that is an older draft, it will delete and stuff like that. So there are things that can run on daily basis or weekly basis or monthly basis. I need to create a monthly accounting report of everything that happened that month. This can run with all the different skills and all the automations and everything that I've built because it will just trigger it from here based on a specific time. Now The other disadvantage is if you run it on a high frequency, you are using more of your subscription, right? So if I'm running a long, complex task once an hour, including on peak hours, I'm going to be using more tokens from my subscription than just me using it. But again, the solution is simple, just reduce the frequency and run it not at the same level, and then there's usually no problem. Or just pay for the higher subscription if it's worthwhile for whatever it is that you're doing. In most cases, it probably will be worthwhile. The two biggest disadvantages is, one, your computer has to be running for this to run, right? So it runs locally on your machine, and if your computer is off, if you have a laptop and you're carrying it with you to a meeting or you're in the airport or whatever it is, then it is not going to run. And also Claude Desktop, the actual application that runs Claude Cowork has to be running as well, which means your computer needs to be running. It can't go to sleep, and Claude needs to be running for this to actually work effectively. Now, it will tell you every time it's not running, and you can trigger it manually straight from Claude, but still it does not do exactly what you wanted it to do. So what is the next option? The next option, which I find it to be maybe the coolest and most relevant and relatively new solution, is what's called Claude Code Routines. So what are routines in Claude Code, and why are they better or different than just scheduled task in Claude Cowork? When you switch from Cowork to Claude Code, then the options become New session, Routines, and Customize. If you go to Routines, it is very similar to what you have inside of the scheduled tasks of Claude Cowork, but with one huge difference that makes a very big deal. So if I click on New routine right now, you can see I have two options. I can choose local and I can choose remote. If I choose local, it is practically almost identical to running Cowork scheduled tasks. I can give it a name, a description and a prompt. I can select a folder, and I can define how frequently it's going to run, whether manually, hourly, daily, weekdays, or weekly, and I can choose the time. That's it. It looks almost identical, and it does the same exact thing. So if that's the case, why just not run scheduled tasks? And the reason is there is no difference. You can run scheduled tasks, and it will be the same thing. You need to connect it to something that it will check on whatever frequency and will be triggered this way just based on a time initiation or trigger. However, if you click on New Routine and you choose Remote, a slightly different menu opens. You still give it a name. There's no description in this particular case, and then there's the instructions. The instructions are still the same. But there's a few things that are dramatically different. First of all, you can see that it says Select repository. A repository or repo for short is a place where you can store information and host it on GitHub, and It allows you to do a lot of things. The main thing is keep history and be able to roll back to previous versions of everything you have. It was built for coders and companies write code, but you can use it for anything such as your entire Claude folder structure or specific folders or specific component on it that you can back up into GitHub, which I do for each and every one of my folders. There's a whole process around that, which we're not gonna get into on exactly what I'm backing up and where, and what I'm not backing up because it's sensitive and so on. But I'm putting that aside for a second and just letting you know you need to think about a few things before you back up everything to GitHub. But once you connect a repository from GitHub, it is actually done, so you can do something else, which we're gonna talk about in a minute. But this one, the Claude version of this Claude routine has three different triggers. One is schedule. This is exactly like the other schedule that we talked about before. The second one is a GitHub event, which to do this, I first of all need to connect a repo. So let's connect any repo, doesn't really matter, and then you'll see I can choose GitHub event. What kind of events can I do? Well, those of you who use GitHub know all these different definitions, such a, pull request or PR for short, and then release or publish, and then issue opened, and then there's a lot of other options and filters that you can apply, which is not something that you should care about unless you are a heavy user of GitHub. But this allows you to trigger different things, such as doing a code review or sending emails to people or creating summaries of everything that was released, whatever you want that related to GitHub, which is probably not what you want to do. The third component is the one that is the most interesting, and that is an API. This allows you to use an API call to trigger whatever this functionality is, which means this is the missing link that was missing before. You can now call this routine from an external source using a simple API call that is a POST request that you don't know how to write, most likely, and I don't know how to write, but Claude does. So if you build this routine and you tell to Claude, "I wanna be able to trigger this through the API," it will know what to do, and now you can trigger this from anything that happens. As an example, I can use anything that happens on my CRM, on my ERP, on ClickUp, on whatever other third-party tool to trigger whatever I want in here. So this is already a huge benefit, but I know what you're thinking. This still does not connect to your skills or to your plugins, or your entire Claude universe is not here because this runs in the cloud and Claude Cowork runs on my computer. Well, this is where it gets really cool and really interesting. If you take your skills folder, whether the main one or a small one that you have created, and you back it up to GitHub, and this is the folder that you're going to connect to, so I just showed you that I connected a GitHub repo into this thing, any skill that exists inside of that repo is now going to be available to this routine. So let me again go back and explain. Skills allow me to do magical things inside of Claude, but Claude lives locally on my computer, and there's no way for me to connect and use these skills from an external source. This enables to do that. Meaning, if you learn how to back up whatever folders or folder from your computer to GitHub, which will take you five to 15 minutes to figure out, it's really, really easy, and again, Claude can walk you through it. You can now back up all your skills or the ones that you need together with additional reference files that you need to do a specific process, such as examples or formats or, styles or guides, whatever you need in order to do this particular thing you needed to do. If you upload all of that to a specific repo on GitHub and you connect it to this routine, now this routine that runs on the web can do all the different things that Claude Cowork can do on your computer that are related to that task, because it has the skills and it has the other reference files all available to it through GitHub. So let's talk about the pros and cons of routines. So the first version of routines, as I mentioned, the local version of routines, is just like running Claude scheduled tasks, and we talked about this just a few minutes ago. But the cloud version has basically everything we wanted. It can be triggered based on a schedule, it can be triggered based on a GitHub repo event, which is less relevant to the vast majority of us. But the most importantly is that it can be triggered by an API call that Claude knows how to create, meaning you can trigger it from any other external tool. And the biggest benefit is that it has access to the relevant skills and the relevant information that you needed to have access to in order to run this as if it is running on your computer. And because it is running in Anthropic's cloud, your desktop or your laptop can be shut off and you don't need to have anything running, and this would still work. So the bottom line is all three options are good. It just depends on the specific use case, and this recent one is probably the most comprehensive as getting the ticks in most or all of the boxes, because you can now allow it to run independent of your computer with a set of instructions, with a set of skills, with a set of context information, all running in the cloud independent of your computer and still achieve whatever sophisticated outputs you want. That's it for today. I hope you will find this helpful. I hope this will solve actual real problems in things you are trying to achieve with Claude. As I mentioned, if you wanna learn a lot more on how to build really complex automations that run Claude Cowork and Claude Code to automate literally anything in your business, come and join our next cohort that is starting in the beginning of August. We are now actively taking registrations to that. And like I said, we sold out of May, June, and July. So if you want to do this, run and register right now before this one sells out as well, and then the next one will probably be in September. And if you are a manager in a company and you wanna do this for your team, or if you are the head of the company and you wanna do this for the entire company and do this as a condensed workshop, please reach out to me and let me know. And that's it for today. We will be back on Saturday with another news episode. Have an amazing rest of your week.