Anthropic Just Killed Claude OAuth in Third-Party Apps. Here Is What to Do.

If you woke up this week and your OpenClaw agents stopped working, you are not alone. Anthropic just confirmed what a lot of builders have been dreading: Claude OAuth is being removed from third-party apps. That means OpenClaw, Zed, and every other tool that let you route through your Claude Pro or Max subscription instead of paying per token. The flat-rate access is going away.

There is a quick config fix if your OAuth broke overnight. But there is also a harder truth here about where this is all heading. This post covers both.

What Is Actually Happening (Plain Language)

For a while, you could connect OpenClaw to your Claude subscription using OAuth. Same technology as "Sign in with Google." Instead of an API key and per-token billing, you authenticated with your subscription and your agents ran against your monthly plan. For a lot of people, that was $20/month for Pro or $200/month for Max. Predictable cost. No surprise bills. Clean.

Anthropic updated their Consumer Terms of Service on February 19, 2026. The key line: "Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted."

Enforcement had already begun in January. A server-side change on January 9 quietly blocked OAuth tokens from consumer plans everywhere except Claude.ai and the official Claude Code CLI. February's documentation update just made it official in writing.

And then last week, a GitHub commit from the OpenCode team showed them removing Claude OAuth support entirely and redirecting users to alternatives. That is what surfaced in the builder community and sparked the current wave of broken OAuth reports.

The Quick Fix If You Are Broken Right Now

Before you do anything else: check your model settings.

Here is what developer Dave Swift documented and tested: if you had switched your OpenClaw model to one running on a one-million token context window, that version breaks OAuth authentication. Full stop. It is a config issue, not a targeted enforcement action against you personally.

The fix:

• Open your OpenClaw model settings
• Switch to Claude Sonnet 4.6 or Opus 4.6 at the 200,000 token context window
• Start a fresh session
• Your stored OAuth credentials should work immediately without re-authenticating

People have confirmed this working in threads and comments. If you changed your context window setting recently, that is probably what broke you. This fix takes two minutes.

But and this is important: the fix restores access that is technically against Anthropic's terms of service. It is not a permanent solution. It is buying you time to migrate properly.

Why Anthropic Killed OAuth (The Real Reason)

I want to give you the honest version here, not the "big company being greedy" take that is flying around Twitter.

Claude Pro and Max were designed for human users doing individual work. The subscription pricing makes sense for someone chatting with Claude throughout their day, getting help on projects, writing things. It does not make sense as the cost basis for an autonomous agent pipeline running 24/7 with scheduled cron jobs, spinning up multiple sessions, making hundreds of tool calls per day.

Anthropic is in the business of training and serving frontier AI models. That costs real money. When power users figured out they could route entire agent pipelines through a flat $20/month subscription, a lot of people did exactly that. Which is understandable. Humans find the cheapest path. But at scale, it breaks the economics of the product.

On top of that: Anthropic is preparing for an IPO and working with enterprise customers. Terms of service that were loosely enforced for indie builders are now actively being tightened. That is just what happens as a company matures.

This was not malicious. It was inevitable.

The Bigger Problem OAuth Was Actually Solving

Here is the thing that frustrates builders, and rightfully so.

OAuth with a subscription gave you a hard cap on what your agents could spend. Your monthly subscription was your ceiling. The agent could run whatever it wanted, and at the end of the month you knew exactly what it cost. No open-ended variable bill. No waking up to unexpected charges because your agent hit a bug loop at 2 AM and made 10,000 API calls.

The Anthropic API does not have a hard spending limit. You can set soft alerts. But if your agent goes rogue in the middle of the night and burns through credits before any alert fires, you are finding out what happened after the fact, not preventing it.

This is a real problem for builders running autonomous agents. I have talked to people in Shipping Skool who had exactly this happen. Agent gets into a retry loop. Nobody is watching. Morning comes and the bill is way higher than expected.

OAuth was the elegant solution to that problem. And now it is gone. So you need to build cost control into your API setup instead of leaning on subscription limits.

How to Migrate to API Billing Without Breaking Everything

Okay, so what do you actually do? Here is a practical migration path.

Step 1: Get Your Anthropic API Key

Go to console.anthropic.com. Log in. Create a new API key. This is different from your Claude.ai subscription login. It is a separate billing relationship based on token consumption.

Set a monthly spend limit in the console. This is not a hard cutoff that prevents all overages, but it is much better than nothing. Set it to something you are genuinely comfortable with.

Step 2: Route the Right Tasks to the Right Models

Not every task your agent does needs Claude Opus. This is the single most important thing to understand about API cost management.

Here is how I think about it for my 13-agent pipeline on my Mac Mini:

• Opus 4.6: Main Atlas session for strategy, complex decisions, creative work
• Sonnet 4.6: Content writing, research, most cron job work
• Smaller models or local Ollama: Logging, formatting, simple transforms, repetitive tasks that do not require frontier intelligence

The math works out. A full day of agent work running Opus for everything might cost $20-30. The same day running smart model routing might cost $4-8. Most builders end up spending similar to what OAuth was costing them, with much better visibility into where the money is going.

