OpenAI joins Amazon, Microsoft and Cursor on portable agent plugin standard

Why it matters

Agent vendors are standardizing the package around AI workflows while keeping marketplaces and permissions under their own control. Developers gain portability without giving any single platform ownership of the format.

OpenAI has joined Amazon, Microsoft, Cursor and Vercel in governing Agent Plugins, a vendor-neutral specification meant to let developers package reusable skills and Model Context Protocol servers once and load them across competing AI agent clients.

The version 1.0 specification defines a common directory structure built around a required plugin.json manifest. Developers can add Agent Skills under a fixed skills/ directory and declare local or remote MCP servers in an mcp.json file. Client makers retain control over installation, distribution, permissions and interface design.

That division is the central bargain. The standard defines the portable core while leaving room for OpenAI, Microsoft, Amazon and other vendors to add their own features through reverse-domain extension namespaces. A plugin can therefore carry instructions and tool connections that work across clients, alongside hooks or other components intended for a particular product.

The specification remains a working draft. Agent Plugins v1 standardizes two component types: skills and MCP servers. It does not create a common marketplace, payment system or authentication layer. The MCP configuration format supports local servers over standard input and output, remote servers using Streamable HTTP, and the older HTTP plus Server-Sent Events transport. OAuth discovery and credential storage remain the responsibility of each client.

One package for competing agents

OpenAI's plugin documentation already describes plugins as installable packages that combine skills, MCP servers and optional interfaces for ChatGPT and Codex. OpenAI operates a shared directory across those two products, allowing developers to publish a plugin once for discovery on supported OpenAI surfaces.

The new cross-vendor format addresses a wider problem. AI coding tools built their plugin systems separately, forcing developers to restructure or duplicate the same skill files and MCP configurations for each client. Agent Plugins establishes fixed locations and schemas that participating clients can recognize without requiring plugin authors to maintain a separate core package for every tool.

The project's compatibility directory lists ChatGPT and Codex, Cursor, Microsoft's Visual Studio Code and GitHub Copilot, and Amazon's Kiro. All five support Agent Skills and at least the standard input-output and Streamable HTTP transports for MCP servers. Support can be adopted incrementally, so a client may load the portable components it understands and ignore the rest.

For OpenAI, the standard extends a plugin push that was already underway. OpenAI added a curated plugin directory to Codex on March 26th, 2026, describing plugins as bundles for reusable workflows. On July 9th, OpenAI moved ChatGPT's app directory into its Plugin directory, making plugins the primary discovery mechanism for workflows across ChatGPT and Codex. Existing app connections continued to work, while workspace administrators gained separate controls over plugin installation and access to the underlying apps.

Those controls matter because a portable package can combine static instructions with software that accesses files, cloud services and internal systems. OpenAI's enterprise documentation separates permission to install a plugin from permission to use an included app. The Agent Plugins specification likewise requires clients to reject files that escape the plugin directory, but it explicitly states that those path checks do not sandbox an MCP subprocess. Each client still has to enforce execution permissions, network access and credential handling.

Governance spreads control across vendors

Agent Plugins is openly licensed and developed through a public specification repository. Its initial Technical Steering Committee consists of core maintainers from Amazon, Cursor, Microsoft, OpenAI and Vercel.

The project's governance charter assigns committee seats to individuals rather than companies and bars any single vendor from controlling a majority of core maintainer positions. The committee oversees changes, compatibility rules, releases, conformance work and technical disputes. Proposals and material technical decisions are intended to remain public.

The membership gives the standard immediate reach across several of the largest distribution points for AI-assisted software development. Microsoft controls Visual Studio Code and GitHub Copilot. Amazon operates Kiro and has already published collections of AWS-focused agent plugins. Cursor runs one of the leading independent AI coding products. OpenAI distributes plugins through ChatGPT and Codex, while Vercel publishes agent tools used by web developers.

That backing does not guarantee identical behavior across products. Client-specific extensions remain part of the design, and marketplaces, review rules and permission systems stay fragmented. The 1.0 draft instead sets a minimum portability layer: a manifest in one predictable location, skills in another, and MCP configuration in a third. If the participating vendors keep those pieces compatible, plugin authors can distribute the same operational knowledge and tool definitions without rebuilding the package around each agent vendor's file layout.