Karna: A Chat-Native, Multi-Channel Architecture for Personal AI Chief-of-Staff Agents

Personal AI assistants are moving from single-window chat interfaces toward systems that operate across messaging channels, web dashboards, mobile clients, voice notes, workflow triggers, and long-running memory. This paper presents Karna, a self-hosted TypeScript monorepo that explores a chat-native architecture for a personal AI chief-of-staff agent. The system is organized around a gateway, an agent runtime, shared protocol types, channel adapters, web and mobile surfaces, a CLI, memory/session modules, workflow hooks, observability surfaces, and deployment artifacts. Repository inspection identifies 13 channel packages, 485 TypeScript or TSX source files, 84 test files, Docker and Kubernetes deployment artifacts, and separate web, mobile, cloud, and command-line applications. The paper describes the design goals, repository structure, channel abstraction, operational safety considerations, and research questions that emerge from using chat as the primary interface for agentic work.