curated · 2026-08-06

This is worth a click because it fixes a real pain point: you build a skill and an MCP server, it works on one client, then you try a second client and the directory layout, manifest format, and MCP config are all different. You fork, maintain two copies, and they drift. Agent Plugins is deliberately minimal—a fixed directory structure, a plugin.json that only declares a name and schema version, skills in skills/, MCP servers in mcp.json with explicit transport types. No discovery paths, no precedence order, no inline declarations. Components fail independently. What v1 deliberately omits is more telling than what it includes: no install mechanism, no permission model, no sandboxing. All left to each client. So you can package and distribute now, but you'll need to handle the security boundary yourself. Google joining as a core maintainer rather than launching their own thing signals they're aligning with the industry here instead of fragmenting further. One detail I appreciate: the post explicitly says a single skill or single MCP server doesn't need a plugin. The format earns its keep only when components must travel together. That's a refreshingly pragmatic line to draw—they're not trying to make this the answer to everything.

This caught my eye because it pushes MCP from "wrap an API" to "wrap an entire website." The approach: use browser rendering to turn pages into structured data, then expose them via MCP to tools like Claude or Cursor — zero changes to the site's own code. For developers, that saves the grunt work of writing scrapers and parsers for every site. I'd discount the "any website" claim a bit though. Sites with clean, predictable structures — e-commerce, SaaS — will probably work fine. Anti-bot measures or complex multi-step flows? Likely still a mess. The thing to watch: if Cloudflare bundles WebMCP pricing with Workers and Browser Rendering usage, it could lower the cost barrier for agents to interact with websites. Right now, that pricing info is exactly what's missing.

The reason this caught my eye: Wan3.0 lowers the bar for video generation by accepting URLs, PDFs, or slides directly—no reformatting. It outputs a 30-second clip with controllable motion, aiming to turn a spec sheet into a concept ad in one step. I'd hold off on getting too excited though. The post doesn't mention generation latency, pricing, or model parameters. If a 30-second clip takes several minutes and costs by the second, the "production-ready" claim gets shaky. Right now we only have a title and a product blurb—real judgment waits for hands-on numbers.