Build an AI Company with different AI Teams | OpenClaw and Paperclip Alternative

If you are using tools such as Cloud Agent Teams, Open Cloud, or an orchestrator as Paperclip, but you feel like it's lacking real power or an actual use case, or you want to set up an AI team or an entire AI company, then this video is for you. Because today I will introduce you to a tool that I built just for that. It's fully local, so you can install it on your own computer, and you can deploy your first team in just two clicks. They will all be wired up with the right skill set, access to the right tools, and working towards a unified goal. Everything is running on Cloud, but you can even hire contractors to get access to models such as Gemini, Grok, and many other. I will walk you through the easy setup process. We are going to cover the core features, and together we are going to deploy our first trade desk team. So, make sure to watch till the end, and you will be ahead of 99% of the people doing agent orchestration. Now, before we start, I want to let you know that this is all open source, so you can add your own ideas, and you can even create your PRs via GitHub if you want to merge it into the source code. You can also find me on X, so make sure to follow me because that's where I will share most of the progress regarding the project. All links will be in the description, and with that said, let's get started. All right. So, in order to install this tool, what you can do is go to the documentation, and that you can find at muxai.io, and then you click on get started or docs here on top. Now, these are all the documentation, and I'm adding more docs every day. However, what you will need is the quick start guide, and here you can see all the steps that are needed. There is a manual setup. So, if you're a bit more dev oriented, you can follow these manual steps. If you want to just install it via a script, also that is possible. And there are a few prerequisites, as you can see here, Node.js, PNPM, and of course Cloud CLI, as we are running on Cloud in the background. And Git you can use in order to clone that Git repository, but that's optional. You can also just download the files via Git. Now, if you want to choose to install via the script, what you can do here, you can select macOS and Linux or Windows. And as we are running on Windows, we are going to use the Windows instructions. So, here you can see what I can do is just a Git clone, and this will download all the necessary files. And for that, I can use PowerShell. I can open, for example, my IDE, and then in a project folder I can go to terminal, new terminal, and that also opens a PowerShell window for me. However, I can also just use a regular PowerShell, it doesn't matter. If I paste this comment here now, or Git clone, this will clone the whole project onto my computer. And as you can see here now, we had it opened here, too. Now, you can see Muxai. And if you then go back to the documentation, it says here we have to change into that directory called Muxai, and then we can run this script here, which is located under install. So, for simplicity's sake, we can use this IDE, so we can also see what happens with these files. But once again, you can also just do that into a PowerShell window if you want to do so. So, let's change into that file. It's called Muxai, like that. As you can see, because the files are now located in Muxai. And then we can use that install script under install, install.ps1. And as you can see, it's located here, install, install.ps1. And there is also a update script, so later if there are any updates, you can always use that script to update your tool. And also, as you can see here, there is an sh and a ps1, so it's for Windows and macOS or Linux. Now, once you run this on a Windows computer, you might encounter an error, and that error is also described here. So, if you got an error that scripts are not allowed to run, then you can just copy this comment here, which will set an execution policy or bypass the execution policy in the current process. So, you can just run this, and then you will be able to run these scripts on your computer. Now, as you can see, it checks these dependencies here, checks for Node, it checks if PNPM is installed, also Cloud. If not, it will try to install it for you. If not, you need to make sure that these are present before you can continue. Next, you have the choice to set up a embedded database. You can also use a Docker as a database or an external database. So, you can bring your own database as long as it is Postgres. And as you can see, the embedded option is the default option because this will just install an embedded Postgres database for you without any additional configuration. If you use Docker, this will spin up a Docker database or a Postgres database better using the Docker Compose file here. Or if you have a Postgres, for example, you use Superbase or you have a database somewhere in AWS, it doesn't matter, you can bring your own Postgres database. And if you choose for option three, it will also explain to you how to enter the connection string. Now, for now, we're just going to use the embedded database because this is the easiest for the install. We don't have to do any configuration ourselves. The whole install script will take care of that for us. So, once we done that, everything will be downloaded, like all the additional libraries that are needed, they will be installed. And also, once that is done, you will get two options here. And then you can run that application once it is ready to go in two ways. You can use PM2, and this will run that application as a background service. So, let's say you're going to install this on a server where you will disconnect from, then you can choose to run this as PM2 because this will be run as a background service. It auto restarts, so if it stops working, it auto restarts the app. And as it says