April 20, 2026
CLI is back, patience isn’t
Anthropic says OpenClaw-style Claude CLI usage is allowed again
CLI is back — but users say their trust isn’t
TLDR: Anthropic greenlit Claude’s command-line use with OpenClaw again, pitching faster toggles and cost‑saving cache features. The crowd’s split: some welcome the convenience, while many say trust is shot, demanding an official tweet and citing canceled plans and switches to competitors—proof that product clarity matters.
Anthropic says the command-line way to use Claude (the CLI) is allowed again for OpenClaw, and the tool now treats it as officially okay unless policies change. On paper it’s neat: API keys are still the most reliable for billing, there’s a simple “fast” switch that picks priority speed when available, and prompt caching can cut costs by reusing recent instructions. In plain English: it should be easier, faster, and cheaper to run Claude through OpenClaw. Docs here.
But the comments? Absolutely on fire. One top voice called Anthropic’s recent moves “more confused and shambolic than Google’s GChat era,” saying they’re getting off the ride. Another demanded an “authoritative source (aka a tweet),” roasting the rollout for feeling like a rumor mill instead of a product update. The vibe: the CLI is back, but confidence is not. Some users say they canceled subscriptions after months of flip‑flops with Claude’s coding tools, and at least one is jumping to rivals out of sheer exhaustion.
There’s also a chorus of “can we get OpenCode back?” and jokes that the new PR strategy is basically Tell HN. Meanwhile, a few pragmatists shrug and say API keys remain the safe path for production while the CLI drama plays out. Drama meter: high; patience meter: low; meme of the day: press releases by tweet.
Key Points
- •Anthropic told OpenClaw that Claude CLI reuse is allowed again, and OpenClaw now treats it as sanctioned for this integration.
- •Anthropic recommends API keys for long-lived production hosts and billing clarity; OpenClaw can reuse existing Claude CLI logins.
- •OpenClaw’s /fast toggle maps to Anthropic service tiers (auto vs standard_only) for direct api.anthropic.com requests and reports effective tiers via usage.service_tier.
- •Claude 4.6 models default to adaptive thinking in OpenClaw, with overrides per message (/think:<level>) or via model parameters.
- •OpenClaw supports Anthropic prompt caching with cacheRetention controls (none/short/long), including Bedrock pass-through for Claude and forced none for non-Anthropic Bedrock models.