One AI agent your users reach on WhatsApp, Slack, Telegram, email or an HTTP webhook. Every channel is an entry point to the same agent — same system prompt, same MCP tools, same conversation history — and the reply goes back out the channel the message arrived on.
An agent is a model, a system prompt and a set of MCP servers. LLM config is per agent, so you point it at any OpenAI-compatible provider with your own key. A channel is an entry point bound to a default agent, with command routes, access guards and rate limits. Everything is scoped to a workspace, which is the tenant boundary.