OpenClaw High Token Usage — How to Diagnose and Fix It

High token usage in OpenClaw usually means unexpected LLM costs or rate limit hits. Here's how to track down the culprit and fix it without crippling your agent's capabilities.

Diagnose First

# Check recent gateway activity:
openclaw gateway logs --tail 200 | grep -i "tokens\|usage\|context"

# See which crons are running most frequently:
openclaw cron list
# Look for crons running every minute or every 5 minutes

# Check your LLM provider's dashboard:
# Anthropic: console.anthropic.com → API Usage
# OpenAI: platform.openai.com → Usage
# Google: console.cloud.google.com → APIs & Services

Common Causes and Fixes

Cause 1: Bloated Workspace Files

Every session loads SOUL.md, TOOLS.md, AGENTS.md, and USER.md. If these files have grown large (code snippets, long tables, historical notes), they inflate every request's token count.

# Check your workspace file sizes:
wc -w ~/.openclaw/workspace/SOUL.md
wc -w ~/.openclaw/workspace/TOOLS.md
wc -w ~/.openclaw/workspace/AGENTS.md

# Targets:
# SOUL.md: under 500 words
# TOOLS.md: under 1,000 words of active content
# AGENTS.md: under 800 words

Move historical content out of active workspace files into archive files that are only loaded on demand.

Cause 2: Overly Frequent Crons

A cron running every minute makes 1,440 LLM calls per day. Most monitoring tasks don't need that frequency.

# Find high-frequency crons:
openclaw cron list | grep -E "\*/[1-9] \*"

# Update to less frequent schedule:
openclaw cron update hex-monitor-check --schedule "*/15 * * * *"
# Every 15 min = 96 calls/day instead of 1,440

Cause 3: Wrong Model for the Task

Using a powerful (expensive) model for simple tasks is a major source of unnecessary cost.

# Use cheaper models for simple crons:
openclaw cron update hex-uptime-check \
  --model "claude-haiku-3-5"
# ~20x cheaper than Sonnet for simple is-site-up checks

openclaw cron update hex-digest \
  --model "google/gemini-2.0-flash"
# Very affordable for summarization tasks

Cause 4: Overly Broad Task Prompts

Tasks that say "check everything" or "analyze the entire codebase" generate massive context requirements. Be specific:

# Instead of:
"Analyze our entire Slack history and summarize trends"

# Use:
"Check the last 20 messages in #sales from today.
Summarize any action items. Keep response under 100 words."

Cause 5: Missing Response Limits

Add explicit output constraints to cron tasks to prevent the LLM from generating long responses when short ones are sufficient:

"[your task]. Respond in under 150 words. No markdown formatting."

Set Up Usage Monitoring

openclaw cron add \
  --name "hex-token-usage-alert" \
  --schedule "0 8 * * 1" \
  --agent main \
  --model "google/gemini-2.0-flash" \
  --task "Check LLM API usage this week from the provider dashboard. If total cost exceeds $50 or is tracking above last week by more than 50%, post a cost alert to Slack DM with breakdown."

Want the full OpenClaw setup guide? The OpenClaw Playbook covers everything — $9.99.