Prompt messages disparity between MCP and Claude API

[headinthebox]

Prompts in MCP represent named parameterized message sequences, that enable reuse by invoking them with concrete arguments.

Currently however, the structure of messages in MCP is not the same as that of messages in the Claude API.

• The structure of ImageContent is different in MCP and Claude (has an extra "source" wrapper)
• The structure of TextContent is different in MCP and Claude (has "citations")
• The Claude API also has "tool_use", "tool_result" which are not present in MCP.
• The "resource" message type in MCP and new "document" type in the Claude API serve similar purposes.

These differences make it impossible to take an arbitrary Claude prompt sequence and expose it in MCP, or take an MCP prompt sequence and pass it to the Claude API.

I think that MCP and the Claude API should use exactly the same structure for messages.