Anthropic clamps down on Claude Code power users

Anthropic sent notices telling Claude Code subscribers that OpenClaw and other third-party harnesses would no longer work on Pro and Max plans, citing 'outsized strain on systems.' The thread on this was pointed: if users are paying for a token budget, why does it matter what tool consumes those tokens? Several people noted it disproportionately hits small startups and solo developers who built workflows on top of Claude Code.

This connects directly to a separate thread examining the VM infrastructure powering Claude Code and Instinct, which found both use Firecracker microVMs with memory stored in git. The underlying systems are clever but clearly not built for arbitrary third-party orchestration at scale. Anthropic is essentially forcing usage through its own interface to protect its infrastructure margins.

The counterpoint in the threads: people building 'agent-agnostic' tools that can route across multiple providers are already on the way, and this move will accelerate that. Anthropic may be buying short-term relief while pushing its most engaged users toward OpenRouter or similar abstraction layers.