AI agents are only as powerful as the tools they can call. On desktop, MCP (Model Context Protocol) lets AI assistants interact with local apps, reading files, querying databases, and sending messages. But on Android? There's no standard way for an AI agent to discover and call tools from other installed apps on the same device. This session introduces a lightweight approach to building on-device MCP servers inside Android apps using HTTP semantics over Binder IPC. No internet. No cloud APIs. No exposed network ports. Just fast, secure, inter-app communication that lets any AI assistant discover your app's capabilities and invoke them — all without data ever leaving the device. I'll live-code an MCP server that exposes tools (like searching contacts or sending messages) as a bound Android Service, and show how an AI agent app can discover, connect, and call those tools at runtime, the same way it would call a cloud MCP server, but entirely on-device.