One process per configured MCP server. It owns a transport, performs the MCP handshake, keeps the server's tools published in Mead.MCP.Catalog, and carries tool calls.

Lifecycle

:disconnected -> :initializing -> :listing -> :ready. When the transport cannot start, closes, dies, the handshake does not finish in :handshake_timeout, or a tool refresh does not finish in :refresh_timeout (which defaults to the handshake timeout), the connection does not crash: it removes its tools from the catalog, fails every pending caller with {:error, :connection_lost}, stops the transport, and tries again after the next delay of :reconnect_delays (the last one repeats). A missing command or a server that is down therefore costs one retry per delay and never becomes a restart loop that could take the application down.

Only messages from the current transport are honoured: a late message from a replaced one is ignored. The connection traps exits so that a transport that dies is an ordinary event.

Calls are carried by call_tool/5; each pending call monitors its caller and has its own timer.