Anthropic Reinstates Sanction for OpenClaw-Style Claude CLI Reuse

OpenClaw’s documentation confirms that Anthropic has given the green light — again — to reusing the bundled Claude CLI (claude -p) as a backend for third-party integrations. After earlier ambiguity around whether wrapping the CLI violated terms, Anthropic staff have communicated that this usage pattern is permitted, and OpenClaw is treating the guidance as settled policy absent a future reversal. Direct API keys remain the recommended path for production gateway hosts that need explicit server-side billing control.

The provider page also documents the operational surface OpenClaw exposes around Anthropic: adaptive thinking defaults for Claude 4.6, a /fast toggle that maps to service_tier: auto or standard_only on direct api.anthropic.com traffic, and prompt caching via a cacheRetention parameter that defaults to the 5-minute short cache when API-key auth is used. Caching is API-only — legacy token auth is explicitly excluded — and the same cacheRetention pass-through works for Claude models hosted on Amazon Bedrock.

The 1M-token context window remains beta-gated behind params.context1m: true and the context-1m-2025-08-07 beta header, and Anthropic currently rejects those requests on legacy sk-ant-oat-* OAuth tokens — OpenClaw falls back to the standard context and logs a warning in that case. Troubleshooting guidance covers 401s from expired token auth, per-agent auth inheritance gaps, and model-scoped rate-limit cooldowns that leave sibling models usable.