here, it is recommended if you're installing this on a server or just a machine that you disconnect from, but you want to make sure that the app keeps running. You can also run it manual, and this means that you have to start and stop it yourself. And for now, we're just going with the default option, which is a manual. And it also says here how you can start it. So, to start it in a production mode, so you want to start this in a built mode, then you can use PNPM start. If you're developing, let's say you're changing things in the code, you can always run in PNPM dev, and then your changes, they will reflect automatically because it has hot reload enabled. However, if you're using PNPM start, this will use the production built that has been created while you installed it via the install script. And of course, changes that you apply, you need to rebuild them then before you can see them. But for now, you can just type PNPM like this, start, and that will start that application for you. And if everything went well, it will also say here the Mux API is now running on localhost 3001. And if you go up one, you can see our app is also running on localhost port 3000. And also, these two are now talking, so our front end app has access to our API, our back end app, you could say, in order to make sure that everything runs smooth. And if you then click on localhost 3000, this should open the app in your browser, and you should see something like this. And if you see something like this, that means that you have installed the app properly. All right. So, one quick thing before we continue, if you go back to your application and you stop this now, what you can do if you notice that there are any changes, you can always use that update script. So, once again, you see the install/update .ps1 or sh if you're on macOS. If you run this, what this will do is pull again that project from Git. It will see if there are any database changes that need to happen, and once again, install that tool including all the latest changes. So, that's also good to know that you have that update script there within the install folder that you can use to update your app in case there are any updates available. Now, in the future, I will make sure that you can update from within the app. For now, you just have to run the install/update.ps1. And also here, you can see we can once again run the PNPM start comment, and that will start the app for us just as it did before. And our web front end and the API are once again available. So, if we go back to the dashboard, we can just refresh it, and we are ready to continue. All right. So, once again, all the core features of an orchestrator, they are there. And I'm speaking of new agent, for example, you have an agent overview. As you can see, you have all the options for these agents. Now, we also have agent templates. As you can see, you can install a team lead, a news analyst, a technical analyst, a data analyst, or you can just start from a blank template. If you want some inspiration, you can always use at some of these templates, and then you can fill your own skill MD, for example, in as you can see here. So, these are predefined templates, they are also tested, and they work quite well. Now, what you have here is, of course, the name, the role of that agent, and you also have a title. You can have a capabilities overview. This is optional, just to show what the agent is capable of. Also, you can response to an other agent. We will look at that in more detail later on. And also here, the model that we want to use. We have Sonnet, Opus, Haiku, and the effort, so we can also set how much effort we want our agent to make. The max turns, and this is when you run an agent, let's say it invokes a tool, and it gets an error back. How many times do we want that agent to retry? It might be with an other tool, but how many times do we want our agent to return to actually get the result that we're after before it gives up. So, these are the max turns for the agent. You can also enable Chrome browser. So, if we enable this, we will allow our agent to use our Chrome browser to do certain things. So, we can instruct it, for example, to go to a chart and actually look at a chart and see what is going on using the actual browser. We can also scrape data from a browser using this Chrome extension, as you can see here. And then you can enable it for any agent. The stream history is enabled by default, and that just mean once we're looking at the the stream, once the bot is doing something, if it's being outputted or not. And then we also have a check here if we want to review the decisions that the bot made in the past before it makes a new decision. And we'll also come back to that later. And then of course the skill MD, and this is what the actual agent is good at, what is his skills, what is expertise is. And the default prompt, which is also always provided once the agent starts. Now, on the other hand, we have here the built-in MCP servers, which are enabled, which means that it will use the built-in MCP services or tools from the platform, and I will show you in a minute what that is. You can also choose to use the global config, and then it will have access to all the MCP servers or the skills that you have provided to your cloud instance. Now, by default, it selects the built-in MCP servers because they are also available once the tool is installed. And then you also can disallow certain tools, so if you want an agent not to run certain tools, like for example, read, write, edit, bash, grep, glob, or agent. An agent is a function to spin up multiple other agents, then you can also disallow these. And as you can see here, we have a number of built-in MCP tools, and these we can also disallow for our agent. So, if you don't want our agent to use the orchestrator MCP server or the wallet MCP server, we can enable and disable these things here. And also docs, for example, so