Model Context Protocol (MCP) is an open standard for exposing tools to an AI. Tool servers is where you connect them: your agents call out to these servers. It is not the same as Developer → AI clients, where outside AI apps such as Claude call in to your account. Connecting a tool server hands your agents every tool that server publishes at once, instead of you defining each one by hand.
MCP or custom tools? #
| Field | What it does |
|---|---|
| Use custom tools | When you are wiring up one or two of your own endpoints. Direct, and you control exactly what is exposed. See Custom tools. |
| Use MCP | When something you already run — or a third party — publishes an MCP server. You get its whole toolset and it stays current as that server changes. |
Connecting a server #
-
Add the server
Give its address and whatever credentials it requires. -
Refresh
This calls the server and reads back the tools it publishes. Refresh is rate-limited, because it makes an outbound request from a browser action. -
Review what it exposes
Read the tool list before attaching it to anything. You are adopting all of it. -
Attach it to an agent
The tools then become available to that agent exactly like custom ones.