Open a granted Agent-Native app (Slides, Design, Forms, …) beside the conversation through Dispatch MCP (https://dispatch.agent-native.com/mcp), then work in it using the app’s live UI and screen state. It ships as both a skill and a /an <app> [request] command (commands/an.md), which calls open_app with embed: true and forwards the rest of the arguments as the request.
Working rules: prefer direct page tools (list-host-webmcp-tools / run-host-webmcp-tool) over the app’s own agent. “You are the model doing the work: … author it yourself and save it with the app’s create/update tools. Never hand off authoring to the app’s own agent when those tools exist.” Call view-screen before any selection-dependent edit and never guess an ID from a title. Make the smallest atomic change and read it back, because “a queued ask_app_status task is not proof an edit completed.” The user signs in inside the rendered surface, and the agent never handles passwords, cookies, or tokens.
Off-stage. It is a product front-end, recorded for completeness and for its read-back rule.
See Also
- builder-webmcp — the general in-browser version for any WebMCP page.
- builder-skills — the framework.