Anthropic’s stance on third-party CLI tools has undergone a dramatic reversal in just three short weeks, but the situation is more complicated than it seems. On April 4 at noon, Anthropic officially announced a ban on third-party tools like OpenClaw using OAuth tokens from Claude subscription accounts to access Claude. By around April 22, OpenClaw’s official documentation quietly updated with a statement saying Anthropic employees had privately informed them that CLI usage was “allowed again.” The news was shared by independent AI engineer Dan McAteer (@daniel_mac8) on the X platform and ignited discussion, with a single tweet generating over 130 replies. However, OpenClaw founder Peter Steinberger (@steipete) subsequently weighed in personally, pointing out that “it’s theoretically allowed but practically still blocked.”
To be clear: This is not an official Anthropic document
Before getting into the details, there’s an important premise that needs to be clarified first,This so-called “unban” – Anthropic hasn’t released any official blog posts, formal documents, or announcements.The full context is: Anthropic employees privately informed OpenClaw maintainers, Claude Code product manager Boris Cherny expressed his stance on his personal Twitter account, and OpenClaw incorporated this information into their official documentation. In other words, it’s true that OpenClaw’s official documentation says CLI can be used, but Anthropic’s official documentation doesn’t explicitly state that CLI can be used.
Timeline: Three Weeks from Lockdown to Loosening
Looking back at the origin of this controversy, it goes back to April 4th. At noon PT that day, Anthropic officially announced restrictions on using third-party CLI tools via Claude subscription OAuth tokens, stating that “the product was never designed for high-intensity usage patterns by third-party tools” and that they need “to allocate compute capacity more sparingly, prioritizing service quality for official products and normal API users.” On April 10th, OpenClaw founder Peter Steinberger’s Anthropic developer account was even temporarily suspended, and was only restored later that day under public pressure.
The twist came less than three weeks later. According to OpenClaw founder Peter Steinberger himself on Hacker News and X issued a statement sayingClaude Code product manager Boris Cherny publicly stated on Twitter that “CLI-style usage is allowed,” and unusually acknowledged that the prior blocking “wasn’t intentional, most likely a false positive from an overactive abuse classifier,” adding that the policy is being clarified. Only after this unofficial statement did OpenClaw incorporate the relevant language into their own documentation.
Since this is blowing up on hacker news.
Boris said that CLI usage is allowed. Thus we added support for it, only to find out that we are still blocked there. It is trival to work around with a few renames, but I don’t wanna play that game. So it’s in a weird limbo where cli use… https://t.co/emiEmF59hf
— Peter Steinberger 🦞 (@steipete) April 21, 2026
How to write OpenClaw’s own files
According to OpenClaw Latest DocumentThe Anthropic integration page currently offers two integration paths for Claude:
- Method One: Anthropic API Key (Recommended for Production):Apply for an API key through the Anthropic Console using a pay-as-you-go pricing model. The documentation describes this as “the clearest production environment path for long-running Gateway hosts”.
- Method 2: Claude CLI Reuse (Subscription Account Transfer):Continue using existing Claude subscriptions (including Pro, Max, Team, and Enterprise plans) via CLI or
claude -pForward the command to OpenClaw.
The key passage in the original document states: “Anthropic staff told us OpenClaw-style Claude CLI usage is allowed again, so OpenClaw treats Claude CLI reuse and claude -p usage as sanctioned for this integration unless Anthropic publishes a new policy.” In plain terms: Anthropic staff privately informed us that CLI usage patterns are permitted again, so OpenClaw considers Claude CLI reuse and claude -p usage officially sanctioned — unless Anthropic releases a new policy later on. It’s worth noting that this statement itself includes the caveat “unless Anthropic publishes a new policy,” making it clear this is based on verbal communication from staff rather than written guidelines.
Steinberger also pointed out that Anthropic subsequently adjusted the classifier, getting caught in a cat-and-mouse dynamic: “To bypass it, you’d just need to change a few names, but I’m not going to play this game, so it’s currently stuck in a strange limbo: theoretically usable, practically not.” This represents Anthropic’s policy announcement that “CLI usage is allowed,” but it’s currently more of a PR statement than a setting that’s actually usable.
Three feature restrictions: not unlimited open access
Even if you believe Boris Cherny’s Twitter statement and are willing to ignore the practical gap Steinberger pointed out, the OpenClaw documentation also lists three technical limitations:
- Legacy tokens don’t support 1M context.Legacy Anthropic Token (starts with
sk-ant-oat-*Unable to enable the 1M context window and other beta features—users who want to read long documents should take note. - Prompt Caching only available with API Key:Prompt caching, a feature that can significantly reduce token costs, is only available through API keys—CLI Reuse and older token versions cannot enable it, making long conversations relatively expensive.
- Feature gap:Claude features available through the CLI backend may differ from those with pure API key authentication, as some advanced features require an API key to unlock.
What does this mean for Claude subscribers?
For Taiwanese users who have already subscribed to Claude Pro (approximately NT$620/month), Max, Team, or Enterprise, the practical implications of this policy reversal are:Nominally you can use your subscription quota to route to OpenClaw, but whether it actually runs smoothly depends on how Anthropic’s classifier is feeling that day.If you’re just using it casually and don’t mind occasional issues with system prompts getting blocked, it is worth trying now. But if you’re hoping to build production workflows with it, honestly this green light isn’t stable enough yet—it could become completely invalid at any time if Anthropic updates their classifier rules.
Additionally, please note that this official recognition is only for the OpenClaw tool.Hermes and other third-party CLI clients are outside the scope of this OpenClaw document.Those using other tools will need to test and verify on their own.
Source: KOCPC Chinese