Your assistant needs a language model behind it. OpenClaw supports any provider that exposes an OpenAI-compatible chat completions endpoint. So which do you pick?
Here are your most common cloud choices:
Anthropic (Claude). Strong at following instructions and refusing unsafe requests. Set your provider to anthropic and supply a Claude API key.
OpenAI (GPT). Broad ecosystem, function-calling support. Use provider openai with your OpenAI API key.
DeepSeek. If you want cost-effective high-volume use, this is your pick. It works through OpenRouter or a direct DeepSeek API key.
You configure the provider and model in ~/.openclaw/model.json. I'll explain the exact format in the Model Refs unit.