Agent Plugins is an open, vendor-neutral specification for packaging reusable components that extend AI agents into distributable plugins. Version 1.0.0 was published on 6 August 2026 and covers two component types: Agent Skills and Model Context Protocol servers (Source: vercel.com). The specification is maintained at agent-plugins.org and developed in the open on GitHub; its own status line records it as "Spec Version: 1.0.0, Status: Working Draft" (Agent Plugins Specification v1.0.0 (Working Draft, August 2026)).
Scope
The specification defines a packaging and discovery contract and nothing beyond it. It standardises the directory layout, the manifest, component discovery, validation and loading, and provides a conformance checklist for client implementers. It does not address marketplaces, permissions, or runtime environments (Source: the-decoder.com). Vercel described the choice of two component types as deliberate, on the ground that Skills and MCP servers are the parts of an agent extension that are portable across clients today (Source: vercel.com).
Structure
An Agent Plugin is a directory containing a plugin.json manifest at its root. The manifest's only required fields are a schema reference identifying the specification version and a name; a version and description are conventional additions (Source: composio.dev). Agent Skills are discovered under a skills/ directory and MCP server configuration is read from mcp.json. Components are validated independently, so an invalid component does not disable the others in the same package (Source: vercel.com).
Version 1.0.0 leaves hooks, commands and sub-agents to individual clients and to possible future versions, and provides a namespaced extension mechanism for client-specific data. The specification, its JSON Schemas, and separate guides for plugin authors and client implementers are published at agent-plugins.org (Agent Plugins Specification v1.0.0 (Working Draft, August 2026)).
Governance
Vercel initiated the proposal, which representatives from Amazon Web Services, Anysphere, GitHub, Microsoft, OpenAI and Vercel refined into version 1.0.0. The initial Technical Steering Committee comprises Core Maintainers from AWS, Cursor, Microsoft, OpenAI and Vercel. Vercel stated that "no single company's product roadmap sets the format's direction" (Source: vercel.com). Google's developer blog described the specification as "published by a TSC of Core Maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel" (Source: developers.googleblog.com). Specification development takes place in a public repository (Source: the-decoder.com).
Anthropic is not among the participating organisations. The Decoder noted the absence, observing that Anthropic created both the Model Context Protocol and Agent Skills as open standards and had separately added a plugin system to its own agent product (Source: the-decoder.com).
Adoption
At the 6 August 2026 launch the format was supported by ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and VS Code (Source: vercel.com). Amazon Web Services shipped support in Kiro Powers and made the AWS Agent Toolkit, which bundles more than 30 curated skills across services including Lambda, S3, DynamoDB and CDK, compatible with the specification (Source: aws.amazon.com). Microsoft had supported its own agent-plugin packages in VS Code since earlier in 2026 and updated its documentation on 6 August 2026 to describe the open cross-client format (Source: virtualizationreview.com).
Commentary has noted the working-draft status alongside the 1.0.0 version number (Source: mlq.ai).
Relation to other frameworks
Agent Plugins sits one layer above the two component specifications it packages. The Model Context Protocol, documented at Model Context Protocol (MCP), defines how an agent connects to external tools and data; Agent Skills define reusable instructions and workflows. Agent Plugins defines how both are bundled and discovered rather than what either can do. It is distinct in kind from assurance-oriented instruments such as NIST AI Risk Management Framework 1.0, ISO/IEC 42001 — AI Management System and Five Eyes Joint Guidance on Secure Deployment of AI Agents (May 2026), which govern risk management and security posture rather than package format.
Provenance note: citations here draw on the specification site and on vendor announcements. They should upgrade to the primary specification text once it is ingested.
Relationships
- depends-on: Agent Plugins Specification v1.0.0 (Working Draft, August 2026) — the primary text.
- depends-on: Model Context Protocol (MCP), Agent Skills, Agentic AI.
- related: Agent Architecture Patterns, Open-Source AI / Open-Weight Models, OpenAI, Microsoft, Five Eyes Joint Guidance on Secure Deployment of AI Agents (May 2026).