webmachinelearning/webmcp

What it solves

WebMCP addresses the difficulty AI agents have when interacting with modern web applications. Traditionally, agents rely on brittle techniques like DOM scraping or simulating mouse clicks, which are unreliable and often bypass the application's internal state and authentication. WebMCP allows web developers to expose specific site functionality as structured "tools" that agents can call directly, ensuring more reliable actuation while keeping the user in the loop.

How it works

WebMCP introduces a client-side API (document.modelContext) that lets developers register JavaScript functions or HTML <form> elements as tools. Each tool includes a natural language description and a structured input schema. When an AI agent (either built into the browser or hosted in an iframe) needs to perform an action, it discovers these registered tools and invokes them with the correct arguments. The browser mediates the call, executing the client-side logic and returning the result to the agent, which then updates the UI or informs the user.

Who it’s for

It is designed for web developers who want to make their applications "agent-ready" without building separate backend integrations, and for AI agent developers who need a stable, developer-defined way to interact with web content.

Highlights

• Client-Side Execution: Reuses existing JavaScript logic and maintains user authentication and session state without needing a separate backend server.
• Human-in-the-Loop: Designed for cooperative workflows where the user maintains visibility and control over the page while the agent performs tasks.
• Hybrid Approach: Supports both imperative registration (via JavaScript) and declarative registration (via HTML forms).
• Secure Delegation: Uses Permissions Policy (allow="tools") to control which iframes or origins can access the registered tools.