Why MCP is dead & How I vibe now

Many people didn't know you can actually use simple open source tools to reduce more than 70% of token consumptions for your coding agents and extend its capability in much more scalable way. So every MCP server you add comes with this bundle of different tools and each tool comes with this JSON schema and all those information is part of context, no matter whether the task that agent is doing at the moment is relevant or not. So it just unnecessary eats up loads of context window. A much

more scalable and easy to use method, in my opinion, is this combination of skills plus CLI tool. It allows you extend your agent capability to hundreds of different integrations with only minimum amount of token consumptions. In my experimentation, some of them can lead to more than 70% of token consumption reductions. And this is what I want to take you through today with step-by-step process of how you can do it and special

agent skills to enable that. But before we dive into that. Entropy released Cloud CodeWorks, which is a special version of a cloud code that aimed to automate not just coding, but knowledge works. It gave us a sneak peek of how white-collar work gonna be done in future. But many of us are still probably early stage experimenting it or figuring out what's the best way to unleash its full power. That's why I want to introduce you to this free research and material HubSpot provided. They did a

whole bunch of experimentations across all the capabilities Cloud Cowork is providing and test across hundreds of use cases in real world. So they figure out 12 advanced prompts and use cases that is programmatically useful across things like automatically analyze content performance data and propose new channels. to create common business documents like job description, SOP, email sequence, ASCAL, as well as getting Cloud Cowork continuously organize your file system for business and

even automated competitive intelligence reports that can be scheduled every week. And each one comes with a ready-to-use prompt that will trigger a predefined Cloud Cowork process. So if you're interested in understanding what the real paradigm shift Cloud Cowork is going to provide, I highly recommend you go take a look. I put a link in the description below for you to download this resource for free. And thanks HubSpot for sponsoring this video.

Now let's get back to how can you use SKILL plus CLI to replace MCPs. So Agent SKILL concept has been introduced end of last year, but it started getting extremely popular past few weeks. And this concept is simple. You basically created SKILL.MD, which including a snippet of prompt that it re-injected to the agent when the SKILL is retrieved, plus the list of resources that agent can use to complete the tasks. Every time when you load up a SKILL, It only adds a minimum amount of token

in the context window. Basically, Cloud Cloud only has this one tool called execute skill. And inside this tool description, it has a section of available skills. So you can see that every single skill you add only add additional 10 to 50 tokens. That means with the amount of token that is taken from MCP tools in this screenshot, you can easily add about 4,000 different skills, which is more than anything you will possibly need. So when agent executes skill, The prompt you load up here will be

added to the context. And based on those context, you can include instruction about how to utilize the rest of resource to get either additional relevant context or execute some predefined scripts. And we already see a whole bunch of developers moving towards the skill-based setup that is creating much more token efficient agent. One very popular example is actually Manus. So Manus actually did a whole bunch of knowledge sharing about how they do

context engineering. And one of the key things they did is that they break down all the agent tools into different levels. There are fundamental tools that is always useful, like read, write, edit file capability, things like that will be always loaded as normal tools. But for the rest of less frequent tools, they're most likely put under a command line that agent can run to execute. For example, if you add an MCP to Manus, the MCP is never exposed to Manus. Instead, Manus agent has this Manus

context engineering. And one of the key things they did is that they break down all the agent tools into different levels. There are fundamental tools that is always useful, like read, write, edit file capability, things like that will be always loaded as normal tools. But for the rest of less frequent tools, they're most likely put under a command line that agent can run to execute. For example, if you add an MCP to Manus, the MCP is never exposed to Manus. Instead, Manus agent has this Manus

MCP CRI tool. The agent can call to execute MCP and they just add a little prompt snip to tell Adrian what type of MCP it can run. So that when Adrian needs to use this tool, it can just run the command line to run the MCP instead. And because it is CRI, it can also run other useful commands to get more detail on debug if there's any error. And because it's CRI, agent outbox can do more complicated actions like piping multiple different actions together in

one go. with some default system function, like wait for a certain amount of time. And even the more complicated action, like passing result A to result B, long as it understands the output schema well. And this is just such a more powerful and token efficient way to extend agents capability. And to showcase why this way better, I just quickly go through this one example. So we all want Cloud Code to be able to use a browser, test application, or review the design. The browser tool like

one go. with some default system function, like wait for a certain amount of time. And even the more complicated action, like passing result A to result B, long as it understands the output schema well. And this is just such a more powerful and token efficient way to extend agents capability. And to showcase why this way better, I just quickly go through this one example. So we all want Cloud Code to be able to use a browser, test application, or review the design. The browser tool like

