How to Bypass Anthropic's OpenClaw Block
Anthropic cut Max subscriptions for OpenClaw. But their official Agent SDK still works. Here's how to get your agents back in 5 minutes.
On April 4, 2026, Anthropic pulled the plug. Claude Max subscriptions no longer work with third-party tools like OpenClaw. Overnight, thousands of users running autonomous AI agents on their $200/month plans hit a brick wall. The Hacker News thread blew up. Blog posts piled on. The community was furious.
But amid all the noise, almost nobody noticed the door Anthropic left wide open: the Claude Agent SDK.
What Happened on April 4
Anthropic announced that OAuth tokens from Claude Pro and Max subscriptions are now restricted to official tools only — Claude.ai and Claude Code. If you were using a third-party harness (OpenClaw, Meridian, CLIProxyAPI) to route calls through your Max subscription, that’s over. Tokens are now scoped and verified server-side.
The stated reason: third-party harnesses put “disproportionate pressure” on infrastructure. Translation: some Max users were burning through hundreds of dollars worth of API calls per day on a $200/month flat rate. The math didn’t add up.
According to TechCrunch, Anthropic is offering a one-time credit equal to one month’s subscription to ease the transition, plus discounts of up to 30% on prepaid API credits. But for power users, the writing’s on the wall: the era of unlimited agent usage on a flat subscription is over.
Why the Old Proxies Are Dead
Every workaround that existed before April 4 — Meridian, CLIProxyAPI, homegrown OAuth proxies — worked the same way: intercept Claude Code’s auth token and replay it against the Anthropic API.
Old proxy: Your script → stolen OAuth token → Anthropic API
↑
Anthropic now checks
the caller is actually Claude CodeAnthropic shut that down by verifying client identity server-side. An OAuth token alone isn’t enough anymore — the request has to come from the real Claude Code binary. Header-replaying proxies are dead.
The Distinction Nobody Noticed
Here’s what makes this interesting. Anthropic blocked OAuth proxies, but explicitly left open its own Claude Agent SDK.
The technical difference is fundamental:
| OAuth Proxy (blocked) | Agent SDK (allowed) | |
|---|---|---|
| Method | Replays OAuth token to the API | Spawns the real Claude Code CLI as a subprocess |
| Auth | Copied token, spoofed headers | Claude Code authenticates itself via OAuth |
| Anthropic sees | Unauthorized third-party call | Legitimate Claude Code call |
| Status | Blocked since April 4 | Allowed (official SDK) |
The Agent SDK doesn’t cheat. It spawns the actual claude binary as a subprocess, which authenticates normally with your Max subscription. From Anthropic’s perspective, Claude Code is making the request — because it literally is.
As The New Stack confirmed, Anthropic encourages experimentation through its official SDK. NanoClaw, a tool built entirely on the Agent SDK, remains authorized. The rule is straightforward: if you go through the official SDK for personal use, you’re in the clear.
claude-max-proxy: the open source fix
I built claude-max-proxy, an open source tool that reconnects your OpenClaw agents through the Agent SDK. Three commands and everything works again:
git clone https://github.com/nmarijane/claude-max-proxy
cd claude-max-proxy
pip install .
claude-max-proxyWhat the script does:
- Reads your existing
openclaw.json— agents, workspaces, models - Loads each agent’s files (SOUL.md, TOOLS.md, skills/) and assembles the system prompt
- Starts an HTTP server on port 3777
- The OpenClaw gateway sends messages → the proxy routes to the Agent SDK → Claude Code responds
No extra config. Your workspaces, your skills, your channels (Telegram, WhatsApp, Slack, Discord, Signal, iMessage, Teams, Matrix, all 20+) — everything is reused as-is.
Before : Channel → OpenClaw Gateway → Claude API ← BLOCKED
After : Channel → OpenClaw Gateway → claude-max-proxy
↓
Agent SDK
↓
Claude Code CLI
↓
Max subscription ← WORKSThe repo is on GitHub: github.com/nmarijane/claude-max-proxy
What this enables beyond OpenClaw
The Agent SDK turns your Claude Max subscription into a programmatic API. Even without OpenClaw, you can call Claude from any script:
from claude_agent_sdk import query, ClaudeAgentOptions, ResultMessage
async def ask_claude(prompt: str, system: str = None) -> str:
result = ""
async for msg in query(
prompt=prompt,
options=ClaudeAgentOptions(
system_prompt=system,
allowed_tools=["WebSearch", "WebFetch", "Read", "Bash"],
),
):
if isinstance(msg, ResultMessage):
result = msg.result
return resultTen lines. No API key. No per-token billing. Opus 4.6 included.
Use cases:
- AI cron jobs — a daily brief, weekly audit, nightly email digest
- Data pipelines — analyze CSVs, generate reports, extract information
- Augmented CI/CD — automated code reviews on every PR
- Multi-agent systems — multiple specialized agents on one machine, one subscription
The Tradeoffs You Should Know
The Agent SDK isn’t magic. There are real limitations:
Same quota as Claude Code. Your Max subscription has rate limits. The SDK shares them with your regular Claude Code usage. If you’re running 8 agents around the clock, you’ll hit the ceiling.
Sequential per session. The SDK launches one Claude Code process per request. No native parallelism within a single session. For multi-agent setups, you need to orchestrate calls sequentially.
No streaming. Responses come back as a single block, not token by token. For a real-time chatbot, that’s a dealbreaker. For async tasks (crons, pipelines), it’s a non-issue.
Risk of future lockdown. Nothing guarantees Anthropic won’t tighten the Agent SDK’s terms next. The April 4 precedent proves they won’t hesitate to cut access when the economics demand it. Anthropic has even stated: “If you’re building a business on the Agent SDK, you should be using an API key.”
What This Tells Us About Anthropic’s Strategy
The April 4 block isn’t just a technical move — it’s a strategic signal. Anthropic is drawing the boundaries of its platform:
- Official tools (Claude Code, Claude.ai, Agent SDK) are promoted and encouraged
- Third-party tools that circumvent billing are actively fought
- The paid API remains the recommended path for any commercial use
It’s the same playbook OpenAI ran with plugins, then GPTs, then its API: build an ecosystem where developers create value inside the platform, not around it.
The difference is that Anthropic is playing the transparency card. The Agent SDK is open source. The Claude Code CLI is free with a Max subscription. Personal and experimental use remains allowed. It’s an “open garden with guardrails” approach — you can experiment freely, but if you want to scale, you pay.
Compared to OpenAI, which charged investors $122 billion with no clear path to profitability, Anthropic is at least trying to build a sustainable business model.
Key takeaways:
- The April 4 block killed third-party OAuth proxies, but the official Claude Agent SDK remains authorized for personal use
- The technical distinction is clear: the SDK spawns the real Claude Code CLI, it doesn’t replay stolen tokens
- You can use your Max subscription as a programmatic API via the Agent SDK — autonomous agents, crons, pipelines, multi-agent systems
- Limitations exist: shared quota, sequential execution, no streaming, risk of future lockdown
- Anthropic’s strategy is pushing developers toward its official ecosystem — the Agent SDK is the encouraged path
FAQ
Is this a hack? No. The Claude Agent SDK is Anthropic’s official SDK, available as open source on GitHub. It uses the Claude Code CLI legitimately.
Does this violate the terms of service? For personal and experimental use, no. Anthropic has explicitly said they encourage experimentation through the SDK. For commercial use, they recommend switching to an API key.
How much does it cost? Nothing beyond your Max subscription ($100 or $200/month). No per-token billing. Opus 4.6 included.
