Major technology companies have joined together to create a unified format for writing AI agent tools and capabilities, allowing developers to build once and deploy across multiple platforms. Vercel released the initial draft of the Agent Plugins specification on Thursday, with contributions from Amazon, Cursor, Microsoft, and OpenAI, all of which committed resources and engineering talent to the initiative. The Linux Foundation's Agentic AI Foundation quickly adopted the standard, though it will function as a separate entity.
The Agent Plugins 1.0 specification packages an AI agent's capabilities and tools into a unified directory structure. It incorporates the AAIF-managed MCP protocol alongside the open Agent Skills standard that Anthropic originally developed in 2025. Agent Skills contains the directions for how the agent executes particular tasks, while MCP delivers the connections to external databases, APIs, shell scripts, and other outside resources. The combination creates modular AI functionality designed for write-once-run-anywhere portability. Platform providers remain responsible for building user interfaces, setting security parameters, and establishing marketplaces for plugin distribution. VS Code, Cursor, GitHub Copilot, ChatGPT, Codex, and Kiro already support the standard.
The specification follows Next.js conventions by using the host computer's directory format to define components. It requires a top-level directory for each plugin containing a compact manifest and schema, with the manifest providing the name, version, and schema in JSON format. Skills go into the skills subdirectory, while the MCP server configuration lives in an mcp.json file. Plugin creators can store additional extensions in the namespace directory. The AAIF noted in a blog post that this format "keeps the shared format predictable without requiring it to absorb every client-specific experiment happening across agent products." Pairing the skills with the MCP configuration creates a portable package that can be copied and used across any compatible agent client.
The specification maintainers intentionally limited the scope for this 1.0 release, even as agent developers worldwide add experimental hooks, rules, and extensions to their capabilities. The strongest additions might be evaluated for future specification versions, though the team isn't guaranteeing anything. Developers can place these extensions in the namespace directory and hope they remain portable. Agent Plugins will operate under an open governance model, with an independent technical oversight committee making decisions based on public input. Engineers from Amazon, Cursor, Microsoft, OpenAI, and Vercel currently staff the core development team, and the specification carries a Creative Commons Attribution 4.0 International License.
Several parallel efforts are attempting to organize the chaotic landscape of agent development. Google integrated its Agent2Agent agentic communications protocol into AAIF and is constructing a specification for locating tools on the web called Agentic Resource Discovery. AAIF runs a related project to Agent Plugins named Skills Over MCP, which shares capabilities by embedding them directly into the MCP server itself—essentially shipping "the manual with the product," as AAIF Vice President Angie Jones described it. The market will ultimately determine whether MCPs should carry skills or whether skills should bring their own MCP connections. The write-once-run-anywhere promise could reshape how companies invest in AI tooling, particularly as talent moves between platforms and organizations seek to avoid vendor lock-in. Standardization typically follows market consolidation rather than preceding it, suggesting the major players see more value in interoperability than in proprietary moats at this stage of AI agent development.