OpenClaw documents a first-class Anthropic integration: you can run against the Anthropic API (usage-based billing, predictable for long-lived gateways) or reuse Claude CLI auth on the host when you already log in that way. Legacy Anthropic token profiles still work if they are already configured.
The page is worth bookmarking for operational knobs: Claude 4.6 defaults to adaptive thinking unless you override it; /fast maps to Anthropic service_tier for direct api.anthropic.com traffic; prompt caching uses cacheRetention (none / short / long) with sensible API-key defaults; and the 1M context beta is gated behind params.context1m: true (with caveats for legacy OAuth-style tokens). Bedrock-hosted Claude gets partial pass-through for cache settings.
For day-to-day debugging, the doc points at openclaw onboard, openclaw models status, and the broader CLI backends, OAuth, and troubleshooting sections-useful when “which credential does this agent actually use?” is the question.
Source: docs.openclaw.ai