OpenClaw Founder Responds to Claude CLI 'Unblocking' Rumors: Verbally Allowed but Still Blocked in Practice

According to monitoring by Beating, someone shared a screenshot from OpenClaw’s latest documentation stating that ‘Anthropic allows the use of Claude CLI in OpenClaw style’ on Hacker News, which quickly gained traction and was also circulated on X. OpenClaw founder Peter Steinberger responded on both platforms, providing the clearest version of the situation: verbal permission is true, but the actual blocking is also true. According to Steinberger, Boris Cherny, the head of Claude Code, previously stated on X that CLI usage was permitted, prompting OpenClaw to re-enable support and default to disabling high token consumption features like heartbeat. However, the team found that the Anthropic server still blocks OpenClaw’s system prompt snippets, resulting in calls being denied. He added that a few renaming tweaks could bypass the classifier, but he does not want to engage in such a cat-and-mouse game. OpenClaw’s integration with Claude Code fundamentally involves calling claude -p, which is a one-time non-interactive mode of the official Claude CLI, without an independent bypass; whether it uses an API key or Pro/Max/Team subscription depends on which account the local Claude CLI is logged into at the time. Anthropic’s current ‘blocking of OpenClaw’ does not prohibit the claude -p command itself, but rather identifies OpenClaw through system prompt fingerprinting on the server side, leading to rejections upon detection. Boris’s verbal ‘CLI allowed’ and the documentation’s ‘reuse of claude -p is tacitly permitted’ refer to the same issue, with the only difference being whether the classifier allows OpenClaw based on subscription limits. OpenClaw removed Claude CLI from the default onboarding for new users in version 2026.4.5, but added it back two days later in version 2026.4.7, while also implementing a series of variable cleanups and permission restrictions for the Claude CLI runtime environment. The ‘re-allowed’ status in this round of documentation corresponds to the changes made in version 2026.4.7.