WebMCP: Connecting Web Apps to AI Agents via JavaScript Tools

The WebMCP API is a new JavaScript interface that allows web developers to expose application functionality as tools for AI agents. By providing natural language descriptions and structured schemas, web pages can function as client-side servers for LLM-based assistants. The API includes robust methods for tool management and ensures user oversight through interactive callbacks during execution.

Key Points

• WebMCP enables web applications to act as client-side Model Context Protocol (MCP) servers by exposing JavaScript functions as tools.
• Tools are defined using a combination of executable code, natural language descriptions for agent discovery, and structured JSON schemas for input validation.
• The API extends the standard Navigator interface to include a ModelContext attribute for managing tool registration and lifecycle.
• A dedicated ModelContextClient interface allows AI agents to trigger user interactions, ensuring users remain in control of actions taken on their behalf.
• The specification is designed to support collaborative workflows between users and autonomous agents within a shared web environment.

Sentiment

The community is cautiously divided. There is genuine intellectual interest and some enthusiastic supporters who see WebMCP as a natural evolution of the web platform. However, significant skepticism centers on the absence of a security model, questionable adoption incentives for businesses, and whether simpler alternatives could achieve the same goals. The overall tone is more curious-but-wary than outright hostile, with constructive debate outweighing dismissiveness.

In Agreement

• WebMCP fills a genuine gap for browser-based AI agents that need to interact with live page state and user sessions, unlike traditional server-side APIs which are disconnected parallel experiences
• This could become the 'responsive design for agents' — the web adapted for mobile, now it needs to adapt for AI agents as the next class of consumers
• Static sites and SPAs benefit most, as they can expose structured tools without needing server-side infrastructure or increased hosting costs
• Having agents call structured tools is vastly preferable to having them scrape DOMs or take screenshots, and WebMCP provides a standard way to do this
• The form-to-tool auto-conversion extension is a practical bridge that piggybacks on existing semantic HTML rather than requiring entirely new markup

Opposed

• The blank security and privacy sections in the proposal reflect rushed, hype-driven development — MCP's protocol provides no authorization method and WebMCP inherits those gaps while exposing them to every page visitor's browser
• The three-party auth problem (user, agent, service) is fundamentally unsolved and must be addressed before WebMCP can be standardized
• Existing approaches like accessibility trees, OpenAPI specs, SKILL.md files, and semantic HTML could serve the same purpose without adding to the web spec
• Businesses have little incentive to implement WebMCP since it makes their data easier for AI agents to consume, mirroring the dynamic that killed RSS
• This could bloat the web spec with something ephemeral — coding agents may evolve past the need for specialized protocols, and DOM agents already handle most navigation effectively