PlayRider or even CloudCode's own Chrome MCP is very complicated because it will load up a whole bunch of different tools for agents to open a tab, click on something, drag something, type in something. And that can easily take at least 2% of context window, which in this case left only 125,000 effective context window that agents can use. Like in this example, I use Chrome Dev2 MCP to do the UI test of SuperDesign platform. And after finished a basic UI test like this, it consumed a

bunch of token and in the end it has 87,000 token left. But on the other side, with Adrian browser, what they did is that they didn't introduce a MCP tools. Instead, they built a CRI package, which is a command that both you and Adrian can run to do all sorts of different browser actions. So it can just run those command line, take actions and get contacts back. And it did a lot of different optimizations to making sure the information in return really help agent understand the

page structure, what kind of interactive elements are. So the content in return here is also a lot more token efficient compared with other browser MCP tools. And more importantly, because this is a CLI there, it can be run directly in terminal. All we need to do is just adding this agent browser skew that clearly lists out how to use this CLI tool. So with this approach, the effective context window you will get is already bigger than what you will get otherwise with MCP tools. And all

you need to do is just load up this context and start running those command line and get the result back. And with this specific example, using agent browser with the same task, when it finished, it got 117,000 token left. That is 70% less token consumptions. And as I mentioned before, this Bash plus CLI tool just provides so much more freedom about how agents can execute tasks. So one thing I think really everyone should be doing is migrate off your MCP to a skill instead.

But to do that, you need a way to run MCP tools through CLI. And this is where I found this open source MCP CLI called MCP Porter. It allows you to run MCP directly through command line. And it has all the helper function that agent need to get more information about tools within this MCP server, like specific schema that I need to use. So this basically allow us to do what Manus has manually implemented on their site before your coding agent. And this is exactly what

I did here. So instead of having all sorts of different MCPs, for each MCP, I will just add one skill file. For example, for context seven MCPs, I can directly run this npx MCPorter called contact7.resolveLibraryId, which is a specific function inside contact7.mcp server and passing on the parameters like this. And if I run this, results will be directly returned back in a markdown format. And instead of adding the contact7.mcp tool, I can just add this one skill.md file

for contact7, where I'm putting in the description about when to retrieve the skills, as well as detailed description about available tools it can call. That's why my Cloud Code doesn't have any MCP tool added. Instead, it just consumes a few hundreds token for all the different SKUs that I've added here. And I'm sure I can actually reduce that token even further by simplifying the descriptions. But it's already much better and can

achieve all the critical actions that we need. Like if I prompt it, help me find the latest doc for Verticell AI SDK using context 7, it will automatically load up the SKU and start executing those different command line to run the MCP. So you can just start immediately save a whole bunch of token by converting your MCP into a SKU together with MCP Porter. And I've already started this process of converting all my MCP over to SKUs. I even created this one SKU called add new

MCPs, where I have this instruction to tell it how to use this MCP Porter, how you should install test MCP server and create a relevant SKU. For example, I can just add any MCP, paste in Cloud Code and tell it, help me add a new MCP SKU. Then it will load up those contacts, start adding the MCP to config and check whether the MCP has been installed properly. Create a folder for Excel skill, which detailing all the functions that the agent can call to use this

Excel MCP. And best thing is after adding this new skills, it only add a few new tokens into the contacts window. So this example of how you can get all the extended capability from MCP into skills. So you can extend agents capability much further without worrying about the performance and agent context window blow up. If you haven't done this, I highly recommend you go and have a try. If you're interested, I also put a full instruction of the add new MCP skill in the AI builder

Excel MCP. And best thing is after adding this new skills, it only add a few new tokens into the contacts window. So this example of how you can get all the extended capability from MCP into skills. So you can extend agents capability much further without worrying about the performance and agent context window blow up. If you haven't done this, I highly recommend you go and have a try. If you're interested, I also put a full instruction of the add new MCP skill in the AI builder

club. So you can just copy and use. I have put the link in the description below for you to join us too. And I adopted similar methods in a recent integration that I'm building between SuperDesign and CloudCode. So many of you probably know I am building this platform called SuperDesign, which is an agent that allows you to vibe design many different options for your products. And we recently built the SuperDesign skills,

club. So you can just copy and use. I have put the link in the description below for you to join us too. And I adopted similar methods in a recent integration that I'm building between SuperDesign and CloudCode. So many of you probably know I am building this platform called SuperDesign, which is an agent that allows you to vibe design many different options for your products. And we recently built the SuperDesign skills,

which allow CloudCode to directly call in specialized design agent and review design in Infinite Canvas. In this case, we didn't go for MCP tool. Instead, we just build a CLI that Cloud Code can call by itself and still doing super powerful things to add a new design to this Canva and retrieve different design prompts from the prompt library to enhance the design output. Setup is extremely simple. You just need to install a package by doing npm install dash g as super design slash CLI at the

latest and then call super design login, which will pop up this screen to allow you to authorize. And in the end, you can do npx skills at super design depth slash super design skill. So you can add to any coding agent that you're using as a skill. All we need to do just slash super design and prompt it. It will load up the instructions contextually and start experimenting with different design directions. All those complicated functions are done just through the skill and CRI itself. So

I highly recommend you go and have check. I hope you enjoyed this video. Thank you, and I'll see you next time.