MCP Tool Descriptions - Not the best update
I think the previous implementation of the MCP we had was better than the current. Discovery I just made syncing my fork.
Before:
After:
As mentioned in my previous post, Claude Code now loads full schema for every mcp tool.
Having the MCP Tool allocation go from 5k tokens to 16k tokens is too much.
https://github.com/coleam00/Archon/issues/404
I'll stick with my old fork for now but we need to take care of this.
I appreciate your thoughts here @Abdo-El-Mobayad! @Wirasm was looking into this yesterday and it seems to be a problem specifically with Claude Code. With other clients, we aren't seeing a token increase and I know Claude Code has had some updates recently. However since Claude Code is the most common coding assistant right now we do want to look into this and see what's up!
I was about to write the same :) but also adding the color that despite the high token consumption, Claude Code doesn't really realise it can go to Archon to get information. It fetches tasks, because I ask it to, but then despite having documents in the knowledge (e.g architecture docs), it doesn't think about using archon instead of searching the code or using its own model data. IMO this worsens the token usage / usefulness ratio. Otherwise Archon is a great idea with some really good things :)
That's the kind of thing you need to configure into your ClaudeMD and dev workflow core files.
For me, Claude Code always creates a task or multiple. Always checks previous completed tasks, then project docs, then RAG (2 types), then internet research if needed.
This is all AI behaviour configuration that you need to do.
You can check my post here where I discuss this more https://github.com/coleam00/Archon/issues/404
I can work on this. @Wirasm