MCP, A2A, SDK or Plain HTTP? Pick by Who Is Calling
New protocol names keep appearing, and people ask which one their integration should use. The action underneath, like "create a ticket", is often the same. What changes is who is calling.
Four ways in
- Plain HTTP: your own backend or script calls the API directly. Nothing extra to learn. The others often run over HTTP underneath too.
- SDK: the same calls, wrapped in a library for your language, like tickets.create(), with types and helpers.
- MCP (Model Context Protocol): an AI app like Claude or ChatGPT asks your server for its tools (tools/list) and calls one (tools/call). Build one MCP server, and any app that speaks MCP can use it.
- A2A (Agent2Agent): your agent hands a whole task to another agent, like "sort out this refund", not a single function call. The task has an ID and a state, like working or completed.
How to choose
Your own code calls it: HTTP, or an SDK if a good one exists. A model inside an AI app calls it: MCP. One agent hands a job to another: A2A. They also stack. An agent reached over A2A may use MCP tools inside, and those tools often make plain HTTP calls. Same action at the bottom, different doors at the top.
Do this in 60 secondsPick one integration you own and write down who calls it today. If the answer includes an AI app, look at MCP. If not, HTTP or an SDK is still the right tool.
Today’s AI angleWith MCP, the model reads your tool descriptions to decide what to call. Write them like notes for a new colleague: what it does, when to use it, and what it changes.