If you're running OpenClaw agents, you're probably paying more than you realize.
The problem: OpenClaw agents run continuously. Heartbeat timers fire every
few minutes. Cron jobs run overnight. Subagents spawn subagents. Every one
of those makes LLM API calls — and without visibility, you won't know where
the money is going until the bill arrives.
## The fix: install CostClaw
CostClaw is a free OpenClaw plugin that captures every LLM call and shows
you a live local dashboard.
Install:
\bash
git clone https://github.com/Aperturesurvivor/costclaw-telemetry.git
cd costclaw-telemetry
npm install && npm run build
openclaw plugins install -l .
openclaw gateway restart
\\
Open http://localhost:3333.
## What to look for
1. Usage by Source
This is always the most revealing. Most people expect "user" to be their
biggest cost. It's usually "heartbeat."
If your heartbeat agent is running GPT-4o or Claude Sonnet for keep-alive
checks, switch it to GPT-4o-mini or Claude Haiku. Same functionality,
~1/20th the cost.
2. Model Breakdown
Sort by cost. If one model owns 80%+ of your spend, ask whether every use
case actually needs that model's capability.
3. Sessions by Cost
Identify your most expensive conversation patterns. Long agentic loops
with many subagent calls show up here.
4. Recommendations panel
CostClaw auto-generates suggestions based on your specific usage.
Common output: "Your most-used model costs $X/1K tokens. Switching
to [cheaper model] for similar tasks would save ~$Y/month."
## Real numbers
In my own setup: $90/mo → $35/mo after seeing the breakdown.
The entire savings came from one heartbeat agent.
GitHub: https://github.com/Aperturesurvivor/costclaw-telemetry (MIT, free)
#openclaw #llm #ai #agents #programming
Top comments (0)