mcp2cli -> Save 96–99% of the tokens wasted on tool schemas every turn; caveman -> laude Code skill that cuts 65%.
I keep coming accross many discussionions around context engineering and how that might save you a fortune. It's either cheking the unprecedented claims from yet anotyher MCP/Skill/CLI pluggable into you agent harness. Or discussing some tricks and approaches. With all the heat arounbd AI costs going out of control those directions lool like a reasonable way forward.
Yet I think focusing on token optimnization wouild be a detour and disctration. Not because those claims are not true... E.g. I tested mcp2cli and I never sanely expected the advertised saving (after all who casres if MCP tool defitntion consume 20k of tokens in system prompt given those will be chached and cost peanuts), I just researched ways to use MCP in agents without MCP support... Or my friend shared his experience with cavemen+graphiphy (Opus for planning, Sonnet for execution, same task) he got $16 wtih stock Claude Code and $13.5 with both tools engaged.
When you focus on token accounting you might be missing the whople point of agentic transtion that happened with year. You might save 20% doing some hard core context engineering yet spend 100x more or long horizone agentic execution runnimg multiple subagents.
