How Agent Plugins Got Made

You may think of the agentic world as the Wild West, but over the past two years, technologists have been developing rules of the road. Anthropic’s Model Context Protocol settled how agents access external tools. Google’s Agent2Agent and Cisco-backed AGNTCY took on the question of how agents reach each other. Now, there’s Agent Plugins, a standard released this month for packaging Agent Skills—reusable instructions for agents—and MCP servers into a folder that any compatible AI assistant can open.

The Agent Plugins standard is designed to be vendor-neutral, meaning no single company controls it and no particular software is required to use it. It also keeps the format compact, specifying a small plugin.json file at the top level, instructions in a folder named skills, and tool connections in a file named mcp.json.

How it came to be is a more interesting story and explains why six of the biggest names in software agreed to make skills and MCP servers portable while leaving room for each to keep building on top of them. One of its architects even compared Agent Plugins to Google’s Chrome extensions.

The idea for the Agent Plugins standard initially came from Vercel but quickly drew support from Amazon Web Services (AWS), OpenAI, Microsoft, GitHub, and Cursor’s parent, Anysphere. “We were building skills for ourselves that we wanted to publish, and we realized there is no good way to do that,” Malte Ubl, Vercel’s chief technology officer, told The AI Economy. “We would have to publish essentially 10 different instructions…there’s a file here, and this is how you get it to Claude, this is how you get it into Codex, this is how you get it into Cursor…they’re all different.”

To solve it, Vercel built Skills.sh in February, a command-line tool and companion site for installing agent skills. Ubl said the idea caught on—visit a site offering a skill today and the “paste this command” instruction is likely pointing at Vercel’s tool. They realized that creating the skills wasn’t the issue, but how to install them. The same gap had appeared with MCP a year earlier, Ubl said, and nobody drew the lesson until it happened again.

“It [would] be much smarter if essentially you would uncouple the distribution from the concept,” he said. In other words, decide what a thing is, and let something else handle how it travels. “We can’t have, every time we invent some new thing that agents want to have, we need to build a whole ecosystem just to get that going.” The Agent Plugins standard is a single, generic way to package new agent ideas without having to redo distribution.

Vercel wasn’t the only one wondering why the process was this fragmented. AWS was seeing the same thing from its enterprise customers. Many were running multiple coding clients and were unable to package a single skill or MCP server for all clients, both internally and externally. “There’d be a plugin written for Cursor, or there’d be a plugin written for Kiro, and you couldn’t install both,” said Clare Liguori, a senior principal software engineer at AWS, in a separate interview. “It’s a real challenge.”

The scale of the problem is already visible. Search GitHub for the folder names different vendors use, and you likely get thousands of results, Liguori said—“a huge constellation of plugins out there today.” Much of it is duplicated effort. A SaaS provider releasing a plugin has to work with each vendor separately to build a configuration for every coding agent, she elaborated. Now, under this standard, they do it once.

Work on a standard began with an approach from Vercel. Someone on the team was also an active member of the MCP community, where Liguori is a core maintainer, and asked if she wanted to join other companies to figure out a solution. But rather than sitting down to design the standard, they looked for common ground they had overlooked.

“The two that really stood out were skills and MCP,” Liguori said. “Everybody had already standardized on agent skills. We’d already seen the success of being able to bring that context into any client that implements skills. We already had MCP servers—a little bit of differences in configuration, but generally all the same. So those were two really clear ones to start with.”

It was a deliberate choice to have the first version of the standard support only agent skills and MCP servers. “It’s really about having a…good baseline that I think [is] uncontroversial and very established, and then a baseline to innovate on,” Ubl said. “We didn’t want to actually waste any time on the things that would be actually, in a way, controversial.” By that he meant something narrower than disagreement. It’s just that nobody had settled on how to do the thing yet.

Agent and hook configurations are not included in the first version of Agent Plugins, though Liguori noted they are on the roadmap. Hooks were proposed early on, she said, but the group would first need a deeper discussion about what types of hooks exist and which ones “likely transcend any one particular vendor’s coding agent.” Then they’d have to sort them into three tiers: the ones everybody must implement, the ones they should but aren’t required to, and the ones best left as vendor extensions for now.

The negotiation among Vercel, AWS, OpenAI, Microsoft, GitHub, and Cursor seemed to be smooth sailing. “It was a very collaborative experience because I think we really all saw the value in this,” Liguori said. “We all wanted to get something out so that people can start building agent plugins, start adding the agent plugin configuration to their existing ones,” she added. “Skills and MCP are so broadly adopted just generally in coding agents and agents in general that we felt that was a great place to start.”

Share

A limited scope hasn’t stopped vendors from building past it. AWS’s Kiro already ships extensions of its own—for steering, the persistent project context that keeps an agent following a team’s conventions, among other behaviors. “One of the things that we made sure to include in this first release was the idea that we shouldn’t hold vendors back from developing new features that aren’t in the standard,” Liguori said. “The vendor-specific extensions were a very important part of that.”

Some of that work can end up in the standard. The steering committee is watching what vendors build, hoping some of it turns out to be the same across products and can be promoted into the standard. Liguori said the group also wants to collaborate on extensions that aren’t tied to any single vendor—“generally useful” ones that a couple of vendors support at first and grow from there.

