You were right — every flat tool was burning context. Took the BloodHound MCP approach and went all in:

207 tools → 26 composite (parameter dispatch)
~6K lines of flat implementations eliminated
Single composite module, no dead files
Token overhead ~4K → 1.8K per query

measure_tokens.py in the repo uses the same cl100k_base methodology if you want to compare. Thanks for the nudge — made the project genuinely better.
