OpenClaw Rate Limiting Explained — How It Works and How to Stay

Rate limiting is the invisible ceiling on what your agent can do per minute, per hour, or per day. Hit the limits and your agent slows down or stops responding. Understanding where limits come from — and how to work around them — keeps your agent running smoothly at scale.

Three Independent Rate Limits

1. LLM provider rate limits: Tokens/requests you can send to Claude, GPT-4o, etc. per minute/day
2. Channel rate limits: Messages per minute your agent can send to Slack, Telegram, etc.
3. External API rate limits: Limits from third-party services (Brave Search, GitHub, Stripe, etc.)

Anthropic Rate Limits by Tier

# Approximate Anthropic rate limits:

# Tier 1 (under $40 spend):
- 5 requests/minute
- 50,000 tokens/day

# Tier 2 ($40+ spend):
- 1,000 requests/minute
- 100,000 tokens/day

# Tier 3 ($200+ spend):
- 5,000 requests/minute
- 500,000 tokens/day

# Check your tier:
https://console.anthropic.com/settings/limits

Diagnosing Rate Limit Issues

# Check gateway logs for 429 errors:
openclaw gateway status --verbose

# Common patterns that cause rate limit hits:
# 1. Too-frequent cron jobs (every 1 minute is usually overkill)
# 2. Sub-agents spawned in tight loops
# 3. Long conversations with many context tokens
# 4. Batch processing without delays between requests

Strategy 1: Reduce Cron Frequency

# Bad: runs LLM every minute
# */1 * * * *

# Good: realistic intervals
# Every 15 minutes for monitoring
# Every hour for summaries
# Daily for reports

Strategy 2: Cache Repeated Queries

# In AGENTS.md:
Before calling the LLM for a query answered in the last hour,
check memory/cache/ for a cached response.
Return cached response for identical queries within TTL window.

Strategy 3: Batch Small Tasks

# Instead of: one LLM call per lead
# Do: collect 20 leads, single LLM call with batched context

Strategy 4: Route to Smaller Models

# In openclaw.json:
{
  "llm": {
    "default": "claude-sonnet-4-6",
    "routing": {
      "classification": "claude-haiku-3",
      "summarization": "claude-haiku-3",
      "coding": "claude-opus-4-6"
    }
  }
}

Channel Rate Limits Reference

# Slack: ~50 messages/minute sustained per channel
# Telegram: 30 messages/second globally, 20/minute per chat
# WhatsApp Business: 250-1000/day on standard tier

# For bulk sends: add 1-2 second delays between messages

Rate Limit Error Recovery

# OpenClaw built-in retry behavior:
# 429 received -> wait (retry-after or 60s) -> retry
# After 3 retries -> log error -> alert agent operator

# In SOUL.md:
When a rate limit error occurs on an important task,
log to memory/YYYY-MM-DD.md and retry in the next cron window.

The OpenClaw Playbook ($9.99) has a full chapter on cost and rate limit optimization — including model routing configs, caching patterns, and how to audit token usage.