#OpenClaw Model Configuration

OpenClaw is model-agnostic and supports auth profile rotation and fallback chains. This chapter explains how to connect and configure models.

#Supported Providers

Models are configured in openclaw.json, with support for auth profile rotation and fallback chains — when the primary model is unavailable, it automatically switches to a backup.

#Context Window & Local Model Requirements

For reliable operation, the project recommends:

• Context window ≥ 64K tokens
• For local models: 32B+ parameters and 24GB VRAM minimum

The agent must fit memory, tool descriptions, skills, and conversation history into context — too small a window leads to unstable behavior.

#Configuring in openclaw.json

Model configuration lives in the main config file openclaw.json (under ~/.openclaw/). Key areas include:

• Provider settings and auth
• Model routing preferences: primary model, fallback chain
• Tool policies and approval gates
• Autonomy level settings

An illustrative snippet (exact fields defer to your version):

{
  // Model routing: prefer Claude, fall back to GPT on failure
  "model": {
    "primary": "anthropic:claude-...",
    "fallback": ["openai:gpt-...", "ollama:llama3.1:70b"]
  }
}

#Credentials & Environment Variables

Keep secrets like API keys in environment variables or a separate config, out of files that might be shared/committed:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
GOOGLE_API_KEY=...

Relevant paths can be overridden via environment variables:

Security tip: never commit keys to Git; keep ~/.openclaw/ private.

#Three Typical Setups

#1) Cloud frontier models (most reliable)

Use strong Anthropic / OpenAI / Google models directly — best tool calling and long-task performance, ideal for stable everyday use.

#2) Local models (most private)

Run open models with Ollama or LM Studio so data never leaves your network. Meet the 32B+ / 24GB VRAM recommendation, or complex tasks may be unstable.

#3) Hybrid + fallback chain (cost and reliability)

Use a cheap/local model for simple tasks and fall back to a strong cloud model when complex or failing:

simple tasks → local Ollama (cheap)
failure/complex → fall back to Claude / GPT (quality)

#Cost Notes

Rough monthly API spend by usage intensity:

Strongly recommended: set spending alerts at the provider level. Shorter heartbeat intervals and pricier models grow the bill faster — see CLI & Automation.

#Next Steps

• Tool System — let the model actually take action
• CLI & Automation — control Heartbeat frequency and cost
• Permissions & Security — credential and port security