OpenClaw's heartbeat runs on a schedule and can add up in API costs if you use a top-tier model for every run. The good news: most heartbeat checks don't need your full SOUL.md, memory, or agent personality. They just need to run a task and optionally report results.
Option 1: Use --light-context
When defining your heartbeat in HEARTBEAT.md, add the --light-context flag to the command. This tells OpenClaw to load minimal context — task instructions only, no workspace files. Perfect for:
- Disk space checks
- RSS or news summaries
- Weather checks
- Simple status probes
## 9:00 — Morning digest
- schedule: every day at 9:00
- command: openclaw run --light-context --message "Summarize my overnight notifications and today's calendar."
Option 2: Route heartbeat to a cheaper model
In openclaw.json, you can use bindings to route specific agents or tools to different models. Create a lightweight "heartbeat" agent that uses a cheaper model (e.g. GPT-4o-mini, Claude Haiku, or local Ollama) for routine tasks, and reserve your main model for interactive chat.
Option 3: Reduce frequency
If you run heartbeat every 15 minutes, consider 30 or 60 minutes for non-critical checks. A daily digest doesn't need to run hourly.
See also: The Cost Optimization guide and the Cost Calculator for more ways to tune your setup.