MCP and AI basics
MCP server
A program or hosted service that exposes a system's abilities (tools), data (resources) and ready-made instructions (prompts) in the format MCP clients understand. Examples: a Google Ads server, a Shopify store endpoint, a Klaviyo server.
An MCP server is the program that sits between an AI assistant and one system, such as an ad account, a store or an email platform. It publishes a list of tools with their input schemas, answers tool calls by talking to the underlying API, and returns the result in a form the assistant can read. Some servers also publish resources (data to read) and prompts (reusable instructions).
Servers come in two shapes. A remote server is hosted by the vendor and reached over HTTPS, usually after an OAuth sign-in; you add it by URL. A local server runs on your own machine or infrastructure, often started with a command, and holds its credentials locally. The directory shows which shape each server takes under 'Hosting'.
The server decides what is possible, not the assistant. A read-only reporting server cannot change a campaign no matter how the prompt is phrased. A server with write tools can, which is why the access level and the vendor's own cautions matter more than the server's feature list when you pick one.
Where it shows up on this site
- Browse every server in the directory →
- Best MCP servers for e-commerce and marketing teams →
- This site's own MCP server →
See also: Tool, Resource, Prompt (MCP)