Agent Plugins 1.0: your AI skills stop being vendor-locked
OpenAI, Amazon, Microsoft, Cursor and Vercel shipped Agent Plugins 1.0, a portable format for Agent Skills and MCP servers. Build the workflow once.
Five companies that compete hard on coding agents just agreed on a file format. Agent Plugins 1.0 packages Agent Skills and MCP server configs into one directory that any supporting client can load. If you have been rewriting the same agent setup for Claude Code, then Copilot, then Cursor, that work just became portable.
What actually happened
On August 6, OpenAI, Amazon, Microsoft, Cursor and Vercel published Agent Plugins 1.0 as a vendor-neutral standard, with GitHub also shaping the spec. Those five hold the technical steering committee seats; development happens in public on GitHub.
The format is deliberately boring, which is the point:
plugin.jsonat the root — identifies the plugin and the spec version it targets.skills/— Agent Skills, the reusable instruction sets that tell an agent how your team does a thing.mcp.json— the MCP servers the plugin needs to reach live tools and data.- Client-specific extras live in reverse-domain namespaces, so one vendor's extension doesn't break another's loader.
Launch-day support spans ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code, per The Next Web. Nothing new was invented here. MCP already handled the connection layer; Skills already handled the instruction layer. Agent Plugins is the envelope that lets you hand both to a different tool without a rewrite.
Why agent portability matters for your business
Every hour your team spends encoding institutional knowledge into an agent — the deploy checklist, the way you categorize a commit, the query patterns for your warehouse — has been trapped in whichever vendor you used that quarter. That is a real switching cost, and vendors have not been shy about it. This spec turns that asset into something you keep.
Three things worth doing about it.
Treat your skills as source, not settings. Put the plugin directory in your repo, version it, review changes to it like code. A skill that lives in a vendor's web UI is a skill you will re-type in eighteen months.
Write the workflow, not the prompt. The portable unit is a procedure with the MCP servers it needs. "Summarize this" is not worth packaging. "Categorize commits against our release-notes taxonomy, using our Linear MCP server, and open the PR" is.
Do not confuse a standard with a guarantee. Five vendors on a steering committee is a strong signal, not a contract. Version 1.0 is where the format is simplest and adoption is thinnest — build on it, but keep the actual business logic in the skill file rather than in whatever client-specific extension shows up next.
The pattern here is the one we keep repeating: the layer you own should be the layer that encodes how your business works. Models change, clients change, pricing changes. A directory of skills and MCP configs sitting in your git history does not.
Key takeaways
- Agent Plugins 1.0 shipped August 6 from OpenAI, Amazon, Microsoft, Cursor and Vercel, with GitHub
- A plugin is a folder:
plugin.json, askills/directory, andmcp.json - Launch support includes ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code
- Your agent workflows become portable across clients instead of trapped in one vendor
- Version the plugin directory in your repo and review it like code
- Package procedures with their tool connections, not one-line prompts
The agent config should outlive the vendor. We build agent workflows as versioned artifacts in your repo — skills, tool connections, and guardrails you can move to whatever client wins next. See how we architect it or tell us what your team keeps re-teaching its agents.
Sources: Agent Plugins specification, The Next Web.
- #agent-plugins
- #mcp
- #ai-agents
- #developer-tools
- #portability
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
NVIDIA's NOOA makes an AI agent one Python class
NVIDIA open-sourced NOOA, an agent framework where methods are actions and type hints are enforced contracts. A 253-line agent hit 82.2% on SWE-bench Verified.
Read itAirbnb shipped 80% more features with AI. Measure what ships
Airbnb says AI cut concept-to-launch time 60% and helped it ship ~80% more features on flat headcount. The metric that matters isn't code written — it's work shipped.
Read it