Revoking a Claude Code OAuth token

Research into how to revoke a long-lived CLAUDE_CODE_OAUTH_TOKEN (the kind claude setup-token mints), prompted by needing to rotate a token baked into a bot-bottle container.

Summary

There is a documented revoke button, but it does not currently provide reliable immediate invalidation. As of early 2026, an open bug shows tokens remaining valid for 3–4 days after explicit revocation, including after "Log out all sessions". Anthropic has not commented on the bug or on the related feature-request issue. For a known-leaked token there is no guaranteed way to make it stop working today; rotate immediately and hope server-side enforcement catches up.

What works (with caveats)

claude.ai/settings/claude-code lists active Claude Code instances with a per-entry "Revoke" control. This is the only path mentioned in any GitHub issue thread; it is not mentioned in the official auth docs, which document claude setup-token but say nothing about revocation.

A second lever is claude.ai → Settings → Account → Active Sessions → "Log out all sessions".

Empirically, neither reliably propagates. Issue #43801 documents a reproducible failure: the reporter shut down their VM completely while offline, performed both revocation actions via claude.ai, waited 3–4 days, then booted the VM cold — Claude Code authenticated without re-login. Because the VM was offline during the revocation window, this isolates the failure to the server side. The issue is open with zero Anthropic staff responses.

Issue #34198 (filed March 2026) requests proper server-side revocation on claude logout and devcontainer shutdown. Also open, also no Anthropic response.

What does not work

claude logout (or claude /logout) only clears local credentials. It makes no server-side revocation call. Do not rely on it.

It is also not confirmed whether changing the Anthropic account password or revoking the broader account session invalidates Claude Code OAuth tokens; no issue thread tested this directly.

Unconfirmed

Whether setup-token-generated long-lived tokens appear on claude.ai/settings/claude-code as a distinct entry type vs. interactive sessions. Plausible but not confirmed.
Whether the 3–4 day server-enforcement gap is consistent or has been silently reduced since the issue was filed. No release notes mention it.
Whether contacting Anthropic support for a leaked token gets faster server-side action than the self-service UI.

Practical rotation procedure

For a known-leaked or suspected-leaked token:

Revoke the entry at claude.ai/settings/claude-code.
Run "Log out all sessions" under Settings → Account → Active Sessions.
Run claude setup-token to mint a replacement, and rotate it into BOT_BOTTLE_CLAUDE_OAUTH_TOKEN immediately.
Email Anthropic support at support.anthropic.com. Security issues sometimes get attention that GitHub issues do not.

The new token is in effect right away; the old token may continue to authenticate for up to several days. There is no client-side mitigation for that — the server is the only authority.

3.7 KiB Raw Permalink Blame History Unescape Escape