OpenClaw: 7 Common Errors and How to Fix Them

1 Introduction to OpenClaw Errors

OpenClaw (formerly Moltbot and ClawdBot) is an open-source AI assistant with over 417,000 GitHub stars. While the official troubleshooting documentation covers basic issues, real-world deployments often encounter more complex problems, especially when running multi-agent configurations with SearXNG, Docker, and external APIs.

This guide documents seven common errors encountered during OpenClaw setup and operation, along with step-by-step solutions verified in production environments. Each error includes the exact error message, root cause analysis, and tested fix.

Error Categories Overview

Understanding OpenClaw File Structure

Before diving into specific errors, it's essential to understand OpenClaw's file structure. Knowing where to apply fixes is as important as knowing what to fix.

Core Configuration Files

Agent Configuration Files (Multi-Agent Setup)

If you're running multiple AI agents, each agent has its own configuration directory with these key files:

How Files Connect: Agent Request Flow

When a user sends a message to an OpenClaw agent, here's how the configuration files work together:

Quick Reference: Where to Apply Common Fixes

2 Brave Search API Rate Limits

OpenClaw's built-in web search uses the Brave Search API by default. While convenient, the free tier has strict limitations that can quickly block research-intensive agents.

Brave Free Tier Limits

SearXNG provides unlimited, self-hosted search by aggregating results from multiple engines including Google, DuckDuckGo, Startpage, and Brave itself.

Step 1: Disable Brave in openclaw.json

"tools": {
  "web": {
    "search": {
      "enabled": true,
      "provider": "brave"
    }
  }
}

"tools": {
  "web": {
    "search": {
      "enabled": false,
      "provider": "brave"
    }
  }
}

Step 2: Add SearXNG Commands to Agent HEARTBEAT Files

Update each agent's HEARTBEAT.md with explicit SearXNG usage:

# USE SearXNG for all web searches (unlimited, self-hosted):
python3 ~/.openclaw/workspace/skills/local-websearch/searxng_search.py "query" --count 10

# Example for research agent:
python3 ~/.openclaw/workspace/skills/local-websearch/searxng_search.py 
  "React best practices 2026" --count 10

# Example for contact research:
python3 ~/.openclaw/workspace/skills/local-websearch/searxng_search.py 
  "John Doe software architect profile" --count 10

Comparison: Brave vs SearXNG

Step 3: Restart Gateway

# Restart OpenClaw gateway to apply changes
launchctl kickstart -k gui/$(id -u)/ai.openclaw.gateway

# Verify SearXNG is working
curl -s "http://localhost:8888/search?q=test&format=json" | head -c 100

3 SearXNG Docker/Colima Issues

Note: OpenClaw itself does not require Docker. This section only applies if you're running SearXNG (self-hosted search engine) as an alternative to Brave Search API. On macOS, SearXNG typically runs in Docker using Colima as the container runtime. According to Colima documentation, it provides a lightweight, free alternative to Docker Desktop. However, Colima doesn't automatically restart after Mac sleep or reboot, leaving SearXNG unavailable.

Root Cause Analysis

This error occurs when:

• Colima VM stopped: Mac woke from sleep or rebooted
• Health check doesn't recover: Script only checks Docker status, doesn't attempt restart
• SearXNG left down: Requires manual intervention

Update your SearXNG health check script to automatically start Colima when Docker is unavailable:

#!/bin/bash
# searxng-healthcheck.sh (v4)

start_colima() {
    log "Colima/Docker not running. Attempting to start Colima..."

    # Check if colima is installed
    if ! command -v colima &>/dev/null; then
        log "ERROR: Colima is not installed"
        return 1
    fi

    # Check if already running
    if colima status &>/dev/null; then
        log "Colima is already running"
        return 0
    fi

    # Start Colima with recommended settings
    log "Starting Colima with cpu=2, memory=4..."
    if colima start --cpu 2 --memory 4 >> "$LOG_FILE" 2>&1; then
        log "Colima started successfully"

        # Wait for Docker socket to be available
        local waited=0
        while [[ $waited -lt 30 ]]; do
            if docker info &>/dev/null; then
                log "Docker is now accessible"
                return 0
            fi
            sleep 2
            waited=$((waited + 2))
        done
        log "WARNING: Colima started but Docker not accessible after 30s"
        return 1
    else
        log "ERROR: Failed to start Colima"
        return 1
    fi
}

ensure_docker_running() {
    if docker info &>/dev/null; then
        return 0
    fi

    log "Docker is not running, attempting to start Colima..."
    start_colima
}

Startup Chain After Fix

Testing the Fix

# Stop Colima to simulate the problem
colima stop

# Trigger health check manually
~/.openclaw/scripts/searxng-healthcheck.sh

# Check the log - should show auto-recovery
tail -20 ~/.openclaw/logs/searxng-health.log

4 LLM Command Hallucinations

One of the most frustrating issues with AI agents is command hallucination - when the LLM invents commands, flags, or API endpoints that don't exist. This happens because LLMs pattern-match to common conventions (like RESTful APIs) instead of following your documented CLI commands.

