webmcp-tool-annotations

WebMCP tool annotations are safety metadata hints like readOnlyHint, destructiveHint, and idempotentHint that inform browser permission prompts and guide agent behavior during agentic browser interactions. They ensure users are appropriately prompted before sensitive actions occur.

To mark browser tools for an AI agent, implement safety annotations by applying readOnlyHint for non-mutating actions like searchProducts, or destructiveHint for sensitive operations like placeOrder. This ensures the browser prompts the user before checkout but allows searches without interruption.

Use idempotentHint when annotating browser tools that can be safely repeated without changing the result beyond the initial application. Combining it with readOnlyHint or destructiveHint requires understanding potential annotation combinations to effectively enhance security and user experience in agentic browser interactions.

No specific dependencies are required to implement safety hints for browser-based tools. You only need an understanding of annotation types, browser usage contexts, and potential combinations to effectively apply readOnlyHint, destructiveHint, and idempotentHint for secure agentic browser interactions.

The best way to configure browser permission prompts for e-commerce agent tools is to use WebMCP tool annotations. Correctly annotate placeOrder as destructive and searchProducts as read-only to inform browser security, ensuring the browser prompts the user before sensitive checkout actions.