we can always granularly select what we want our agent to use as tools, what tools he should have access to, and which tools should be forbidden. We can also set up a job, so we can run it every X amount of minutes, hours, or custom, and this is just regular cron. So, if we set a schedule, our agent will start on a schedule. Now, once again, one of the features of our tool is to deploy a whole team at once, so we don't have to create all these agents one by one. And if you go to deploy team, you can see here we have already a trade desk, and more teams are coming. These teams, they are created by myself. I created all these agents. I carefully tested them, and I made sure that this trade desk, for example, works properly together without any errors. And more teams are coming, but as you may know, all the skills, they need to be written carefully, they need to be wired up with each other. So, more teams are coming, but for now you can just, for example, use the trade desk as an example. So, if we click on trade desk, you will see the team members that are in this team. For now, we have a lead agent, a news analyst, a technical analyst, and a data analyst. And if we click on deploy team, this will deploy our team for us. It will populate the skill sets for all these different team members. It will wire them up, so they know who to report to who, and all the things like, for example, assign roles, assign MCP servers to some of these individuals. So, as you can see, we end up here now on the team lead page, and we can already see that our team lead has three team members, the news analyst, the technical analyst, and the data analyst. We also have a default prompt, and here it says that he has to instruct his team doing a detailed analysis of BTC versus USDC on the 4-hour time frame, and it also needs to instruct each team member to do their assigned tasks. And also for the team lead, it has a default prompt including the name, which is team lead, and also the role of that team lead, which is in this case getting all the information from the news analyst, the technical analyst, the data analyst, and then come up with a synthesized response, a decision, and that decision is also defined here. So, in the result card, we can select certain decision types. And for now, we want a trade decision, which means that the decision will be outputted in a certain JSON way for us, and our results engine or our decision engine will pick up that format and output it in a nice card, but I will show you in a minute how that actually looks like. Now, there are many card types, as you can see, we can output raw JSON, no decision, we can have a task decision, and alert, metrics, and so on. But for now, since this is a trade desk, we want a trade decision, which includes long, short, or wait, and also the entry, take profit, stop loss, the consensus, the invalidation, and also things that we need to wait for. Now, you can always copy these instructions to go to give to agents. So, let's say you selected this trade decision, you can copy the instructions and then just paste them in the skill MD, and that will say to that agent, you need to always output your decision in this format, and then the decision engine will pick up that format and create a nice card. We also have notifications, but for now, we don't have any notifications enabled. That we can do in settings, and then we can choose here to enable these notifications per agent. So, it can, for example, send that decision to Discord. Now, we also have the invocation here, so we can see what that agent actually going to do, what the environment variables are set up for that agent, what MCP server and tools he had access to, and then the prompt and the default prompt. So, everything is here, and that's really great for troubleshooting if something goes wrong. Now, if you have a look at an other agent, so as you can see here, if we click on agents, now we have four agents here. If we, for example, look at a technical analyst, also here, as you can see, there are no recent runs, there are no teams, but also here we have a default prompt, and that prompt is a bit more extensive because we actually say to that agent what tools to use, what MCP servers to use to actually conduct a review or actually do a analysis, which was instructed by the team lead. So, also this is a bit more extensive, and you can update that under edit. And also here is the skill MD. And also here you will see that this agent, it doesn't have access to, for example, news analyst MCP servers, contractor MCP servers. It doesn't have to orchestrate, it doesn't do anything with wallets. It only needs to do a technical analysis, and this and therefore it has access to the crypto data, like the candles, as you can see here, the candles, the indicators, and also the short analysis. So, it can use these MCP servers, but nothing more than that. So, if you go back to agents, or we can click on teams, here we can also see that team that we just created. We can also have a flat overview or a hierarchy. As you can see here, our team lead, and these are the members. So, back to the team lead, if we start this now, this will instruct its team to start doing that analysis and eventually give us a result. Now, before we do that, let's have a quick look at some other features, for example, the MCP servers. So, these were the MCP servers that I talked about and which are built-in. So, we have a news analysis MCP server, short analysis MCP server. We can also invoke contractors, and contractors, for example, are just other models that we can use to actually get a second opinion, for example, and I will show you in a minute how that looks like. We have the orchestrator, and that is used by team leads to orchestrate a team to get the decisions. We have a wallet MCP, which is used to create a wallet and also to pay over X402. And we also have crypto related MCPs, as you can see here, the