Why LLMs Hallucinate Commands

LLMs are trained on vast amounts of code and documentation. When they encounter a CLI tool, they often:

• Apply RESTful conventions: get-resource, list-items, fetch-data are common patterns
• Invent familiar flags: --search, --name, --filter seem logical even when unsupported
• Skip required arguments: Positional arguments get dropped when the model focuses on flags
• Confuse similar tools: Commands from one CLI tool bleed into another

General Signs of Command Hallucination

The fix is to provide explicit command documentation that includes both valid commands AND common invalid patterns. Add this section to your agent's HEARTBEAT.md file:

## CRITICAL: Valid Commands Reference

**ONLY use these exact commands. Do NOT invent or guess commands.**

### VALID COMMANDS:
  list-items --type [type] --status [status]
  create-item "[name]" --type [type]
  update-item "[id]" --field [value]
  delete-item "[id]"

### INVALID COMMANDS (NEVER USE):
  get-item [X]       (use: list-items --id [id])
  fetch-data [X]     (use: list-items)
  --search [X]       (not supported)
  --name [X]         (use: --title instead)

### COMMAND STRUCTURE:
  [subcommand] [positional-args] [--flags]
  [OK] create-item "My Item" --type project
  [X]  --type project create-item "My Item"

Reference Example: Notion Helper Script

Here's a real-world example from a Notion integration where agents hallucinated REST-style commands:

# Missing subcommand (common mistake):
python3 notion-helper.py --organization "Acme" 
  --role "Manager" --email "user@example.com"

# Subcommand FIRST, then positional, then flags:
python3 notion-helper.py add-contact "John Smith" 
  --organization "Acme" --role "Manager" 
  --email "user@example.com"

Prevention Checklist for Any CLI Tool

Add this verification pattern to your agent's workflow:

5 Agent Workflow Deadlocks

Even with correct commands, multi-agent systems can experience workflow deadlocks where tasks get stuck and production output flatlines despite agents reporting "healthy" status.

Symptoms of Workflow Deadlock

• Tasks stuck in inbox: 10+ tasks with no progress
• Agents report "HEARTBEAT_OK": But produce nothing
• Pipeline appears healthy: Coordinator shows "2 in_progress" but both are waiting
• Activity log full of heartbeats: But no actual work entries

Root Cause: Missing Status Query

Agents typically query only --status assigned and --status in_progress, missing their inbox tasks entirely:

Fix 1: Check Inbox FIRST

Update each agent's HEARTBEAT.md to query inbox before other statuses:

### Step 1: Check Your Tasks

# Check inbox FIRST - these are tasks waiting to be started
python3 ~/.openclaw/scripts/notion-helper.py query-tasks --assignee ResearchBot --status inbox

# Then check assigned and in_progress
python3 ~/.openclaw/scripts/notion-helper.py query-tasks --assignee ResearchBot --status assigned
python3 ~/.openclaw/scripts/notion-helper.py query-tasks --assignee ResearchBot --status in_progress

Fix 2: Mandatory Status Updates

Add explicit status update requirements:

### Step 2: If Tasks Found - START WORKING

**CRITICAL: Before doing ANY work, update task status to in_progress:**

python3 ~/.openclaw/scripts/notion-helper.py update-task "[task-id]" --status in_progress

# ... do the actual work ...

# When COMPLETE, update status:
python3 ~/.openclaw/scripts/notion-helper.py update-task "[task-id]" --status done

Fix 3: Coordinator Task Activation (Step 0)

Add a "Step 0" to the coordinator agent (like Coordinator) to activate idle tasks:

### Step 0: ACTIVATE IDLE AGENT TASKS (DO THIS FIRST!)

**Check for agent tasks stuck in inbox:**

python3 ~/.openclaw/scripts/notion-helper.py query-tasks --assignee ResearchBot --status inbox
python3 ~/.openclaw/scripts/notion-helper.py query-tasks --assignee WriterBot --status inbox
python3 ~/.openclaw/scripts/notion-helper.py query-tasks --assignee OutreachBot --status inbox

**For EACH agent task in inbox, ACTIVATE IT:**

python3 ~/.openclaw/scripts/notion-helper.py update-task "[task-id]" --status in_progress

**This ensures agents pick up their work on next heartbeat!**

Fix 4: Realistic Pipeline Health Check

Update health check to count only ACTIVELY working tasks:

