How do I call an MCP server manually?

Updated October 2026 · How we answer

Short answerCalling an MCP server manually means sending JSON-RPC messages yourself, usually over HTTP or stdio. You start with initialize, confirm the server responds, then list and call tools.

The basic message flow

An MCP session follows a fixed order. The client sends initialize, the server replies with its capabilities, and the client sends an initialized notification. After that you can ask for tools/list and tools/call. Skipping a step often leads to errors that look like server bugs.

Each message is a single JSON object with jsonrpc set to 2.0. Requests include an id so you can match replies. Notifications leave out the id because no reply is expected. Keeping these rules straight makes manual testing much easier.

  • initialize: start the session
  • notifications/initialized: confirm setup
  • tools/list: see available tools
  • tools/call: run one tool with arguments

Local versus remote servers

Local servers often use stdio, where the client launches a process and writes JSON messages to its input. Remote servers usually use HTTP. Manual calls over stdio are harder to script because you need to keep the process open between messages.

Keep a log of each request and response. It makes it much easier to spot a wrong protocol version, a missing field, or a session problem. Logs also help when you compare two servers that seem to behave differently.

  • stdio: keep the process running
  • HTTP: one POST per message
  • Log every request and response
  • Compare results against the server docs

Common mistakes

  • Skipping the initialized notification and then getting errors on later calls.
  • Mixing transports, such as sending HTTP bodies to a stdio server.
  • Using an outdated protocol version string from an old tutorial.
From our shopsCaseMorph: Type an idea, see a custom phone case in seconds, then print a one-of-one.