Hey folks, really excited to see an official remote SSE server from you!
I've been testing it out and wanted to share some observations about the current architecture. The tool-based approach works well, though I think there might be some value in incorporating resources for certain use cases.
A couple of things I've noticed in my workflow:
Client behavior differences: In Claude Code, tool calls need approval while resources are treated like workspace files. Some other clients have similar patterns where resources flow more naturally.
Model efficiency: Since some servers end up exposing resources as tools anyway, using native resources can reduce the total tool count presented to the LLM.
Relationship mapping: Resources work really well for showing ticket relationships through labeled links.
Common patterns: You probably see this already from internal usage, but there are recurring queries like "What's the latest on TKT-1234?" that could benefit from the MCP prompt system for templating.
The MCP architecture seems to break down along access control lines - resources for contextual data managed by the host, tools for actions with side effects, and prompts for templated workflows.
Really appreciate you making this available in public beta. Curious to hear how you're thinking about these architectural choices as you continue developing it!