Chat-Client Agent Interface
Chat-client agent interface is the pattern of using everyday messaging apps as the front door for an AI agent that performs work on the user’s computer or services.
Key points
- OpenClaw began as a WhatsApp message relay into Claude Code: a message arrived, a CLI command ran, and the result was returned to chat [src-064].
- Steinberger added images because screenshots and photos are efficient context for agents, especially when mobile users need calendar, translation, event, or place decisions [src-064].
- Voice support emerged when the agent inferred an audio file type, converted it, found an OpenAI key, and used speech transcription without being explicitly taught that path [src-064].
- Discord, Telegram, WhatsApp, Signal, and iMessage shift the agent from terminal tool to ambient companion, but public chat surfaces also increase security and prompt-injection risk [src-064].
- Interaction design matters: a no-reply token lets an agent stay quiet in group chats, making the agent feel more socially appropriate [src-064].
- Hermes provides Nate’s Telegram-first version of the pattern: voice notes, quick research, scheduled reminders, comment monitoring, and business check-ins can happen from the phone while the heavier Claude Code workflow stays on desktop [src-074].
- Nate cautions that chat clients hide some of the context and control surfaces that a CLI exposes, so chat-driven agents are best for lower-risk tasks unless their permissions and workflows are tightly bounded [src-074].
Related entities
Related concepts
- System-Level AI Agents
- Agent Personality Files
- Agent Security Boundaries
- Human-Agent Collaboration
- Hermes Agent
- Mobile Agent Work Surface