OpenClaw Configuration Deep Dive

OpenClaw is the backbone of my AI agent setup. After months of tweaking, here's what I've learned about getting the most out of it.

Architecture Overview

OpenClaw runs as a Gateway daemon on your server, providing:

• Multi-provider LLM routing
• Tool/ skill system for agent actions
• Session management
• Various channel integrations (Telegram, Discord, etc.)

Core Configuration

The main config lives at ~/.openclaw/config.yml:

gateway:
  bind: "0.0.0.0:18789"
  auth:
    enabled: true
    tokens:
      - name: "main"
        token: "sk-your-token-here"
        
model:
  providers:
    - name: "lmstudio"
      type: "openai"
      endpoint: "http://100.107.153.118:1234/v1"
      model: "unsloth/qwen3.5-9b@q6_k"
      priority: 1
      
    - name: "openrouter"
      type: "openrouter"
      api_key: "sk-or-your-key"
      default_model: "anthropic/claude-sonnet-4-6"
      priority: 2

Model Routing

The priority system is clever:

1. OpenClaw tries providers in priority order
2. If primary fails (rate limit, timeout), it falls back
3. You can also specify model per-conversation

I've set up my local LM Studio as primary for routine tasks, with OpenRouter as fallback for complex reasoning.

Skills System

Skills extend what agents can do. They're defined in ~/.openclaw/skills/:

skills/
├── calculator/     # Math operations
├── web-search/     # Brave Search API
├── code-interpreter/ # Run code
└── custom/        # Your own skills
    ├── SKILL.md
    └── scripts/

Each skill has a SKILL.md that describes:

• When to use
• Input/output format
• Implementation details

Session Management

Sessions are persistent conversation contexts:

session:
  storage: "sqlite"  # or "redis" for clustering
  db_path: "~/.openclaw/sessions.db"
  max_history: 50    # messages per session
  ttl: 86400         # 24 hours

For long-running agents, I recommend external session storage.

Environment Variables

Don't hardcode secrets! Use env vars:

model:
  providers:
    - name: "openrouter"
      api_key: "$ENV_OPENROUTER_API_KEY"  # reads from env

Troubleshooting

Common issues and fixes:

"No available model" error

# Check LM Studio is running
curl http://100.107.153.118:1234/v1/models

# Check OpenClaw logs
journalctl -u openclaw -f

High latency

• Switch to local models for simple tasks
• Reduce context window
• Enable response streaming

Rate limits

• Implement exponential backoff
• Queue requests
• Use multiple API keys rotation

Production Checklist

Before going to production:

• Enable authentication
• Set up monitoring/logging
• Configure rate limiting
• Set up backup for sessions
• Test failover between providers
• Document your configuration

Conclusion

OpenClaw is incredibly flexible. The config above is just a starting point — read the docs, experiment, and find what works for your use case.