Anthropic Blocked OpenClaw: Claude Max Workarounds That Still Work

Last week, Anthropic dropped a bomb. With almost no warning, they cut off all OpenClaw agent access through Claude Pro and Max subscriptions. If you were relying on your Claude sub to power an OpenClaw agent, that door is closed. But Marcelo did not rebuild from scratch. He engineered around it.

The Architecture That Saved Him

The key point in the video is not just that Anthropic changed the rules. It is that Marcelo had already separated his workflow logic from the provider access path. Long before OpenClaw, he was building repeatable /command workflows inside Claude CLI. When those projects moved into ClearMud OS, the backend logic stayed the same.

That decision is what made the workaround possible. He did not need to reinvent the workflow. He only needed to change how the system called it.

Workaround 1: Claude CLI + REST API Wrapper

This is the first and most important workaround. Keep the Claude CLI workflow intact, wrap it in a local REST API, and have the OpenClaw agent call that local endpoint instead of trying to use Claude directly through the old subscription path.

Marcelo shows this with his YouTube script generator. The front end is just a control surface. When he clicks generate, the API calls Claude CLI in the background. Opus handles the deep research and title-thumbnail-hook strategy. Then Sonnet writes the full script on the next turn.

The important takeaway is that the core workflow did not change. The wrapper changed. That is why the system survived the policy shift.

Workaround 2: Build an IDE-Like Terminal Into Your OS

The second workaround is for builders who already have a custom dashboard or operating system around their agent stack. Instead of pretending the terminal is unnecessary, Marcelo brings it directly into the interface.

That creates two paths. You can do the work yourself from the embedded terminal, or you can tell your agent to open the dashboard, invoke Claude in that terminal, and perform the work on your behalf. In both cases, Claude Code still runs through the local CLI path you control.

He demonstrates this with Muddy OS, using Claude to make UI changes from inside the embedded terminal. No extra SSH dance. No leaving the product context.

Workaround 3: Switch the Model for the Agent Layer

The third workaround is simpler. If Claude can no longer power the agent personality layer directly, move that layer to another model. Marcelo points to GPT as the practical choice for the agent shell while keeping Claude Max for the actual CLI-based execution workflows.

That separation matters. The personality layer and the work execution layer do not need to be the same thing. Once you accept that, the system becomes far more resilient.

Key Takeaways

Build CLI-first workflows, then wrap them in an API when needed. Decouple the agent layer from the model access layer. Use different models for different jobs. And if you live inside a custom OS, do not hide the terminal. Bring it into the product honestly.

The deeper lesson is architectural. If your workflow only works with one provider path, you do not really own your workflow. Marcelo's setup kept working because the useful parts were his.