Token Cost Optimizer — Smart Model Routing

The default OpenClaw setup sends every message to the same model. If that model is Sonnet or Opus, you are burning expensive tokens on tasks that Haiku handles just as well. Combine that with heartbeats firing every 30 minutes with your full session context loaded, and some operators are spending $150–600/month without realising most of it is waste.

This skill is the fix. It includes a complete model routing framework — a decision tree your agent uses to choose Haiku for simple tasks, Sonnet for standard work, and Opus only when complexity genuinely demands it. Community testing shows this alone reduces costs by 60–70%.

The heartbeat section is where the biggest savings live. With isolatedSession: true and lightContext: true, a heartbeat turn drops from 170K–210K tokens to under 3K. At 24 heartbeats per day that is a 98% reduction on heartbeat cost alone.

Also included: QMD (Quick Memory Database) setup guide — OpenClaw's local SQLite semantic search that injects only relevant memory snippets instead of loading your full MEMORY.md on every turn. Expected memory token savings: 60–97%.

Everything here is from a live production deployment. These are the exact configs. Setup time for the heartbeat fix: under 5 minutes.