OpenAI, Google, Amazon, and Vercel’s new “Agent Plugins” standard might make devs’ lives easier

Four major AI companies have set competition aside to develop “Agent Plugins 1.0.0,” a new standard for agents.

The new standard, developed by OpenAI, Google, Amazon, and Vercel, is described as “open” and “vendor neutral,” meaning it’s open source and can be used across a range of different AI models.

These AI companies decided to create a plugin that acts as a directory, which could include skills, MCP servers, hooks, and commands.

When creating a model context protocol (MCP) server or agent skill, developers will package them for a specific client, making them harder to transfer across different products.

This is because each company once developed its own agent platform separate from others, due to a lack of standardization. But that’s starting to change.

Standard starts with MCP and agent skills

This first version specifically focuses on agent skills and MCP servers.

MCP, an open standard originally developed by Anthropic, lets agents connect with external data to solve problems independently using a single protocol instead of writing a script for each AI product.

Agent skills are reusable expertise packages that help AI agents perform specific tasks by equipping them with the necessary instructions, workflows, and context, allowing the developer to work elsewhere.

These AI companies decided to tackle agent skills and MCP servers as they are the 2 types “with the most adoption today,” according to Amazon Web Services.

While developers have flexibility when it comes to which AI model they use, whether they choose Anthropic’s LLM or Google’s, they don't have much choice when it comes to choosing what tools the AI uses.

This standard allows developers to pick which tool the AI uses without committing to one specific client or provider.

Why the change?

The original process was complex, inefficient, and error-prone, as devs were forced to adapt the same piece of software.

For example, if a developer wanted one AI agent to access GitHub, they’d end up writing different instructions for Claude, Cursor, Gemini, and ChatGPT.

It doesn’t matter whether the developer wants to use the same integration across multiple agents – they would still need to reformat it due to a lack of standardization.

It’s as if a developer is writing the same application for a different operating system, as each system works differently.

Now, the Agent Plugin aims to streamline the process by making sure each system works in a similar way.

This is similar to standardizing hardware, as the European Union did with USB-C support for all new devices.

Isn’t meant for every task

This standard can’t be used for everything. It's more about having things organized in one location for transport.

While developers might be interested in trying things for themselves, there might not be an immediate need.

“Before you reach for a plugin, ask whether you need one. If you're shipping a single MCP server to a single client, ‘mcp.json’ on its own is still the simpler answer. If you have a single skill, you don’t need a plugin,” Google advises.

Agent Plugins work if devs have components that work together and need to be transported as a unit.

This version is nothing more than a package format, as it “defines no install mechanism, no distribution protocol, no permission model, no sandboxing requirements, no trust or provenance verification, and no user experience.”

This is a deliberate move by OpenAI, Google, Amazon, and Vercel, as these companies are not promising that packages will be safer, nor are they providing guidelines for downloads or permissions, since the use cases are likely to be broad.

What it’s saying is that this is a formatting guideline rather than a “how to use” guide. How enterprises, independent developers, or other organizations use it is up to them.