Cut Claude Code token bills with tier routing, a thinking-token cap, and MCP pruning

Build to Launch·high signal

Agentic coding sessions routinely burn 40K–80K tokens before a line is written (subagent fan-out, autocompact cascades, and MCP servers loading 18K+ tokens per turn each), pushing average spend to $400–$1,500/mo. The four levers that actually work: per-user budget caps, prompt caching of system instructions, model-tier routing (Haiku for grunt work, Opus for hard reasoning), and aggressive context pruning. Lower the ceiling directly with `MAX_THINKING_TOKENS=8000`, drop `/effort`, or disable thinking in `/config` for task classes that don't need it.