crypto, the candles, and we also have crypto data, and also the docs, and this is tied into the actual docs that we see here, the Mux AI docs. These are available to our agents via that MCP server, and we can also add our own MCP servers as well, as you can see here, or we can do it uh like by pasting a JSON file, or we can add it manually. And if we paste that JSON file, it will also start exporting all the values from it and create a valid MCP connection for us, and we will also be able to test it. So, if you want to add additional MCP servers besides the one that are built-in, we can also do that. Now, real quick, contractors, as you see here, we can sign a new agreement, and here we can give that contractor a name, and for now, we only have the open router provider. We can select a model, for example, Croc, or we can select a Gemini, and then we can give an API key for open router, and then via that MCP, that contractor MCP server, our agents, they will have access to a contractor. Now, we're going to add more contractors later, like we can add open claw, we can add many other things that can actually be triggered, that can actually be seen as contractors. For now, we're just using open router that we can invoke other LLMs, other models to get opinions from. Now, besides that, we have a chat, so we can also chat with all our agents individually, and they know what their tasks are, they know what tools they have access to, so we can ask them, just as in a real corporation, why they did certain things, we can steer them, we can give them a bit more context, so that is also available. Now, let's just continue with starting our team. So, if we go back to our agents, we go to our team lead, we can just run it, and then you will see that it starts. So, now we have one running tasks, and if you go to the stream history here, we can clear the old history, we can show all the columns. You can see that our team lead started. And also here, you will see that it starts by invoking its team, and it knows that how to do it because it can use that orchestrator, that MCP orchestrator tool that it has, to invoke the news analyst, the technical analyst, and the data analyst. And as you can see, all three started here now, and once they conducted their own investigation, they will feed back the information to the team lead, and then the team leads will also check for additional information if for example that the technical analyst needs more information from the data analyst but once he is happy with all the information it got it will create a decision for us and output that decision and it will be available for us here under results and once that's there I will come back and we will have a look at it. All right so as you can see now every analyst has returned his findings as you can see here the technical analyst of course it returns stuff that has to do with the technicals for example VWAP it's talking about RSIs all that kind of things and it feeds that back to the team leads. The news analyst on the other hand he uses an MCP server that is created to fetch news feeds and based on these news feeds it can extract data that has to do with Bitcoin in this example and also feedback his analysis to the team lead and then the data analyst of course he looks for data things like open interest funding all that great stuff that can provide us some extra confluence and then send it to the team lead. Now the team leads job is just to get all that information see if there are any gaps asks for additional information from any of these analysts and then come up with a trade decision and as you can see that trade decision is in a form or in a certain format that we requested because that format that will be used by our decision engine to create these nice results as you can see here. The team decided to go short on the BTC USDT they are medium convinced and also we have an entry a take profit and a stop loss their risk reward for our trade the consensus that has been given by the three analysts also the invalidation for this trade so if certain actions or triggers happen it can invalidate this trade and also things that we need to watch for so if these things happen it can give more confluence again it can invalidate our trade but at the end we have a nice trade decision that has been given by our three analysts and our team lead. So as you can see also the decision is shown here and this we can run for example every four hours if we look at the four hour time frame short for example and we can also allow our team lead to review [music] his decisions because that was an option remember in the agent edits before it's create a new decision. So also that is possible and again we can fine tune these agents very well because there are many configuration options as you can see under the team lead and then under edit but that team lead that trade panel that is auto deployable I really curated that I looked at it I tested it a lot and it seems to work well. However you can always add your own strategy your own insights and you can add more agents. Now I hope that this was a bit insightful that you can see the power of the system and again more teams come soon that can be deployed just by two clicks. For now we have the trade desk but as I said these teams they are carefully curated so that they are wired up correctly that they have the right skill sets and that they can actually come up with a nice decision in any sort of field it doesn't only have to be trading. So if you think this was interesting check out the documentation pull in the repository start playing with it again you can add me on Twitter I also have a Discord server I will put all the links in the description I'm happy for people to reach out to me that want to work on this project that have ideas they can also be merged in the code base via Git. Again just reach out to me I hope you think it was interesting and I'll see you in the next one.