WebMCP: Exposing Web UI as Structured Tools for AI Agents

AI Engineer2026-06-11go watch the original →

the gist

WebMCP is a proposed browser standard that allows developers to define web page actions as structured tools, enabling AI agents to interact with sites via APIs rather than brittle DOM-scraping.

The Breakthrough

WebMCP replaces unreliable AI screen-scraping and pixel-coordinate clicking with a structured tool-calling interface, allowing agents to execute specific, typed actions directly on web pages.

Implementation Approaches

Declarative API: Developers add tool-name and tool-description attributes to standard HTML forms. The browser automatically generates a JSON schema that agents use to map user prompts to form parameters.
Imperative API: For complex, multi-step flows, developers use the registerTool function. This requires defining a custom JSON schema and an execute block containing JavaScript that performs the DOM manipulation and returns a state result to the agent.

Integration and Testing

Chrome Canary: WebMCP is currently an experimental feature available in Chrome version 146 and above.
Tool Inspector: The Chrome DevRel team provides a Model Context Tool Inspector extension that lists available tools on a page and allows for manual testing of tool calls.
Evaluation: Developers can use the provided eval CLI to test agent performance and reliability on their own sites before full deployment.

AI agents currently struggle to navigate the web because they rely on parsing the entire DOM, accessibility trees, and screenshots, which is token-intensive and prone to failure when layouts shift. WebMCP treats the web as a high-performance API, allowing agents to trigger specific functions like searching, filtering, or purchasing without needing to visually interpret the page. The standard is currently in early preview and intended to complement the broader Model Context Protocol (MCP) by focusing specifically on client-side browser interactions.