Show the full transcript (1,257 words)
A lot of big names in the AI industry have just announced a new open standard called Agent Plugins, which means that you can bundle together multiple agent skills and MCP servers that will work out of the box across lots of different agents. So, what does this all mean? How do you get started? And how is this different to skills and plugins that you might be already using?
To set the scene here, generally within mainstream coding agents, we have access to plugins, where we can extend the coding agent's capabilities. So, by selecting any of these kind of plugins, for example, these will have reusable skills and connections to external services. This new standard aims to make it so that one of these kinds of plugins will work across lots of different agents. This standard is being pretty widely adopted by a lot of players in the industry, as well as more recently, Google. Unsurprisingly, there's one big player absent from this standard, and that's Anthropic. But I will be showing you a skill you can use to migrate Anthropic plugins to this new standard, so they'll work for all of those other coding agents.
A plugin under this standard is just a directory. We have a skills folder, which can hold multiple skills, and then we have a plugin.json file, which is a manifest, that's a very simple file. And then we have the mcp.json file, which can store the MCP servers. So the plugin here really is just a collection of skills, as well as an mcp.json, which will allow us to integrate with other services. So really, the only thing that this spec actually covers are the skills and MCP. Everything else is specific to your client or your coding agent. A lot of the stuff you might lean on when using a coding agent, like using commands and hooks, doesn't come along within the main part of this spec. That has to be stored within a client-specific folder. That's likely going to be the biggest criticism of this spec, but likely the creators of this specification felt that all of these elements are still too client-specific to pin down across all of the different coding agents, which is probably fair enough.
So within the agent plugin folder structure here, we have this portable core, the plugin.json, the skills, and the mcp.json, but then you can let individual clients add behavior without changing the portable core. For example, hooks that might be used by a specific coding agent. Right now I have two plugins up and running within CodeX using this new open standard. Right now, compatible clients are GitHub, Kiro, ChatGPT and CodeX, Cursor, VS Code, now Google. Also, if you use the Pi coding agent, there is a Pi agent plugins package within their package catalog. I haven't used that, by the way, so definitely have a look at it first if you're planning to use it.
To test using plugins using this open standard, what I did was I went to this Agent Plugins Example repo, which was created by the maintainers of this spec. And this repo contains a skill where you can migrate existing Claude, Copilot, CodeX, Cursor, specific plugins to this more portable spec. And separately to that, Anthropic have some very popular knowledge work plugins on GitHub. Some of these were spooking the stock markets earlier on in the year. But to start with, I'm going to select that migration plugin that I talked about, and I'm going to provide that Anthropic legal knowledge work plugin. From there, I'm just going to press enter and let CodeX work away. So after a few minutes, it looks like it's ported this over to the new standard. And it looks correct. We have the plugin.json, mcp.json, and all of the skills within here.
Now, if we go to plugins within CodeX, I'm going to refresh, and now we can see we have this new legal agent plugin here. If I go into it, we see we have these skills and MCP servers. If I click install, then that should install it. Because that plugin has MCP servers out of the box, I just got prompted with one of these sign-in screens, which I'll skip for the moment. But if I go to a completely different project now and press forward slash, if I type in legal, then we now have all of these different skills popping off that we can use. So, for example, compliance check, I can click into that, click into it again, and we can see the skill pops up.
So that's all great, but I was initially running into problems when trying to even install these plugins locally, because agents like CodeX and Cursor both expect you to be installing plugins from marketplaces. So if a developer uploads one of these open standard agent plugins to the online marketplace for the coding agent you're using, then it should be very seamless to install. And the developer will then be able to upload the same plugin to other marketplaces potentially without making many changes to it or any at all. But for me to get this to work locally on my computer, I had to get CodeX to create a marketplace.json file and tell CodeX where to pick up the plugins, which is definitely not ideal. So from my understanding, this does not seem to be as portable as just loading an agent skill into a project.
The open standard is defining the actual folder structure that will be portable across different agents, but it doesn't include the distribution or loading of those plugins for you. And of course, something I have to say, if you're using externally sourced plugins, then we have the standard disclaimer that we'd also have for skills and MCP servers to only use ones that come from reputable sources or ones that you understand very well. So is this standard going to persist? Well, that's the big question at the moment. Standards live and die based on how widely they're adopted. I think it's a very positive move to have some consistency between agents so we can jump back and forth between them. For users of AI agents, the more open standards we have, the better. Anyone that's used Claude Code or CodeX within the same project, will have these niggely issues whereby you have to have an agents.md file and a claude.md file. You might need to have two sets of skill files and one that references the other. And this can lead to extra work and redundancies and inconsistencies that can cause issues in projects.
This is a really simple standard that has the potential to be adopted pretty widely. If Anthropic started using it, then that would be pretty guaranteed. But there's a good chance that in the agent plugin world, we'll have one format for Anthropic and another format for everyone else. If you want to learn more about that standard, you can check out the links in the description. If you want to learn how to build expert-level AI systems from the ground up, then check out the link in the description to our AI Architects course in our community, where you'll get over 50 lessons covering everything from agentic retrieval to harness engineering. You can get access to all of our other resources, starter apps, and a global network of AI builders. Thanks for watching, and see you in the next one.
