Summary

This SEP proposes adding an optional conversationEvents capability that allows MCP servers to subscribe to conversation-level events. When a user sends a message, subscribed servers receive the content and can respond with context to inject into the LLM's prompt.

Motivation

Building a memory system for AI assistants, I hit a wall: MCP servers only receive data when the LLM explicitly calls a tool. If it doesn't call recall(), the memory system is blind.

Workarounds like strong tool descriptions, piggybacking on other tools, or client-side hooks don't reliably solve this. The protocol needs to support automatic context injection.

Use Cases

• Memory systems that surface relevant past conversations
• Knowledge bases that inject relevant docs based on the question
• Project context that reminds assistants about codebase conventions
• User preferences that personalize responses automatically

Key Points

• New conversationEvents capability with onUserMessage subscription
• Servers respond with context to inject (plain text or structured)
• 500ms timeout to prevent blocking
• Purely additive - backward compatible with existing servers/clients

Full specification in the SEP file.

Happy to iterate on the details. The core ask: let servers see the conversation so they can provide automatic context.