Half of the Open Claw users I talk to are running their Open Claw instance using their Anthropic Max subscription via an OAuth loophole to avoid paying per-token API costs. It’s technically against Anthropic’s terms of service. Everyone knows it and has been hoping they can continue skating under the radar undetected.
Today Anthropic finally issued this ultimatum that confirms they are officially closing this loophole:
They’re now actively blocking third-party harnesses from using subscription tokens. This thread on Hacker News is hovering just under 1000 points and 700+ comments in hours. The community is furious.
I don’t blame Anthropic for this move at all. Any subscription model relies on some level of breakage and a vast majority of users not consuming more than their subscription is worth. Open Claw as an always-on, proactive, autonomous agent is throwing the curve, breaking their subscription economics and they know it. So they’re doing the rational response here.
If you’ve been running Opus through Open Claw on your Max plan, the math just changed dramatically.
Your options now:
Switch to API billing. Variable costs. Horror stories of $200+ inference bills from runaway agent loops are no longer hypothetical - they’re your new reality.
Downgrade to a lesser model that fits within whatever free tier you can cobble together. Good luck getting meaningful work done reliably.
Give up on the whole autonomous agent idea.
None of those are great paths.
I went through all of this already last month. I spent 30 days and 40+ hours trying to make Open Claw work for me. Tried every model. Local models on my Mac Mini weren’t powerful enough. Frontier models meant runaway costs because you have one runaway heartbeat checking snafu that eats up a bunch of tokens and don’t realize it until it’s too late. The memory was broken. The context kept falling apart. It couldn’t even parse my RSS feed to figure out my writing voice - a task that Claude Code handles trivially.
So I went a different direction.
I know from experience Claude Code is an incredibly capable platform and its Opus model is super competent at complex, goal-oriented, multi-stage tasks. My hypothesis was that I could setup a persistent, autonomous agent running on a Mac Mini using Claude Code to replace Open Claw as the always-on agent. Discord as the interface via their new channels feature. Voice notes in via Discord, trustworthy work out. Persistent memory that actually works and gets smarter over time. A heartbeat system for scheduled tasks. MCP tools wired into Gmail, my production database, my note system, browser automation, analytics - all the key systems.
This thing runs legally under the Max plan. Fully compliant with Anthropic’s terms of service. Fixed monthly cost. No OAuth hacking. No wondering if you’re going to get banned or forced to API-based billing.
Where Open Claw struggled to perform any kind of meaningful work reliably, my Claude Code-based agent is working with me like a dependable teammate.
Every day it drafts my outreach emails, scores leads from my production database, researches blog topics in my actual voice, monitors 30+ community groups and surfaces what matters in a morning briefing I can read in two minutes all while slaying issues on Github. I interact with it the way I’d converse with an employee: high-level goals instead of micromanaged instructions.
I’m demoing this setup live on Wednesday, April 8th here in Lisbon. Small group. I’ll show exactly how it works, the repo structure, the tools, and a side-by-side comparison of Open Claw fumbling a task that this setup handles cleanly.
If you’re in Lisbon, come see it: RSVP here
If you’re not in Lisbon, the recording will be available after. If there’s enough demand I’ll sanitize my repo and make an installer skill for it and put it out there for all the Open Claw Opus refugees who are up in arms today.
Anthropic’s move may have just handcuffed everyone’s Claws but it doesn’t have to mean the end to your subscription-based, fixed-cost, Opus-powered agent.
Sean Tierney builds software with AI at Vibecode Lisboa. He runs Build School, a 30-day cohort where non-technical founders ship their first product.