Source: GitHub Changelog — 2026-08-12
Summary
Agent Plugins 1.0 is a cross-industry, vendor-neutral packaging spec — backed by core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel — that defines a single directory format for bundling AI agent skills and MCP servers so one plugin package works unmodified across VS Code, Copilot CLI, Copilot app, Cursor, Kiro, and more. GitHub's changelog post rolls the spec out to its own Copilot surfaces, with existing pre-1.0 Copilot plugins remaining supported through the transition.
Key Takeaways
- A plugin is just a directory rooted at
plugin.json, a closed-schema manifest permitting only name, version, description, author, and a handful of other top-level fields — no room for vendor-specific extensions to sneak into the core manifest. - Only two component types are defined: Skills (each a subdirectory with a
SKILL.md, non-recursive) and MCP servers, declared entirely in a separatemcp.json— neither can be inlined into the manifest itself. - MCP servers in
mcp.jsonsupport three transports (stdio, streamable-http, and deprecated SSE), and the spec explicitly warns that headers or environment variables are not a secrets mechanism — credentials must not be embedded there. - Two runtime placeholders,
${PLUGIN_ROOT}and${PLUGIN_DATA}, let a plugin reference its own bundled files and a persistent, client-managed data directory that survives updates — both get expanded into subprocess environment and arguments at launch. - Governance is explicitly structured to prevent any single company from controlling the standard: the spec's charter bars any one vendor from holding a majority of Technical Steering Committee seats, with five listed core maintainers spanning Amazon, Cursor, Microsoft, OpenAI, and Vercel.
Reel Script
Hook: Right now, if you build a skill or tool for one AI coding assistant, it usually only works in that one assistant. Five of the biggest names in AI tooling just agreed on a packaging format meant to end that — build once, and it runs in Cursor, Copilot, and half a dozen others.
Core Concept: Every AI coding tool — VS Code Copilot, Cursor, Kiro, and others — has its own way of letting you extend it with custom skills or connect it to external tools via MCP, the standard protocol agents use to call things like databases or search APIs. That fragmentation means a developer building a reusable capability has to package it differently for every client it might run in. Agent Plugins 1.0 fixes that with a deliberately minimal, shared package format: a directory with a manifest file describing the plugin, a folder of skills, and a separate file declaring any MCP servers the plugin needs to talk to. It's intentionally narrow — only two kinds of components are allowed, skills and MCP servers, with a locked-down schema that doesn't let individual vendors bolt on proprietary extensions to the core format, which is exactly what keeps it portable instead of turning into another fragmented ecosystem.
Hands-On: The structure worth sketching is genuinely simple: a plugin.json manifest sits alongside a skills/ folder — each skill just a subdirectory with a SKILL.md file inside — and an mcp.json declaring any external tool servers, specifying whether they connect over stdio, or the modern streamable-http transport. Two placeholder variables do real work here: ${PLUGIN_ROOT} points at the plugin's own bundled files, and ${PLUGIN_DATA} points at a persistent directory the client manages for the plugin's cached data, so an update to the plugin doesn't wipe out what it's learned. And the part that matters just as much as the technical format is the governance structure sitting on top of it — the steering committee has named maintainers from five different companies, with an explicit charter rule that no single company can hold a majority of seats, which is the actual mechanism keeping this a genuinely shared standard instead of one vendor's format with other logos attached.
Takeaway: If you're building anything meant to extend an AI coding assistant, target this spec instead of a single vendor's proprietary plugin format — you get every major client for the same packaging work. Go look at the spec repo before you build another single-platform integration you'll have to rewrite in six months.