TLDRocket
Sign in

Agent Plugins

Agent Plugins

Amazon, Cursor, Microsoft, OpenAI and Vercel just agreed on a shared plugin format for AI agents. Translation: build a plugin once, and it should work across way more agent tools without extra rework.

Based on reporting by Agent Plugins — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Every AI agent maker has been quietly reinventing the same wheel. Cursor has its plugin format, OpenAI has its own, Microsoft and Amazon each cooked up something slightly different, and if you wanted your Agent Skill or MCP server to work across all of them, you were stuck copy-pasting and restructuring files client by client. That busywork is exactly what a new spec called Agent Plugins is trying to kill off.

The 1.0.0 release, backed by a Technical Steering Committee that includes core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel, lays out a plain directory structure any compatible client can read. A plugin.json manifest sits at the root, declaring what version of the spec it targets. Below that, a skills/ folder holds Agent Skills defined by their own separate spec, and an mcp.json file describes MCP servers, whether they talk over stdio, Streamable HTTP, or the older HTTP+SSE transport. It's a deliberately thin layer — just enough shared scaffolding that the components themselves don't need to be rebuilt for every tool.

What's notable is what the spec doesn't try to standardize. Installation, permissions, distribution channels, and the actual user experience stay entirely in each client's hands. Vendors that want to bolt on client-specific behavior can do it through reverse-domain namespaced folders, like a com.example.client/hooks directory sitting alongside the portable parts, without touching or breaking the shared core. It's a narrow interoperability floor rather than a full platform, which is probably why five competing companies could actually agree on it.

The project is developed in the open on GitHub, with new feature ideas required to start as Discussions posts that demonstrate real portability needs and actual implementer interest before anything gets merged. That's a fairly deliberate choice — it keeps the spec from ballooning with speculative features nobody will implement, a trap a lot of standards efforts fall into once they get momentum and a big-name steering committee behind them.

My take — AI-written commentary, not fact-checked reporting

Getting Amazon, Microsoft, OpenAI, Cursor, and Vercel to agree on anything is rare enough that it's worth paying attention to what they agreed on: almost nothing, by design. That's the smart move — a thin, boring interop layer survives; an ambitious one gets abandoned the moment a vendor wants to differentiate. Plugin authors should be thrilled, because the alternative was an ecosystem where every agent client quietly became its own walled garden.

Read more about this at: Agent Plugins

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.