The AI app opens one connector for each service.
Each connector is an MCP client. It keeps a separate connection to one MCP server.
Visual guide 02
MCP gives an AI application a standard way to connect to tools and data. Follow one request from the user to an external system and back.
MCP standardizes the connection between the AI app and the server.
Follow one request
Each connector is an MCP client. It keeps a separate connection to one MCP server.
Local connections use standard input and output. Remote connections use Streamable HTTP.
They exchange protocol versions and supported features before the app sends a request.
A server can offer actions, information, and templates. MCP calls them tools, resources, and prompts.
Find incidents
Read a handbook
Review a support case
For this request, it chooses Find incidents with checkout and open. The inputs must match the tool schema.
incidents_searchIt validates the inputs, applies its access rules, and uses its own credentials to call the incident tracker.
The app adds it to the conversation. The AI can now answer the user.
Here are the 3 open incidents.
Control points
The host, transport, and server each control a different part of the request.
Local stdio servers usually get credentials from their environment. MCP OAuth applies to HTTP connections.
The protocol boundary
Official sources
This guide follows the stable MCP specification dated November 25, 2025.
MCP in production
See how Amelfi adds policy, credentials, and audit controls around MCP traffic.