# Just counts tasks
in_progress_count = tasks in progress

# Reports "healthy" even if all
# tasks are waiting on external factors

# Evaluates task quality
For EACH in_progress task:
  - ACTIVELY WORKING? → Count
  - WAITING? → Don't count

IF active_work_count < 2:
  → ACTIVATE INBOX TASKS

Fix 5: Proactive Work When Idle

Add proactive work sections for when agents have no assigned tasks:

### Step 3: If NO Tasks Found - PROACTIVE WORK

**Do NOT just reply HEARTBEAT_OK. Find new opportunities!**

# For research agents:
Search for trending topics in your domain
Monitor competitor activities and industry news
Find new data sources or APIs to integrate

# For writing agents:
Check content calendar for upcoming deadlines
Draft outlines for planned blog posts or docs

# For outreach agents:
Research potential partners or collaborators
Find contact information for key stakeholders

Expected Behavior After Fixes

6 Environment Variable Issues

Skills and tools often require environment variables that may not be passed correctly to agent subprocesses, even when configured in openclaw.json.

Root Cause: LaunchAgent Doesn't Pass Env Vars

Even with correct configuration in openclaw.json, the LaunchAgent may not inject environment variables into agent subprocesses:

// In openclaw.json - correctly configured but not working:
"searxng": {
  "enabled": true,
  "env": {
    "SEARXNG_URL": "http://localhost:8888"
  }
}

Step 1: Edit LaunchAgent plist

Edit ~/Library/LaunchAgents/ai.openclaw.gateway.plist and add the environment variable:

<key>EnvironmentVariables</key>
<dict>
    <key>SEARXNG_URL</key>
    <string>http://localhost:8888</string>
</dict>

Step 2: Full LaunchAgent Reload (Not Kickstart)

# WRONG - doesn't reload env vars:
launchctl kickstart -k gui/$(id -u)/ai.openclaw.gateway

# CORRECT - full reload:
launchctl bootout gui/$(id -u)/ai.openclaw.gateway
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/ai.openclaw.gateway.plist

Step 3: Verify Environment

# Check if env var is now present:
launchctl print gui/$(id -u)/ai.openclaw.gateway | grep SEARXNG

# Expected output:
# SEARXNG_URL => http://localhost:8888

Another Common Issue: save-document Missing --content

Fix by adding explicit wrong/right examples to the writing agent's HEARTBEAT.md:

# Missing --content argument:
python3 notion-helper.py save-document "Title" 
  --type report_draft --author WriterBot

# --content is REQUIRED:
python3 notion-helper.py save-document "Title" 
  --type report_draft 
  --content "Full document content here" 
  --author WriterBot

7 AI-Powered Error Monitoring with ErrorMonitor

Traditional error monitoring using hardcoded patterns often misses novel error types. For example, LLM command hallucinations weren't caught because the pattern "exec failed.*not found" doesn't match argparse errors like "invalid choice".

Why Pattern-Based Monitoring Fails

# Traditional CRITICAL_PATTERNS miss argparse errors:
CRITICAL_PATTERNS=(
    "Config invalid"
    "exec failed.*not found"     # ← Requires "not found"
    "No such file or directory"
    "ECONNREFUSED"
    "Error.*API"                 # ← It's "error:" lowercase!
)

# Actual error format - NOT MATCHED:
notion-helper.py: error: argument command: invalid choice: 'get-page'

Instead of bash scripts with hardcoded patterns, create a dedicated OpenClaw agent (ErrorMonitor) that uses Claude's semantic understanding to detect and diagnose errors.

ErrorMonitor Agent Architecture

ErrorMonitor's Monitoring Workflow

Benefits of ErrorMonitor

• Uses Claude Max subscription: No additional API costs
• Semantic understanding: Catches novel error types
• Intelligent diagnosis: Provides actionable fix suggestions
• Deduplication: Same error won't notify for 6 hours
• 8-hourly progress reports: Agent health summaries

Error Categories Detected

Verification Commands

# Verify ErrorMonitor agent registered
openclaw agents list
# Should show your configured agents including error-monitor

# Check cron jobs
openclaw cron list
# Should show: error-monitor-system, error-monitor-8hr-report

# Force run ErrorMonitor (test)
openclaw agent error-monitor --message "Check HEARTBEAT.md and monitor system health now."

# View ErrorMonitor logs
tail -50 ~/.openclaw/logs/error-monitor.log

8 Frequently Asked Questions

9 Abbreviations & Glossary

Related Reading

• OpenClaw Complete Guide: Installation, WhatsApp Setup, and Best Practices
• Notion MCP Server: Complete Integration Guide for Claude