enhancement: dynamically loaded MCP config for subagents #114
Description
So, still kind of chewing on this idea, but thought it might be good to start some discussion on the topic.
With the advent of subagents, many mcp users (myself included) are now leveraging them as isolated context windows with specialized prompts and toolsets. I didn't see anything in the documentation about it, but it'd be excellent if I could leverage mcphub utilities to help sandbox my subagent toolkits.
In thinking about it, at least for Claude Code, it may be more efficient to just circumvent mcp-hub and directly list the prompts/tools in the agent config, but i think the ideal scenario would be something like allowing for a given client request to come alongside some config headers, which request a given set of tools/prompts/resources/the other thing i always forget to be enabled for that agent, which could be assigned a UUID when requesting a given config as a way to manage the config state.
My goal with this is to minimize excess token usage in any given subagent; as it stands, mcp-hub is an excellent tool when I'm only reconfiguring a single agent for a task, waiting until completion, and changing config as needed, but it'd be double-excellent if I could use it to solve this problem too.
Curious what others' thoughts on the matter might be, or any alternative approaches I haven't considered. I feel like tracking state per uuid miiight get a little messy, if we're tracking client connections as well? But could use sqlite or something to manage it if it gets bulky.
Anyway, open discussion thread, I guess, lol.