Step 3: Add Rate Limiting to Your Agent Config

Your agents should not be able to make unlimited API calls in any given window. Add rate limits to your agent configuration. Set a max tokens per session. Set a daily budget cap in your orchestration logic.

This is the architecture work that OAuth was letting you skip. Now you have to do it. But it actually makes your whole system more robust and predictable.

Step 4: Consider Local Models for Light Tasks

Ollama is free. You run a local model on your machine, zero API cost. For lightweight tasks like summarizing, reformatting, simple classification, and checking logs, a local 7B or 13B model does the job fine.

Hybrid architecture: local models handle the boring stuff, API handles the smart stuff. This is where the experienced builders are landing. You end up with a very efficient pipeline and API costs that stay predictable because you are only calling expensive models when they actually add value.

What My Own Migration Looks Like

I am running 13 agents on a Mac Mini in Silverton, Oregon. The full Wu-Tang pipeline: The RZA for scripts, Ghostface for intel, Inspectadeck for X posts, Method Man for clips, the GZA for newsletters, and Atlas as the main brain I talk to all day.

My setup was on Claude OAuth for the main Atlas session and API for everything else. So the OAuth change affects one agent for me directly.

What I am doing: migrating Atlas to API with Opus 4.6, adding a daily token budget cap, and using Sonnet more aggressively for sessions where I am not doing anything that genuinely needs Opus-level reasoning. Which is honestly most of them.

The bigger change is adding the rate limiting logic I had been lazy about because OAuth made it unnecessary. The subscription cap was doing that work for me. Now I have to do it properly. Which is fine. It is better architecture.

Not For You If: When This Migration Is Genuinely Hard

I want to be straight about who this is difficult for.

If you built your entire agent setup around the assumption of a hard cost ceiling and you have agents running at high frequency all day, this migration requires real work. Not impossible work. But real work. You need to audit your token usage, understand what each agent actually costs at the API level, and build in the controls that OAuth was handling automatically.

If you are a solo builder who runs agents occasionally and does not have complex automation, the migration is simple. Switch to API, set a reasonable spend limit, probably spend $10-20 a month, done.

The people who need to think hardest about this are running continuous, high-frequency agent pipelines where the cost ceiling really mattered. For that group: smart model routing and local model integration are the path forward.

The Bigger Picture: What This Means for Builders

This is what maturing platforms do. The grey area that benefited early adopters gets formalized as the company grows. Anthropic is building a real business with real investors and real legal obligations. The practices that existed in a grey zone are now officially off the table.

That is not the end of the opportunity. It is the beginning of the mature phase of the opportunity. The builders who establish compliant, sustainable foundations now are the ones who will not have to scramble and rebuild later.

I saw the same thing with Whatnot reselling. Early grey-zone practices that made things easier all got shut down as the platform grew. The sellers who had already built clean operations barely felt the changes. The ones running on workarounds had to restart from scratch.

Build the right way from the start. Even when the wrong way is cheaper right now. Especially then.

FAQ: Claude OAuth Removed From Third-Party Apps

Why did Anthropic remove Claude OAuth from third-party apps?

Anthropic updated their Consumer Terms of Service to explicitly ban OAuth tokens from Claude Free, Pro, and Max plans in any third-party tool. Subscription plans were designed for individual human use, not autonomous agent pipelines. They want builders to use the pay-per-token API for agent use cases.

Is there a quick fix if my Claude OAuth stopped working in OpenClaw?

If you recently switched to a model with a one-million token context window, that breaks OAuth. Switch back to Sonnet 4.6 or Opus 4.6 at 200,000 tokens, start a fresh session, and your credentials should restore. This is a temporary fix, not a permanent solution to the TOS situation.

Can I still use my Claude Pro or Max subscription with OpenClaw?

Technically a workaround exists. But Anthropic's Consumer Terms of Service explicitly say it violates their terms. The official path is using the Anthropic API with a pay-per-token key.

How much does it cost to run OpenClaw on the Anthropic API?

Depends on model choice and usage volume. Most builders using smart model routing report spending $20 to $60 per month for a full agent pipeline. The key is not running every task through Opus when Sonnet or a local model will do.

What is the best alternative to Claude OAuth for running AI agents?

Three options: (1) Anthropic API with smart model routing, (2) local open-source models via Ollama for lightweight tasks at zero cost, or (3) a hybrid of both. Most serious operators are landing on option 3.

Will Anthropic ban my account for using Claude OAuth with OpenClaw?

Active account bans have been rare so far, but the enforcement direction is clear. Anthropic has explicitly reserved the right to enforce their terms. Do not build your business on a foundation that depends on enforcement staying lax.

---

If you want to work through this migration alongside other builders figuring out the same thing, that is exactly what we do inside Shipping Skool. We have live calls six days a week. 170 members building right now. Real questions, real answers, no fluff.

Come build with us.