One developer's pre-W3C attempt to make websites MCP-native

What it does

WebMCP is a JavaScript widget and local websocket relay that lets website owners expose tools, resources, and prompts to a user’s local LLM client. The site behaves as an ad-hoc MCP server, scoped by domain, while a localhost-only websocket server shuttles requests between the browser and the MCP client. The author explicitly notes this is an early, non-compliant proposal that predates the official W3C effort.

The interesting bit

The clever part is the token-exchange dance: the LLM generates a one-time registration token, the website consumes it to claim a session channel, and then all tool calls flow through a localhost websocket scoped to that domain. It is essentially glue code wrapping the Model Context Protocol for a browser environment, with the heavy lifting happening in a local relay rather than inside the browser itself.

Key highlights

• Websites expose tools, resources, and prompts to any connected local LLM via a drop-in widget
• No API keys change hands; authentication uses ephemeral registration tokens minted by the MCP client
• Supports multiple simultaneous website connections, with tool names prefixed by domain to avoid collisions
• Ships with built-in helpers: a token generator and an MCP Tool Definer for scaffolding schemas
• Keeps the relay entirely on localhost, so traffic flows between your browser and local MCP client without a third-party broker

Caveats

• The author states this implementation is not compliant with the current W3C WebMCP spec and points to the official repository instead
• Security is explicitly flagged as unfinished; the README warns of prompt-injection risks and malicious extension threats
• Some MCP clients require restarts to pick up newly registered tools

Verdict

Worth a look if you want to see what client-side LLM tooling looked like before standards bodies took over, or if you are building a personal MCP bridge. Skip it if you need a production-ready, spec-compliant implementation.