the part that stood out: local models go from basically unusable (8%) to 77% tool accuracy at 100 tools.
cuts input tokens 70-85% at normal catalog sizes.
comment RATEL for the install + github
stops dumping every tool into context every turn.
70-85% fewer input tokens.
local models jump from 8% → 77% tool accuracy at 100 tools.
comment RATEL on the original for the guide
local models with 100 tools go from ~8% to ~77% tool accuracy when you stop dumping the whole catalog every turn.
70-85% fewer input tokens at realistic sizes.
comment RATEL on the original for the install guide + github