Claude CLI Misunderstanding: OpenClaw's Status Remains Unchanged

According to monitoring by Dongcha Beating, discussions among OpenClaw users about whether ‘Anthropic has allowed Claude CLI again’ have clarified the facts: Anthropic has never relaxed its policy; the only change is the misunderstanding by OpenClaw founder Peter Steinberger regarding a statement from Boris Cherny, the head of Claude Code. This misunderstanding has been denied twice by the behavior of Anthropic’s server-side classifier. Boris’s reply on April 6 was posted under Steinberger’s experimental tweet. The experiment itself was simple: a request was made using the official Claude CLI with the -p parameter, and by including the phrase ‘running inside OpenClaw’ in the system prompt, it was classified as a third-party application, resulting in charges from Extra Usage. Boris acknowledged that this was an overreaction by the classifier and promised to improve the usage terms for -p. This statement was aimed at ensuring that ‘developers using the official CLI for scripting should not be misclassified,’ not that tools like OpenClaw would be exempt from classification. However, Peter directly restored Claude CLI as the default backend for new users in version 2026.4.7, stating in the documentation that ‘OpenClaw considers the reuse of claude -p as implicitly permitted,’ and he even disabled high-consumption features like heartbeat to demonstrate compliance. This step extended his interpretation of ‘CLI use is allowed’ to mean ‘OpenClaw fundamentally operates as claude -p, so it counts as CLI usage.’ Anthropic has never publicly recognized this extension. The server-side classifier continues to reject requests based on OpenClaw’s characteristics (currently confirmed mainly by system prompt fingerprints) because the prompt injected by OpenClaw is seen as the strongest signal of ‘third-party tools driving the CLI.’ While the official CLI is being called, the actual entity initiating the request remains OpenClaw, thus it is still charged as a third party. This is what Steinberger acknowledged today as ‘theoretically should be usable, but practically is not.’ In other words, he thought the permission he received would cover OpenClaw, but it has proven not to be the case. The so-called ‘gray area’ is not due to Anthropic’s ambiguous stance, but rather a disagreement obscured by vague wording: Boris aimed to fix the boundary of the classifier’s misclassification, while Peter interpreted it as ‘OpenClaw being within the boundary.’