webmachinelearning/webmcp

What it solves

WebMCP addresses the challenge of AI agents struggling to interact with complex web applications. Currently, agents often rely on "brittle" methods like scraping the DOM or simulating mouse clicks to perform tasks. WebMCP allows developers to expose specific, reliable client-side functions as "tools" that AI agents can directly invoke, enabling more efficient and accurate collaboration between users, web pages, and AI.

How it works

Web developers use the document.modelContext.registerTool() API to define JavaScript functions with natural language descriptions and structured input schemas. When an AI agent is active on the page, it discovers these registered tools. When the user requests a task, the agent calls the tool, the browser executes the page's own client-side logic, and the page returns a structured result to the agent. This keeps the UI in sync with the agent's actions without requiring a separate backend server.

Who it’s for

This project is for web developers who want to make their applications "agent-ready," allowing AI assistants to perform actions within their sites more reliably, and for browser/agent developers looking for a standardized way to interact with web content.

Highlights

• Client-side focus: Allows reuse of existing application logic and state management directly in the browser.
• Improved reliability: Replaces error-prone UI scraping with structured, developer-defined tool calls.
• Cooperative workflows: Designed for human-in-the-loop scenarios where the agent and user work together within the same interface.
• Security-conscious: Integrates with browser permissions policies to control which agents or iframes can access registered tools.