Ubl expressed similar thoughts, saying extensibility was a “very conscious” decision. “Any vendor can add proprietary extensions in a safe way to the standard, and that’s really important.” He added that’s what made it easier to get a large group of vendors on board. A company with a feature the standard doesn’t cover puts it in a folder labeled with its own domain name, tells developers how to use it, and ships without waiting for anyone. “This is definitely different from standards like HTML where there is an expectation that you support the standard and you should not invent more stuff,” Ubl said. “This is different in Agent Plugins. This is actively encouraged—be creative, have ideas. We are still inventing the world.”

At launch, Agent Plugins works with OpenAI’s ChatGPT and Codex, Anysphere’s Cursor, Microsoft’s GitHub Copilot and VS Code, and AWS’s Kiro. Notably missing is Anthropic’s Claude. Its absence is conspicuous, especially since it’s not only one of the more popular coding agents but also the one that originated both MCP and Agent Skills. Anthropic did not respond to a request for comment. Ubl said conversations with the company took place, and that sometimes companies want to be named in a blog post—sometimes they don’t. He also acknowledged “some interesting dynamics going on.”

He pointed out that regardless of whether a company supports Agent Plugins, “ultimately, everyone benefits,” because it removes each vendor’s responsibility to build out that ecosystem alone. But the benefit isn’t evenly distributed. It matters more, Ubl said, “if you’re not the market leader”—the market leader can assume developers will find their own way into its product. “The kind of thing we are concerned about is mostly how easy it is for the Next.js, Svelte, Nuxt, [and] AI SDK team to provide integration with the coding agents of the choice of their users,” he added.

Share

Ubl compared Agent Plugins to Chrome extensions. Google’s system allows users to install add-ons that change what the browser does, such as a password manager, an ad blocker, or a grammar checker—all with a couple of clicks from the online store. Chrome had the users, and Microsoft and Mozilla eventually built their browsers to accept the same basic format. It was where the developers already were. A W3C group took up a formal specification years later, working from what the browsers had already converged on.

Agent Plugins is attempting to achieve the same thing, but instead of building first and letting everyone catch up, it’s about agreeing first, then getting people to use it. One critical difference, though: Chrome’s format spread because Google had the users. No member of the Agent Plugins consortium holds that kind of position on its own, though collectively the six ship most of the coding assistants developers actually use. “This is literally like a Chrome extension, but for your coding agent,” Ubl said.

However, the analogy runs into a limitation that Ubl himself raised: Chrome extensions update quietly in the background, whereas an agent skill stays at whatever version was installed until someone replaces it. That falls to whoever installed it. “In practice, people don’t do that,” Ubl said, before adding that he was exaggerating and that it’s simply rare. “You have to actively do it.” Auto-updating was discussed, he said, though whether it belonged in scope was never settled. Plugins delivered through OpenAI and Anthropic marketplaces already get it. Because those companies run the marketplaces, that’s precisely what makes it hard to write into a standard.

With Agent Plugins formally launched, Liguori expressed hope that the standard would make it easier for people to share their work and configurations, and for vendors to share skills and MCP servers. The result, Liguori said, is that “the end users start to get more superpowers out of the box, regardless of what agent [they’re] actually using.”

But one part that hasn’t been discussed, which Liguori raised on her own, is that the standard could encourage users to share their setup. “People think of plugins as…something that SaaS companies or just vendors in general are going to create,” she said. “What we’ve seen is that some of the most popular vendor-specific plugins on GitHub right now are not from companies. They’re from individuals [who] say, ‘this is what works for me.’”

Communicating that information can be difficult. It requires determining which skills someone uses, which MCP servers are needed, and the combination that actually makes you faster. “I’m excited about…being able to see what are other people’s setup, what’s their configuration,” Liguori said. “Maybe I’ll install theirs, and it’s pretty good.”

Ubl offered a different perspective. Most people who use plugins won’t think about how they work, he said, and shouldn’t have to. The audience he has in mind is smaller—thousands of open-source maintainers rather than millions of users. They’re the ones weighing how easy something is to build and distribute, and how many users they can reach without shipping a version per agent. “If I don’t have to make 15 versions of something, it’s much more likely that the one I make is actually really good. That’s how you get a win-win-win situation out of it,” Ubl said, referring to plugin users, those who make plugins, and the vendors whose agents run them.

Enterprises also stand to benefit from this standard. Vercel publishes plugins so people can use its platform inside their agent, for example. Alternatively, if a company operates an agent platform, such as Salesforce, ServiceNow, or Notion, embracing Agent Plugins makes it easier to quickly inherit thousands of existing plugins. Ubl isn’t concerned about the standard gaining traction. “This looks so similar to what has already been in the market. It’s not really a question for me whether the standard is going to be adopted. It’s so simple,” he said, describing it as the “most obvious thing you can possibly do” and also “not very controversial.”

Agent Plugins is already garnering the attention of some organizations. SAP told The AI Economy that it is looking more into the standard. “SAP welcomes open initiatives and continuously evaluates opportunities that align with our strategy and customer needs,” said Sean Kask, SAP’s chief AI strategy officer, in a statement. Salesforce declined to comment, and requests to ServiceNow went unanswered.

Adoption is one open question. Who ends up owning the standard is another. Both A2A and AGNTCY were donated to the Linux Foundation, and Anthropic transferred MCP to the Agentic AI Foundation. Does the consortium plan on turning control over to an independent organization? “We really have just been focused on the 1.0 launch, so we haven’t had that conversation yet,” Liguori said.