Agent Plugins 1.0
Skills you build once, running everywhere.
OpenAI, Amazon, Cursor, Microsoft, and Vercel have published "Agent Plugins," a shared packaging standard for AI agent extensions. Within days, Google joined the technical steering committee too, bringing the count to six. Skills and MCP servers you build once are starting to need less rebuilding every time you switch tools.
Why Now
Rebuilding for every client
used to be the norm
Anthropic's Agent Skills spec already defined a package format that bundled instructions and code together. But when a skill depended on a specific MCP server, wiring the two together still required separate — and often undocumented — setup steps for every single client. Until last month, that meant rewriting the same wiring once for ChatGPT, once for Cursor, once for VS Code, and so on for every tool you supported.
To close that gap, OpenAI, Amazon, Cursor (maker Anysphere), Microsoft, and Vercel published version 1.0.0 of a shared packaging standard called "Agent Plugins" on August 6, 2026.The specification and its governance are published and maintained on GitHub at agentplugins/agent-plugins-spec. Its Technical Steering Committee lists Clare Liguori of Amazon, Roshan Sadanani of Cursor, Harald Kirschner of Microsoft, Gav Verma of OpenAI, and Jonathan Hefner of Vercel as Core Maintainers, with Hefner serving as Lead Core Maintainer.
Anatomy
The core is simple:
one plugin.json
The spec is deliberately kept small. A plugin is just a directory with a root manifest, plus a skill and an MCP config in fixed locations.
Per the published plugin manifest schema (v1.0.0), the only required field is name (1-64 characters, lowercase alphanumeric with hyphens/dots, no consecutive separators). version, description, author, homepage, repository, license, and keywords are all optional, alongside an extensions field that lets clients store their own data under a reverse-domain namespace. The spec text is licensed CC-BY-4.0 and the code Apache 2.0.
Write plugin.json
One manifest at the root. Since name is the only required field, it's easy to bolt onto an existing project.
Drop SKILL.md files under skills/
Package the instructions and workflows you want to reuse as Agent Skills.
Add mcp.json if needed
If the skill connects to external tools or data, bundle the MCP server config alongside it.
By the Numbers
Agent Plugins,
right after launch
Google joined the technical steering committee as a Core Maintainer shortly after launch, bringing the count to six companies.The six clients that announced day-one support are OpenAI's ChatGPT and Codex, Cursor, Microsoft's GitHub Copilot and VS Code, and Amazon's Kiro — each with its own official announcement of the integration.
Who It's For
Three angles on the impact
Same news, different stakes depending on where you sit.
Engineers
Skills and tool wiring you build now can potentially move across ChatGPT/Codex, Cursor, GitHub Copilot, VS Code, and Kiro without a rebuild. The fastest way to check is to migrate one existing skill to the plugin.json schema and confirm it actually runs on more than one client.
Product / PM
An agent extension you build once internally becomes easier to roll out across multiple AI tools. It's a genuine new data point to bring into vendor-selection and tool-comparison conversations about switching cost.
Business / Leadership
Investing in one vendor's agent platform now looks less likely to lock your team's skill assets in with it. But the spec is still being shaped — switching costs did not drop to zero today.
| Before Agent Plugins | After Agent Plugins 1.0 |
|---|---|
| Skill/MCP wiring configured separately per client | One plugin.json distributable across multiple clients |
| Wiring steps often informal and undocumented | skills/ and mcp.json placement defined in a public CC-BY-4.0 spec |
| Each vendor held the format on its own | No single company can hold a majority of TSC Core Maintainer seats |
| Client support was scattered and inconsistent | ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and Kiro support it at launch |
Skills go from something you rebuild
to something you ship and run.
What's Next
Between "it works" and
"it's production-ready"
Over the past year, Agent Skills grew popular fast, and MCP servers made connecting to external tools routine. Both pieces were maturing on their own, but there was no shared container to bundle and ship them together — that's the exact gap this spec fills. What sets it apart from the usual pile of vendor-specific extension formats is that Amazon, Cursor, Microsoft, OpenAI, Vercel, and now Google — the major developer-tool vendors — all lined up behind the same format at once, rather than one company shipping its own standard.
Still, this isn't unambiguously good news. The spec just hit v1.0.0, and real-world cases of the exact same plugin.json running unmodified across multiple clients are still limited. The moment a plugin leans on the client-specific data allowed under the extensions field, some of that portability disappears. And as Google's late arrival shows, this wasn't locked down by the original five companies alone — scope and emphasis can still shift. It's also unclear when, or whether, vendors absent from the launch list and governance docs — Anthropic among them — might join.
Three concrete next steps: (1) teams with existing skills or MCP configs should migrate one to the published plugin.json schema as a pilot. (2) Keep tracking which clients beyond the initial six announce support. (3) Check for any dependency on the extensions field before migrating, and confirm a plugin genuinely runs across more than one client before committing fully.