Anthropic Charged Devs Extra For Mentioning OpenClaw In Their Code
If your repo contained the word 'Hermes' or 'OpenClaw,' Claude Code quietly billed you for overage. Anthropic is now refunding affected users — but only after the tweets went viral.
Anthropic spent the week being the ethical AI company that quietly grepped your codebase for words it didn't like and charged you extra when it found them.
Matt Wolf broke this one down on his weekly news roundup, and the receipts are public. Here's what happened, why it matters, and what it tells you about the harness wars heating up between Anthropic, OpenAI, and the third-party agent ecosystem they all claim to support.
The bug
A tweet from Theo Brown lit the fuse:
"Fun fact, if you have a recent commit that mentions OpenClaw in a JSON blob, Claude Code will either refuse your request or bill you extra money. This is an empty repo. I'm just calling Claude Code directly."
Then a developer named Patel posted his own story. He was on Claude Max 20× — the $200/month plan. Out of nowhere, Claude Code hit him with an "out of extra usage" error. His dashboard showed 13% weekly usage, 0% current session, 86% of his plan untouched. But he'd been billed $200.98 in extra usage that should have been covered by his subscription.
He binary-searched his repos and commits to find what triggered it. The trigger was a single string: hermes.md in a recent git commit message.
No Hermes harness running. No OpenClaw running. Just the word in a commit message his repo happened to contain.
Anthropic's first response was worse than the bug
When Patel reported it, Anthropic support acknowledged the bug three times, called it "an authentication routing issue," thanked him for finding it — and then refused to refund. The exact response:
"I sincerely apologize for the disruption you experienced with the billing routing issue. We take service reliability very seriously. However, I need to let you know that we are unable to issue compensation for degraded service or technical errors that result in incorrect billing routing."
In other words: we know we charged you wrong, we know why, we're not giving the money back.
That's the response that broke containment. The original Reddit post, then Theo's tweet, then Patel's tweet — together they pulled in over 2.4 million views.
The walk-back
Once the numbers got loud enough, Tariq from Anthropic posted the official acknowledgment:
"Sorry, this was a bug with the third-party harness detection and how we pull git status into the system prompt. We're reaching out to affected users and giving them a refund plus another month's worth of credit."
Which is the right outcome. But the question Matt raises — and it's the right question — is whether Anthropic does any of this if the original tweets get 200 views instead of 2 million.
What this actually tells you
Four things worth pulling out of this:
1. Anthropic is grepping your code.
The billing bug is a symptom, not the disease. The real story is that Claude Code pulls your git status into the system prompt and the platform actively scans for keywords like hermes and openclaw to detect third-party harness usage — agent runtimes that aren't Anthropic's own Claude Code or Claude Co-work. The fact that string matching was wired into the billing path at all is the disclosure.
2. The "ethical AI" branding is taking damage.
One of the top replies to the original thread, quoted in Matt's roundup: "It's a good thing Anthropic tells us how ethical they are. Or we might worry about their morality." Anthropic has spent two years marketing itself as the safety-first lab — the one with red lines for the military, the one that publishes Constitutional AI papers. Charging users extra when their codebase mentions a competitor is a hard image to square with that pitch.
3. Theo's response is the line worth remembering.
"There's a certain class of bugs that suggests the thing you're trying to do is a bad idea. Worth reflecting on that."
That's the most useful frame here. The bug isn't really the bug. The bug is that the system was designed to do this in the first place. A different implementation would have caught it before it shipped, but a different design would have meant it couldn't ship at all.
4. The harness wars are real.
Anthropic, OpenAI, and Google are all watching the third-party agent layer — Hermes, OpenClaw, custom MCP setups — and trying to figure out where the value capture lives. This incident is what it looks like when one of them tries to draw a line and gets caught doing it. Expect more of this, more subtly executed, from every frontier lab over the next year.
What to do if you're an Anthropic customer
If you're on a Claude Max plan and you've seen unexplained overage charges in the last month, check your git status output. Search your recent commit messages and any JSON config files in your repo for the strings hermes, openclaw, or related agent harness names. If you find any and your bill looks wrong, screenshot the dashboard and reach out to support. Per Tariq's tweet, affected users are getting refunds plus a month of credit.
If you're not on Anthropic but you're building anything agentic on top of frontier models, read your provider's system prompt logs if you have access to them. If you don't have access, ask why. The amount of metadata that gets pulled into a closed-source system prompt is becoming a real cost-and-privacy variable.
The Bottom Line
Anthropic shipped a billing path that actively scanned customer code for strings it didn't like, charged users for having those strings, and only refunded after the tweets hit a million views. Even if every individual decision in the chain was made in good faith — auth bug, scoring heuristic, support script — the architecture that produced the outcome is the disclosure. Read the system prompt. Watch the bill. And don't let your harness loyalty be inferred from a commit message you wrote at 2am.