WebMCP for browser agents

WebMCP is the third leg of the discovery stool, after agents.json and agent-card.json. The pitch: web pages expose a manifest at /.well-known/webmcp that describes their MCP-compatible tool surface. An in-browser agent reads it, calls the tools, executes intent without scraping the DOM.

It is a W3C Community Group Draft as of late 2025. Chrome 146+ supports it experimentally. Firefox and Safari are observing. Adoption is early.

I ship it on every site I build because the cost is one JSON file. The bet is that one of the major model labs (OpenAI, Anthropic, Google) ships a browser-side agent that uses it before the end of 2026. If they do, every site without a webmcp manifest is invisible to that agent. If they don't, I lost zero engineering time.

The thing webmcp does that MCP and A2A don't: it sits in the page. An MCP server has to be deployed somewhere, an A2A skill has to be advertised in a registry. WebMCP is found by visiting the page. The discovery surface is the URL itself.

If you are building now: spend the 30 minutes, ship the manifest, move on. Revisit when a browser agent that consumes webmcp ships in stable. Until then it is a hygiene play.