Anthropic's Selective ToS Enforcement: Ships Apple-Violating Plugin While Sending Legal Threats to OpenCode
Roughly 1.5 weeks before publishing an official iMessage plugin that violates Apple's terms, {{Anthropic}} sent legal threats to the {{OpenCode}} project forcing removal of functionality that let users access their Claude Code subscription from a third-party harness. The selective-enforcement pattern — strict to outsiders, lenient to itself — is the structural complaint.
Anthropic's public posture on third-party developers is encouraging. Its actual enforcement pattern in early 2026 looks different. Roughly 1.5 weeks before the iMessage plugin announcement of April 2026, Anthropic's legal team forced OpenCode — an open-source coding agent harness — to remove a feature that let paying Claude Max subscribers use their subscription quota inside OpenCode rather than inside Claude Code itself. The basis cited is Anthropic's Claude Code legal compliance documentation, which states that OAuth tokens issued through Claude Code, Claude.ai free, Pro, and Max accounts 'are intended exclusively for cloud code and cloud.ai. Using OAuth tokens that are obtained through those methods and through those accounts in any other product, tool or service, INCLUDING THE AGENT SDK, is not permitted and constitutes a violation of the consumer terms of service.' The asymmetry critics highlight: Anthropic enforces its own vague terms aggressively against third parties (legal threats, account bans, removal demands) while simultaneously officially distributing a plugin that violates Apple's explicit terms. Developer Matt PCO reportedly spent almost a month attempting to get Anthropic to clarify whether his open-source wrapper was permitted; the responses were apologetic but contained no actionable rule. One working hypothesis — articulated by critics including Theo Browne — is that the vagueness is the point. Deliberately ambiguous terms plus selective enforcement let Anthropic appear pro-ecosystem in public statements, eliminate competitive harnesses individually as they become threatening, and preserve the moat around its subscription product without ever publishing a rule that could be characterized as anti-competitive. The contrast with OpenAI, GitHub Copilot, and similar subscription products — which generally allow cross-tool use of paid quota — is the comparative evidence offered.