This section covers exposing your agents outward — making an agent's tools or its whole self callable by external clients, over two standard protocols.
Overview
An MCP Server exposes a chosen set of an agent's tools and prompts to external Model Context Protocol clients — apps like Claude, an IDE, or any MCP-compatible tool — so they can call those tools directly. An A2A Server exposes a whole agent (or a Team) to other agent systems over the Agent-to-Agent protocol, publishing an agent card that advertises its skills so another agent can discover and call it. This is the reverse direction from External Agents (in Creating & Managing Agents), which is how you consume a third-party agent instead.
What's in This Section
Topic | Covers |
|---|---|
Provisioning a server that fronts tools and prompts for MCP clients. | |
Exposing connector actions as MCP tools. | |
Exposing reusable prompts to MCP clients. | |
Activating the server and its endpoints. | |
How an external MCP client finds and connects to your server. | |
Provisioning a server that exposes an agent to other agent systems. | |
Declaring the capabilities your A2A server advertises. | |
Deploying the server and managing authentication. |
MCP vs. A2A
Use MCP when the caller is a generic tool-using client that wants to invoke specific capabilities. Use A2A when the caller is another agent that wants to hand off a whole task to yours.