OpenClaw Founder Responds to Claude CLI "Unban" Rumors: Verbal Permission Granted But Still Permanently Banned

According to Dynamic Insight monitoring, someone took a screenshot of the sentence "Anthropic allows OpenClaw-style Claude CLI usage" from the latest OpenClaw document and posted it on Hacker News. The post reached the front page and sparked a discussion in the X community. OpenClaw founder Peter Steinberger responded on both platforms, providing the clearest clarification to date: verbal permission is indeed granted, but practical blocking is also in effect.

As per Steinberger, Claude Code manager Boris Cherny had previously publicly stated on X that CLI usage was allowed. Based on this, OpenClaw reintroduced support and by default disabled high-token-consumption features like heartbeat. However, during team testing, the Anthropic server continued to block the system prompts from OpenClaw, and the calls were still rejected. He added that a few simple renames could bypass the classifier, but he did not wish to engage in this cat-and-mouse game.

OpenClaw interacts with Claude Code by invoking `claude -p`, which is the official Claude CLI in a one-time non-interactive mode, with no separate channel to bypass it. Whether it uses an API key or a Pro/Max/Team subscription quota depends on which account the local Claude CLI was logged into at the time. The current "ban on OpenClaw" by Anthropic is not a prohibition of the `claude -p` command; instead, it is the server-side fingerprint detection of OpenClaw's system prompts, with rejections upon detection. Boris's verbal "CLI allowance" and the document's "`claude -p` reuse by default" refer to the same thing, with only a classifier decision standing between the ability to run OpenClaw based on subscription quotas.

In version 2026.4.5, OpenClaw removed the Claude CLI from the default onboarding of new users, only to add it back two days later in version 2026.4.7, along with a set of environment variable cleanups and permission restrictions for the Claude CLI runtime. The "re-permissioned" in this round of documentation corresponds to this change in version 2026.4.7.